JP7510375B2 - Information processing device, information processing method, and program - Google Patents

Description

One embodiment of the present invention relates to an information processing device, an information processing method, and a program.

Against the backdrop of the liberalization of the electricity market, methods have been proposed for predicting the market price of electricity. Conventional methods are mainly based on actual data and regression models (statistical models), and while they have high prediction accuracy for short-term, high-frequency events, they have a problem in that they have low prediction accuracy for long-term, low-frequency events.

Japanese Patent No. 4329644

One embodiment of the present invention provides an information processing device, information processing method, and program that can accurately predict market prices.

In order to solve the above problems, according to one embodiment of the present invention, a classification unit classifies past data of a market price of an object into a plurality of data according to a frequency with which the market price is adopted in a trading market of the object,
An acquisition unit that acquires the plurality of data within a period according to a frequency;
An information processing device is provided, comprising: a supply and demand curve estimating unit that estimates a supply and demand curve indicating a relationship between a bid volume when buying and selling the object and a buying and selling price of the object based on the acquired period.

FIG. 1 is a block diagram showing a schematic configuration of an information processing apparatus according to an embodiment. 11 is a flowchart showing an outline of a processing procedure of a performance learning phase. 11 is a flowchart outlining the processing steps of the predictive analysis phase. 11 is a flowchart showing details of a processing procedure of the performance learning phase. 11 is a flowchart showing details of the processing steps of the predictive analysis phase. Frequency distribution chart of major supply and demand components in the Tokyo area. A scatter plot showing the relationship between demand and price in the Tokyo area. A scatter plot showing the relationship between thermal power generation volume and prices in the Tokyo area. A waveform diagram of the supply curve. A scatter plot of thermal power supply and area prices created from actual values of the JEPX spot market. Frequency distribution chart of prices in the Tokyo area. A diagram showing learning data and a supply curve model for a supply curve learned from actual values. A graph showing the results of visualizing 365-day supply curve parameters. A visualization example of actual prices in the Tokyo area, supply volume, and forecast values calculated from the supply curve for each day. A graph showing the hourly/monthly averages of actual and forecast market prices for 2019. A graph showing the hourly/monthly averages of actual and forecast market prices for 2019. A graph visualizing the average parameter values by year/month to capture changes over time in the thermal power supply curve. A graph visualizing the average parameter values by year/month to capture changes over time in the thermal power supply curve. A chart showing the supply and demand curves published by JEPX in its trading performance data. Scatter plot of buying and selling bid volumes and system prices. Scatter plot of buying and selling bid volumes and system prices. A scatter plot of buy/sell execution rate (execution rate) and system price. A scatter plot of buy/sell execution rate (execution rate) and system price. A visualization example of actual data for fiscal years 2019 and 2020 and the learned supply curve. A diagram showing the calculation results of the system price by combining the learned supply curve and the execution rate. A visualization example of a scatter plot of actual data and a supply curve with different selection periods. A graph showing predicted values based on the supply curve for each training data. A graph showing predicted values based on the supply curve for each training data. FIG. 13 is a graph showing the system price predicted based on different supply curves for the learning data selection period. FIG. 13 shows an example of evaluation of prediction accuracy focusing on price increases/declines. FIG. 13 is a diagram showing the relationship between the price threshold and the F value for price hike prediction. FIG. 13 is a graph showing the relationship between the price threshold and the F value for crash prediction. A diagram showing an example of forecasted market prices visualized over time. A diagram showing the estimated values of renewable energy obtained from actual supply and demand data in the Tokyo area and the assumed values based on forecast analysis. A graph showing the results of visualizing the average market price by hour/month assuming an increase in renewable energy. A graph showing the results of visualizing the average market price by hour/month assuming an increase in renewable energy.

Below, an embodiment of an information processing device, an information processing method, and a program will be described with reference to the drawings. The following description will focus on the main components of the information processing device, but the information processing device may have components and functions that are not shown or described. The following description does not exclude components and functions that are not shown or described.

Fig. 1 is a block diagram showing a schematic configuration of an information processing device 1 according to one embodiment. The information processing device 1 in Fig. 1 has a function for predicting the market price of an object. The object is any item that is bought and sold in the market, such as electricity, water, gas, etc. In the following, an example in which the object is electricity will be mainly described.

The information processing device 1 in FIG. 1 includes a classification unit 2, an acquisition unit 3, a learning unit 4, and a supply-demand curve estimation unit 5. Of these, the classification unit 2, the acquisition unit 3, and the supply-demand curve estimation unit 5 are essential components.

The classification unit 2 classifies the actual data of the market price of the object into a plurality of data according to the frequency with which the market price is adopted in the trading market for the object. More specifically, the classification unit 2 classifies the actual data of the market price of electricity into a plurality of data according to the frequency with which each market price is adopted. The actual data is data including the market price of electricity, the supply amount of electricity, the area in which electricity is used, and the time period, date, season, etc. in which electricity is used. In the electricity market, the market price of electricity changes according to the demand and supply of electricity. In general, among the actual data, data with a low market price of electricity appears frequently, and data with a high market price of electricity appears infrequently. Therefore, the classification unit 2 classifies the actual data of the market price of electricity into high-frequency data (first data) and low-frequency data (second data). The high-frequency data is actual data with a low market price of electricity, and the low-frequency data is actual data with a high market price of electricity.

In this embodiment, an example is described in which the classification unit 2 classifies performance data into high-frequency data and low-frequency data, but the data may be classified into three or more types of data each having a different frequency.

The acquisition unit 3 acquires each of the multiple data classified by the classification unit 2 within a period according to the corresponding frequency. For example, the acquisition unit 3 makes the period for acquiring high-frequency data shorter than the period for acquiring low-frequency data. High-frequency data is data that appears frequently within a short period of time, such as by time period or day, so by acquiring the high-frequency data within a short period of time, the trend of the data can be understood. On the other hand, low-frequency data is data that appears infrequently within a long period of time, such as by year, so by acquiring the low-frequency data over a long period of time, the trend of the data can be understood. The acquisition unit 3 may temporarily store the data acquired within a set period of time in a storage unit (not shown) or the like.

The learning unit 4 estimates a supply and demand curve that indicates the relationship between the bid volume and the buying and selling price of the target object when buying and selling the target object based on the acquired period. More specifically, the learning unit 4 generates learning data for predicting performance data based on the data acquired by the acquisition unit 3. More specifically, the learning unit 4 generates learning data based on high-frequency data and low-frequency data acquired within a specified period. If learning data is generated using only high-frequency data, performance data when electricity demand increases temporarily cannot be reflected. Performance data when electricity demand increases temporarily is low-frequency data, and the learning unit 4 can generate learning data that reflects the actual supply and demand situation of electricity by generating learning data using both high-frequency data and low-frequency data.

The supply and demand curve estimation unit 5 generates supply and demand curve parameters for estimating a supply and demand curve showing the relationship between the bid quantity and price of the target object based on the learning data. The supply and demand curve is used to predict the market price of electricity. The supply and demand curve parameters are used to generate a supply and demand curve model.

As described below, the supply and demand curve includes a supply curve and a demand curve, and the intersection of the supply curve and demand curve is the market price.

The information processing device 1 of FIG. 1 may include a performance data storage unit 6. The performance data storage unit 6 stores past performance data on electricity. For example, the performance data storage unit 6 stores the amount of electricity supply, the amount of electricity demand, and the market price in association with each other for each area. The classification unit 2 described above may acquire performance data from the performance data storage unit 6 and classify the data into high frequency data and low frequency data.

The information processing device 1 of FIG. 1 may include a supply curve generating unit 7. The supply curve generating unit 7 generates, for example, a step function as a supply curve that is a part of the supply and demand curve. The step function is a function in which the price increases stepwise according to the amount of electricity supplied. The step function may be, for example, a sigmoid function. On the other hand, the demand curve may be approximated to a straight line according to the purchase bid amount. The supply and demand curve estimating unit 5 may generate demand curve parameters based on the step function and the demand curve.

The information processing device 1 of FIG. 1 may include a fuel cost prediction unit 8. The fuel cost prediction unit 8 predicts the fuel cost. The supply and demand curve estimation unit 5 may estimate the supply and demand curve based on the fuel cost predicted by the fuel cost prediction unit 8 and learning data. When the fuel cost changes, the price of the electricity supplied by the power supplier changes, and the shape of the supply curve changes. Therefore, by estimating the supply and demand curve taking into account the predicted fuel cost, the prediction accuracy of the supply and demand curve can be improved. For example, the above-mentioned supply curve generation unit 7 adjusts the height of the steps of the step-like function based on the fuel cost predicted by the fuel cost prediction unit 8.

The information processing device 1 in FIG. 1 may include an area supply and demand data acquisition unit 9, a power source configuration prediction unit 10, and a residual demand estimation unit 11.

The area supply and demand data acquisition unit 9 acquires actual supply and demand data for each area based on the actual data stored in the actual data storage unit 6.

The power source configuration prediction unit 10 predicts the configuration of power sources that will generate electricity. Power sources include thermal power, hydroelectric power, nuclear power, pumped storage, and renewable energy. Renewable energy is power generation from solar, wind, geothermal, solar heat, biomass, and the like. The power source configuration prediction unit 10 predicts how much electricity each power source will generate.

The residual demand estimation unit 11 generates purchase bid volume forecast data based on actual supply and demand data for each area and the predicted power source configuration. Since the price per unit of power varies depending on the power source and also varies depending on the time period and area, the residual demand estimation unit 11 forecasts the purchase bid volume based on actual supply and demand data for each area and the predicted power source configuration.

The information processing device 1 in FIG. 1 may include a power supply bidding prediction unit 12 and a power supply operation estimation unit 13.

The power source bid prediction unit 12 predicts the amount of selling bids for each type of power source. For example, solar power only generates electricity during the day, whereas wind power is predicted to generate electricity day or night. Coal and nuclear power, which require a lot of effort to stop generating electricity, are cheap per unit of electricity, but because the policy is to generate electricity at all times, bids are predicted to be accepted at all times. Thermal power plants, such as those using natural gas, which can start generating electricity quickly, are predicted to only accept bids during times when electricity demand peaks. Thus, the power source bid prediction unit 12 predicts the amount of bids for each type of power source, taking into account the characteristics of each power source.

The power source operation estimation unit 13 generates selling bid volume forecast data based on actual supply and demand data for each area and the predicted bid volume for each power source. The amount and price of electricity supplied from each power source differs depending on the area, time period, etc., and electricity demand also differs depending on the area, time period, etc. Therefore, the power source operation estimation unit 13 predicts the selling bid volume taking into account the characteristics of each area and power source, etc.

The information processing device 1 in FIG. 1 may include a model generation unit 14 and a market price prediction unit 15.

The model generation unit 14 generates a demand curve model based on the supply and demand curve parameters. The market price prediction unit 15 inputs actual supply and demand data for each area, buying bid volume, and selling bid volume into the demand curve model to generate market price prediction data.

The information processing device 1 of FIG. 1 may include an output unit 16. The output unit 16 generates market price prediction data. The output unit 16 may also output the selling bid volume, the buying bid volume, and the supply and demand parameters in association with each other as prediction basis information for the market price prediction data.

Figures 2A and 2B are flowcharts showing an outline of the processing procedure of the information processing device 1 of Figure 1. The information processing device 1 of Figure 1 first performs a performance learning phase to learn a supply and demand curve based on past performance data on the market price of electricity. Next, a forecast analysis phase is performed to predict and analyze the market price of electricity based on the learned supply and demand curve. Figure 2A is a flowchart showing an outline of the processing procedure of the performance learning phase, and Figure 2B is a flowchart showing an outline of the processing procedure of the forecast analysis phase.

In the performance learning phase, as shown in FIG. 2A, the classification unit 2 classifies performance data into high-frequency data and low-frequency data (step S1). Next, the learning unit 4 generates learning data based on the high-frequency data and low-frequency data (step S2). Next, the supply and demand curve estimation unit 5 generates supply and demand curve parameters for estimating the supply and demand curve (step S3).

After step S3, the supply and demand curve estimation unit 5 may modify the supply and demand curve parameters for estimating the supply and demand curve based on the fuel cost predicted by the fuel cost prediction unit 8 and the learning data generated in step S2 (step S4).

In the prediction analysis phase, as shown in FIG. 2B, the residual demand estimation unit 11 predicts the purchase bid volume (step S11). The residual demand estimation unit 11 generates purchase bid volume prediction data based on the actual supply and demand data for each area and the predicted power source configuration.

Next, the power source operation estimation unit 13 predicts the selling bid volume (step S12). The power source operation estimation unit 13 generates selling bid volume prediction data based on the actual supply and demand data for each area and the predicted bid volume for each power source.

Next, the market price prediction unit 15 inputs the actual supply and demand data for each area, the buying bid volume, and the selling bid volume into the demand curve model to generate market price prediction data (step S13).

Figure 3A is a flowchart showing an example of a detailed processing procedure for the performance learning phase, and Figure 3B is a flowchart showing an example of a detailed processing procedure for the predictive analysis phase.

In the performance learning phase, as shown in FIG. 3A, performance data on market transactions and area supply and demand is acquired (step S21). Next, the performance data is classified into high-frequency data and low-frequency data based on the value obtained by dividing the performance data by the price (step S22). Here, low-frequency data is data with a higher price among the performance data, and high-frequency data is data with a lower price than the low-frequency data. The high-frequency data is acquired within a shorter period of time than the low-frequency data.

Next, it is determined whether learning has been performed over the entire period of the actual data (step S23). If there is actual data that has not yet been learned, high-frequency data and low-frequency data are combined to generate learning data (step S24). Next, initial values of the supply curve parameters are generated (step S25).

Next, it is determined whether the error between the learned supply curve and the supply curve obtained from the actual data has converged (step S26). If the error has not converged, the error is evaluated using the learned data (step S27), and the supply curve parameters are corrected based on the evaluation results (step S28). Then, the processing from step S26 onwards is repeated. If it is determined in step S26 that convergence has occurred, the processing from step S23 onwards is repeated. If it is determined in step S23 that learning has been performed over the entire period of the actual data, the final supply curve parameters are output (step S29).

In the predictive analysis phase, as shown in FIG. 3B, actual supply and demand data for each area, the supply curve parameters obtained in the actual learning phase in FIG. 3A, and a predictive analysis scenario are obtained (step S31).

Next, it is determined whether all components of electricity for each area, specifically the demand data and supply data, have been corrected (step S32). If there is data that has not yet been corrected, it is corrected (step S33). Then, the process returns to step S32.

When it is determined in step S32 that the correction of all components is complete, the remaining demand is calculated (step S34). Here, the buying bid volume is predicted.

Next, the supply curve is calculated (step S35). The supply curve is then used to convert the residual demand into a market price (step S36).

Next, it is determined whether or not market prices have been predicted for the entire period to be predicted (step S37). If predictions have not yet been made for the entire period, the statistical value of the market price for the relevant period is calculated (step S38), and the process returns to step S37. If it is determined in step S37 that predictions have been made for the entire period, the supply and demand components and market price predictions are output (step S39).

The following is a detailed description of an overview of the information processing device 1 according to this embodiment. The following describes a method for predicting a supply and demand curve based on actual data on electricity demand and supply published by the Japan Electric Power Exchange (hereinafter, JEPX), and predicting the market price of electricity based on the predicted supply and demand curve.

1. Relationship between market prices and supply and demand components Figure 4 is a frequency distribution diagram (histogram) of the main components of supply and demand in the Tokyo area. The horizontal axis of Figure 4 is power supply or demand (GW), and the vertical axis is frequency. Figure 4 shows histograms g1 to g6 of demand in the Tokyo area, thermal power, hydroelectric power, solar power generation performance, pumped storage, and interconnection lines. As Figure 4 shows, thermal power and solar power are the main components and account for a large proportion on the power supply side. Interconnection lines are generally on the positive side (import side), and although pumped storage is short in duration, it can be seen that the plus/minus amplitude is large.

Figure 5A is a scatter plot showing the relationship between demand and price in the Tokyo area. The horizontal axis of Figure 5A is demand in the Tokyo area (kWh), and the vertical axis is price in the Tokyo area (yen/kWh). Figure 5B is a scatter plot showing the relationship between thermal power generation volume in the Tokyo area and price. The horizontal axis of Figure 5B is thermal power generation volume in the Tokyo area (kWh), and the vertical axis is price in the Tokyo area (yen/kWh).

Both the scatter plots in Figures 5A and 5B show a tendency for prices to rise as demand increases, but the relationship with price is clearer for thermal power generation than for demand. As can be seen from Figures 5A and 5B, the market price of electricity is largely determined by the supply of thermal power.

1-1. Overview of the supply and demand curve model Electricity demand and solar power generation are strongly correlated with time of day, month, temperature, and solar radiation intensity, so they can be predicted from calendar information and weather conditions. Hydroelectric power generation is difficult to predict, as it has a weak correlation with season (month) and precipitation. Thermal power generation, pumped storage, and interconnection lines have a relatively strong correlation with market prices.

From the above, it seems possible to create a regression equation to calculate electricity demand and solar power generation from calendar information and weather conditions, and a regression equation to calculate market price from supply volumes of thermal power, pumped hydro, and interconnection lines, and then combine this with the supply and demand balance equation to realize a forecast analysis of market prices. Note that components that are difficult to predict, such as hydroelectric power, are not major components, so for simplification they can be considered as constants such as average values.

The amount of electricity in an area is balanced between supply and demand for each hour. The relational equation showing the balance between supply and demand for the main components of area supply and demand is expressed by the following equation (1). This relational equation means that the demand Qdemand on the left side is balanced with the supply from thermal power Qthermal, hydroelectric power Qhydro, solar power Qpv, pumped storage Qpumped, and interconnection line Qlink on the right side.
Q demand = Q thermal + Q hydro + Q pv + Q pumped + Q link ... (1)

Considering the controllability of each term in equation (1), quantities that are difficult to control, such as the demand amount and the amount of renewable energy power generation, are moved to the right-hand side, and quantities that are easy to control, such as the amount of thermal power generation and pumped storage power generation, are moved to the left-hand side. The right-hand side is considered to depend on calendar conditions and weather conditions, while the left-hand side is considered to depend on the market price. Thinking in this way, it can be interpreted that the right-hand side is determined when a forecast analysis scenario is given, and the market price is determined by the balance with the left-hand side when the supply curves of thermal power generation and pumped storage are given.
(Qthermal + Qpumped)(Price)
= (Qdemand - Qhydro - Qpv - Qlink) (Calendar, Weather) ... (2)

Thermal power is currently the largest supply component in the area's supply and demand, and the market price is determined mainly based on the relationship between fuel costs and thermal efficiency. It is thought that operation of pumped storage is decided depending on the market price, which is determined by the supply volume of thermal power. Thinking about it in this way, it can be thought that the market price is determined mainly once the supply volume of thermal power is decided. Below, we will explain an example of how the supply curve model, which determines the relationship between the supply volume and market price of thermal power, is learned from past performance, and then used to perform predictive analysis when an increase in renewable energy is assumed.

2. An example of a supply curve model: Area thermal power vs. area price 2-1. Supply curve function The supply curve P is modeled as a step-like function including parameters taking into account the trend of the actual data, as shown in the following formula (3). Each parameter {p}, {q}, r plays a role in describing the position of the step, and is defined as "when the supply volume exceeds q _k , the price rises from p _k-1 to p _k ." When learning parameters from actual data, it is convenient to be able to differentiate the modeling function with respect to the parameters, so as shown in FIG. 6, a smooth sigmoid function θ _r (x) is used as the step function. The smoothness parameter r should be selected to be a small value relative to the domain of the supply volume.

Fig. 6 shows the waveform shape of equation (3). The horizontal axis of Fig. 6 is the supply amount Q (GW), and the vertical axis is the price P (yen/kWh). When the performance data is given as a set of the supply amount Q and the price P {( _Qt , _Pt )|tεT}, the parameters {p}, {q} of the supply curve model are found as a solution to the problem of minimizing the mean squared error (MSE) with constraints, as shown in the following equation (4). Note that the symbol |T| is the number of elements in the timestamp set T related to the performance values.

That is, when the parameters are written as β≡(p, q, r), the supply curve model P learned from the sample set T of actual results is expressed by the following equation (5).

2-2. Learning the supply curve Figure 7A is a scatter plot of thermal power supply volume and area prices created from actual values of the JEPX spot market, and Figure 7B is a frequency distribution diagram of prices in the Tokyo area. The horizontal axis of Figure 7A is thermal power supply volume (kWh), and the vertical axis is price in the Tokyo area (yen/kWh). The horizontal axis of Figure 7B is price in the Tokyo area (yen/kWh), and the vertical axis is frequency.

The scatter plot in Figure 7A shows that the thermal power supply curve changes by season (month). The frequency distribution diagram in Figure 7B shows that there are area prices that occur frequently, and that the higher price side is in regular increments of 5.00 yen/kWh. Taking these characteristics into consideration, the supply curve parameters described above are determined.

The supply curve in the spot market may be different for each product block (time period). However, for product blocks that span consecutive times, power sources with similar operating conditions participate in the market under similar bidding conditions (quantity and price), so the supply curves resulting from the accumulation of bids are also likely to be similar. Therefore, the supply curves for the same day are assumed to be common across product blocks, and the parameters of the supply curve are determined for each day.

There are 48 actual values for each day, but they do not include samples where prices soared on the target day. If samples from when prices soared are not included when learning the parameters of the supply curve, the upper right part of the supply curve (where prices soar) cannot be learned correctly. The shape of the upper right part is important when considering price spikes. For example, if you want to predict market prices from the supply volume and supply curve, prediction errors will be large if the upper right part is inaccurate. Therefore, when learning the parameters of the supply curve, samples from when prices soared in the same month are added to the samples for the target day (48 points). Specifically, samples that are above the 99th percentile of prices from the actual values for the target month (48 x 30 ÷ 100 ≒ 15 points) are added to the learning data for each day.

Figure 8 shows the learning data and supply curve model for the supply curve learned from actual values using the above method. The horizontal axis of Figure 8 is thermal power (GW), and the vertical axis is price in the Tokyo area (yen/kWh). Plot JEPX in Figure 8 shows the actual value for the target day (48 points), plot P99 shows the top price actual value for the target month, curve SC1 shows the supply curve model learned only for the target day, and curve SC2 shows the supply curve model learned by adding the target month to the target day. The difference between curve SC1 and curve SC2 is in the upper right part where prices soar. Curve SC1 has characteristics that make it more susceptible to price surges than curve SC2.

Figure 9 shows the visualization of the supply curve parameters P and Q for 365 days in 2019. The horizontal axis of Figure 9 is the date, and the vertical axis is the value of the supply curve parameters P and Q. The high side of the supply quantity parameter Q {q0 to q6} changes roughly every month and is thought to be determined by additional data on price increases. The high side (40 yen/kWh or more) of the price parameter P {p0 to p6} was fixed at 10.00 yen/kWh increments because there is a regular bidding pattern in increments of 5.00 yen/kWh in the actual data. The parameter r, which determines the smoothness of the steps, was set to r = (Qmax - Qmin)/100 based on the width of the learning data. This supply curve model has only 10 parameters (7 x 2 - 4 =) for a learning data of 63 (48 + 15 =) and is also restricted by a monotonically increasing function, so overfitting is not thought to occur.

2-3. Evaluation of the supply curve Figure 10 is a diagram showing an example of visualization of forecast values calculated from actual prices P and supply volume Q in the Tokyo area and the supply curve F for each day. The horizontal axis of Figure 10 is the date, and the vertical axis is the price (yen/kWh) in the Tokyo area. Compared to model SC2, model SC1 is more prone to price spikes. In order to evaluate the quality of such a supply curve model, the forecast error Error of the area price is evaluated as shown in equation (6). Note that the forecast referred to here is "a forecast when the supply volume is known," and forecast errors arise only from inaccuracies in the supply curve.

Figures 11 and 12 are diagrams showing the hourly/monthly average values of actual and forecasted market prices in 2019. The horizontal axis in Figures 11 and 12 is the period (month), and the vertical axis is the average price (yen/kWh). The symbol SC2x is a forecast value using the supply curve parameters for 2018. Compared to forecast SC2, forecast SC1 predicts higher prices, which are more likely to spike during the evening peak. Forecast SC2 predicts an average value close to the actual value. Forecast SC2x predicts prices that are higher than the actual value overall. The difference between forecast SC2 and forecast SC2x comes from the differences in the supply curves for 2018 and 2019.

2-4. Changes in the supply curve Figures 13 and 14 are graphs that visualize the average parameter values by year/month in order to capture the time change of the thermal power supply curve. The horizontal axis of Figures 13 and 14 is thermal power (GW) and the price in the Tokyo area (yen/kWh). For example, the supply curve for 2019 shows a decrease in price on the low supply side compared to 2018, while the price increases on the high supply side. In addition, if we look at the supply volume where the price rises in the monthly supply curve, we can see that the supply volume tends to be low in the spring and increases in the summer and winter.

3. An example of a supply curve model: Selling contract rate vs. system price Figure 15 shows the supply and demand curves published by JEPX in the transaction performance data. The horizontal axis of Figure 15 is the total contract amount Q, and the vertical axis is the price (yen/kWh). The demand curve w1 is a step-like curve determined by the accumulation of buying bids, and the supply curve w2 is a step-like curve determined by the accumulation of selling bids. JEPX does not publish the details of the shape of the supply and demand curve, and only publishes the selling bid amount Qs, the buying bid amount Qd, the contract amount Q, and the system price P. As a result of the ``conditions for the occurrence of extreme events'' on the previous day, the rise and fall of the system price depends not only on the contract amount but also on the buying and selling bid amount. In other words, the functional shape of the supply curve is modeled by expanding the one-variable function ``area thermal power → area price'' of the previous period to the two-variable function ``selling bid amount, contract amount → system price'' of the current period.

To consider the functional shape of the supply curve, Figures 16A and 16B show scatter plots of buying and selling bid volumes and system prices, while Figures 17A and 17B show scatter plots of buying and selling contract rates (contract ratios) and system prices. Figures 16A and 17A are scatter plots for JEPX for the period from 2019/04/01 to 2020/03/31, while Figures 16B and 17B are scatter plots for JEPX for the period from 2020/04/01 to 2020/05/31. The horizontal axis of these plots is the total contract volume (kWh), and the vertical axis is the system price (yen/kWh).

The buy/sell contract rate is a dimensionless quantity (0-1) that normalizes the contract volume by the buy/sell bid volume. If we look at the conditions on the horizontal axis that correspond to the rise/fall of the system price, Figures 17A and 17B have narrower intervals than Figures 16A and 16B, and show clearer conditions. For example, Figures 17A and 17B show that when the sell contract rate is 0.55 or below, the price crashes, and when the sell contract rate is close to 1.0, the price soars. This result can be naturally interpreted as meaning that when the sell contract rate is high, there are fewer unsold items and the price rises, and when the sell contract rate is low, there are more unsold items and the price falls.

From the above, the supply curve used to predict price crashes/rises is the functional relationship between the selling execution rate learned from actual data and the system price.

Figure 18 shows an example visualization of actual data for fiscal years 2019 and 2020 and the learned supply curve. The left side of Figure 18 shows an example that includes price spikes/plummets in the data for the entire year of fiscal year 2019, while the right side shows an example that includes only price plummets over two months of fiscal year 2020 (2020/04/01-2020/05/31). The horizontal axis of Figure 18 is the contract ratio, and the vertical axis is the system price (yen/kWh). 2020 in Figure 18 does not include price spikes, so the top right of the supply curve cannot be accurately determined.

Figure 19 shows the calculation results of the system price by combining the learned supply curve and the contract rate. The horizontal axis of Figure 19 is the month or date, and the vertical axis is the system price (yen/kWh). The left side of Figure 19 shows the case for the whole of fiscal year 2019, and the right side shows the case for two months of fiscal year 2020 (2020/04/01-2020/05/31). For fiscal year 2019, the model captures price increases and falls, but for the case of fiscal year 2020, the model captures a fall but incorrectly predicts a price increase. This error comes from the fact that the actual data for 2020 does not include any price increases, and the shape of the supply curve is inaccurate on the side with a large contract rate.

3-1. Supply curve estimation method High-frequency data is selected by time period, season, etc. and used as learning data for the supply curve. In order to analyze the quality of the high-frequency data selection (acquisition) method, the acquisition period is changed to yearly, monthly, weekly, and daily, and evaluated. Figure 20 shows a scatter plot of actual data on a specific day (May 3, 2019) and a visualization example of the supply curve with different selection periods. The horizontal axis of Figure 20 is the selling contract rate, and the vertical axis is the system price (yen/kWh). Figure 20 shows the relationship between the annual system price y, the monthly system price m, the weekly system price w, and the daily system price d selling contract rate. Generally speaking, if the acquisition period of high-frequency data is selected to be short, the actual data can be explained well, but on the other hand, the number of parameters (sets) of the supply curve increases, which leads to overlearning and a decrease in predictability (generalization).

Figures 21 and 22 show predicted values (cases of price rise and crash) based on the supply curve for each learning data. The horizontal axis of Figures 21 and 22 is month, and the vertical axis is system price (yen/kWh). Figures 21 and 22 show examples of error evaluation of market price predicted from the supply curve with the sell contract rate known, and show evaluation examples where the acquisition period (y: year, m: month, w: week, d: day) is changed. Figure 21 shows the state of price rise, and the characteristics of the rise are common, and there is no significant difference in time periods other than the rise. Figure 22 shows the state of price crash, and the characteristics of the crash are common, and there is no significant difference in time periods other than the rise.

3-2. Evaluation of quantitative forecast accuracy Figure 23 is a graph showing the system price error predicted based on supply curves with different learning data acquisition periods. When the high-frequency data acquisition period is long, fewer parameters are required, but the error is somewhat large. The forecast error for "yearly" is somewhat larger than that for "daily", but is the same as that for "monthly" and "weekly", and can be said to be good considering the small number of parameters.

3-3. Evaluation of classification prediction accuracy Figure 24 is a diagram showing an example of evaluation of prediction accuracy focusing on price rise/fall. The horizontal axis of Figure 24 is the system price threshold (price) (yen/kWh). First, a price threshold that defines a price rise or fall is determined, and then a binary classification table (mixed table) is created from the predicted value and the actual value. For example, in the case of price rise prediction, the number of cases where the predicted value and the actual value are both equal to or greater than the threshold is classified as TP, and the number of cases where the predicted value and the actual value are both below the threshold is classified as TN. Then, the precision rate (the proportion of matches in the prediction) and the recall rate (the proportion of matches in the correct answer) are calculated, and the F score, which is the harmonic mean value of the precision rate and the recall rate, is calculated. The classification accuracy is evaluated based on this F score. Figure 24 shows the recall rate w4, the F score (F value) w5, and the precision rate w6.

The reason for using the F-score for evaluation is as follows. In general, precision and recall are in a trade-off relationship. For example, in the case of price rise prediction, if the price threshold is set high to increase precision, the proportion of correct answers that are correct will decrease, and the recall will fall. In this way, a good classification is achieved when precision and recall are high and balanced. The F-score of the harmonic mean can be interpreted as twice the resistance value of a parallel connection when precision and recall are considered as electrical resistance, and if the two are about the same, the F-score will also be about the same, but if one is close to zero, a short circuit will be created and the F-score will also be close to zero. In this way, the F-score is an index that provides a good balance between precision and recall, and in the case of price rise prediction, the threshold just before the F-score suddenly drops is an appropriate classification threshold.

Fig. 25 is a diagram showing the relationship between the price threshold and the F value for price hike prediction. The horizontal axis of Fig. 25 is the system price threshold (yen/kWh). Fig. 25 shows how the annual F value y, monthly F value m, weekly F value w, and daily F value d change depending on the system price threshold.
For example, if the price threshold for judging a price rise is 20 yen/kWh, the F-measure for all models is about 0.5. In other words, the recall and precision of whether the price is 20 yen/kWh or higher can be interpreted as being roughly 0.5.

Figure 26 shows the relationship between the price threshold and the F-value for crash prediction. The horizontal axis in Figure 26 is the system price threshold (yen/kWh). Figure 26 shows how the annual F-value y, monthly F-value m, weekly F-value w, and daily F-value d change depending on the system price threshold. For example, if the price threshold for a crash is set to 7.5 yen/kWh, the F-values for each model are approximately 0.8. In other words, the recall and precision rates for predicting whether the price will be 7.5 yen/kWh or less can be interpreted as being roughly 0.8.

4. Example of application of the supply curve model: Predictive analysis of the impact of renewable energy 4-1. Method of predictive analysis The change in market price when the amount of renewable energy introduced changes is evaluated using the supply curve model described above. If the amount of thermal power supply Q and the amount of renewable energy supply R are assumed to be α times the amount of renewable energy introduced, the amount of renewable energy supply R' ≡ αR and the amount of thermal power supply Q' ≡ Q-(α-1)R, respectively. In other words, the total value of thermal power and renewable energy is the same before and after the change (Q + R = Q' + R'). In this case, if we assume that the thermal power supply curve F is the same before and after the change, the market price after the change will be P' ≡ F (Q'). For example, we consider that the market price will fall by the amount that the amount of renewable energy increases and the amount of thermal power supply decreases. The decrease in the amount of thermal power supply is expressed by Equation (7).

Figure 27 shows an example of market prices predicted using this method visualized over time. The horizontal axis in Figure 27 is the month, and the vertical axis is the price (yen/kWh) in the Tokyo area. Curve SC1 is the predicted value of the supply curve model trained only on the target day. Curve SC2 is the predicted value of the supply curve model that takes into account times of price rise. Curve PV is the predicted value assuming twice the amount of solar power (α = 2), and shows a tendency for market prices to fall during the day. Curve WF is the predicted value assuming 20 times the amount of wind power (α = 20), and shows a tendency for market prices to fall throughout the day and night.

In the following, in order to more systematically examine the effect of introducing renewable energy on market prices, we will analyze the impact by time of day and season, assuming the future amount of renewable energy introduced.

4-2. Assumptions for forecast analysis Figure 28 shows the estimated values of renewable energy obtained from the actual supply and demand data for the Tokyo area and the assumed values for forecast analysis. If the renewable energy power generation is R, the capacity C, and the operating rate η, the annual average value of the renewable energy power generation is R = ηC. The estimated values in Figure 28 were obtained using the average operating rate of renewable energy (photovoltaic: about 13%, wind: about 18%). The assumed values in Figure 28 assume a photovoltaic capacity of 0 times and 2 times (PV0 and PV2) and a wind capacity of 20 times and 40 times (WF20 and WF40) based on 2019. The assumed PV0 was set for the purpose of investigating the "price without photovoltaic power" and the assumed WF20 for the purpose of investigating the "price of 14 GW wind power".

4-3. Results of forecast analysis Figures 29 and 30 are visualizations of the average market price by hour/month assuming an increase in renewable energy. The horizontal axis of Figures 29 and 30 is the period (hours), and the vertical axis is the average price (yen/kWh). For example, in the case of assumed PV2 (2x solar power), the market price falls only during the day, so the August peak is suppressed, but the 18:00 peak is not suppressed. On the other hand, in the case of assumed WF20 (220x wind power), the market price falls both during the day and at night, so the 18:00, April and August peaks are all suppressed. In this way, unlike solar power, wind power generates electricity at night, so it has a large effect on suppressing the peak of the market price.

The increase in renewable energy is expected to be approximately 14 GW for both the estimated PV2 and estimated WF20. In other words, when comparing the effect of suppressing market prices with the same installed capacity, wind power is expected to have nearly twice the effect of solar power.

The annual market price for the assumed PV0 without solar power will be approximately 4.51 yen/kWh higher. For the assumed PV0, market prices tend to soar during the day, in April, and in August. In general, demand is high during the day and in August, but low in April. The price hike in April in particular is caused by the seasonal decrease in thermal power supply under the assumption that the supply curve is the same as in 2019, while the demand for thermal power is higher than in the assumption without solar power. In reality, if the price hike is clearly a period in which the supply of thermal power increases and the supply curve changes, the previous assumption does not hold. In other words, the price hike in April for the assumed PV0 is unavoidable given the assumptions of the forecast, but is considered an overestimate.

In this way, in this embodiment, performance data is classified into high-frequency data and low-frequency data, learning data for predicting performance data based on these data is generated, and the supply and demand curve is estimated based on the generated learning data, so that the supply and demand curve can be estimated more accurately than when the supply and demand curve is predicted using only high-frequency data.

When estimating the supply curve, which is one of the supply and demand curves, a step function is used in which the market price increases in steps according to the amount of power supplied, making it possible to predict the supply curve easily and accurately.

By forecasting fuel costs and adjusting the height of the steps in the step function based on the forecasted fuel costs, we can estimate the supply curve taking fuel costs into account.

In addition, since the buying bid volume is predicted based on actual area supply and demand data and a prediction of the future power source composition, the buying bid volume can be predicted with high accuracy.In addition, since the selling bid volume is predicted based on actual area supply and demand data and a prediction of the bid volume for each power source, the selling bid volume can be predicted with high accuracy.

The market price of power sources can be predicted with high accuracy based on the predicted power source configuration, purchase bid volume forecast data, selling bid volume forecast data, and the supply and demand curve model. In addition, by outputting the corresponding selling bid volume, purchasing bid volume, and supply and demand curve parameters that serve as the basis for the prediction, this can be useful for estimating future supply and demand curves and forecasting market prices.

At least a part of the information processing device 1 described in each of the above embodiments may be configured with hardware or software. If configured with software, a program that realizes at least a part of the functions of the information processing device 1 may be stored on a recording medium such as a flexible disk or CD-ROM, and may be read and executed by a computer. The recording medium is not limited to removable recording media such as magnetic disks and optical disks, but may also be fixed recording media such as a hard disk device or memory.

The program that realizes at least a part of the functions of the information processing device 1 may be distributed via a communication line (including wireless communication) such as the Internet. Furthermore, the program may be encrypted, modulated, or compressed and distributed via a wired or wireless line such as the Internet, or stored on a recording medium.

The aspects of the present disclosure are not limited to the individual embodiments described above, but include various modifications that may be conceived by a person skilled in the art, and the effects of the present disclosure are not limited to the above. In other words, various additions, modifications, and partial deletions are possible within the scope that does not deviate from the conceptual idea and intent of the present disclosure derived from the contents defined in the claims and their equivalents.

1 Information processing device, 2 Classification unit, 3 Acquisition unit, 4 Learning unit, 5 Supply and demand curve estimation unit, 6 Performance data storage unit, 7 Supply curve generation unit, 8 Fuel cost forecasting unit, 9 Area supply and demand data acquisition unit, 10 Power source composition forecasting unit, 11 Residual demand estimation unit, 12 Power source bidding forecasting unit, 13 Power source operation estimation unit, 14 Model generation unit, 15 Market price forecasting unit, 16 Output unit

Claims (18)

a classification unit that classifies historical data on the market price of the object into first data, and classifies historical data having a market price of the object that is adopted in a trading market with a frequency higher than a predetermined threshold into second data, and classifies historical data having a frequency equal to or lower than the threshold into second data ;
an acquisition unit that acquires the first data and the second data within a period according to a frequency;
a learning unit that generates learning data for predicting the performance data based on the first data and the second data acquired by the acquisition unit;
a supply and demand curve estimating unit that estimates a supply and demand curve indicating a relationship between a bid volume when buying and selling the object and a buying and selling price of the object based on the learning data ;
a market price prediction unit that predicts the market price based on the supply and demand curve;
An information processing device comprising: The information processing device according to claim 1, wherein the acquisition unit sets a period for acquiring the first data to be shorter than a period for acquiring the second data. The information processing device according to claim 1 or 2, wherein the supply and demand curve estimation unit generates supply and demand curve parameters for estimating the supply and demand curve. The supply and demand curve includes a supply curve and a demand curve,
A supply curve generating unit that generates a step function as the supply curve,
The information processing device according to claim 3 , wherein the supply and demand curve estimating unit generates the supply and demand curve parameters based on the step function and the demand curve. The information processing device according to claim 4 , wherein the step function is a function in which the selling price increases stepwise according to the supply amount. The information processing device according to claim 4 , wherein the step function is a sigmoid function. The information processing device according to claim 4 , wherein the demand curve is approximated to a straight line corresponding to a buying bid amount. the target is electricity,
A fuel cost prediction unit for predicting a fuel cost is provided,
The information processing device according to claim 1 , wherein the supply and demand curve estimating unit estimates the supply and demand curve based on the predicted fuel cost and the learning data. the target is electricity,
A fuel cost prediction unit for predicting a fuel cost is provided,
The information processing device according to claim 4 , wherein the supply curve generating unit adjusts a height of a step of the step function based on the predicted fuel cost. the target is electricity,
A power source composition prediction unit that predicts a composition of power sources to be generated;
The information processing device according to claim 1 , further comprising a residual demand estimating unit configured to generate buying bid volume forecast data based on actual supply and demand data for each area and the forecasted power source configuration. The information processing apparatus according to claim 10 , wherein the power source includes at least one of thermal power, hydroelectric power, pumped hydroelectric power, and renewable energy. The information processing device according to claim 11 , wherein the renewable energy is power generation including at least one of sunlight, wind, geothermal, solar heat, and biomass. the target is electricity,
A power supply bid prediction unit that predicts the selling bid amount for each type of power supply;
The information processing device according to claim 1 , further comprising a power supply operation estimation unit configured to generate selling bid volume prediction data based on actual supply and demand data for each area and the predicted selling bid volume of the power source. the target is electricity,
A power source composition prediction unit that predicts a composition of power sources to be generated;
A residual demand estimation unit that predicts a purchase bid amount based on actual supply and demand data for each area and the predicted power source configuration;
A power supply bid prediction unit that predicts the selling bid amount for each type of power supply;
A power source operation estimation unit that generates selling bid volume prediction data based on actual supply and demand data for each area and the predicted selling bid volume of the power source;
a model generating unit that generates a demand curve model based on supply and demand curve parameters for estimating the supply and demand curve ,
10. The information processing device according to claim 1 , wherein the market price prediction unit inputs actual supply and demand data for each area, the buying bid volume, and the selling bid volume into the demand curve model to generate prediction data for the market price. The information processing device according to claim 14 , further comprising an output unit that outputs the market price prediction data. The information processing apparatus according to claim 15 , wherein the output unit outputs the selling bid volume, the buying bid volume, and the supply and demand curve parameters in association with each other as forecast basis information for the market price forecast data. The computer
Classifying historical data on the market price of the object into first data, data in which the market price is adopted in the trading market of the object more frequently than a predetermined threshold, and classifying historical data equal to or less than the threshold into second data,
Acquiring the first data and the second data within a period according to a frequency;
generating learning data for predicting the performance data based on the acquired first data and second data;
estimating a supply and demand curve indicating a relationship between a bid volume and a purchase and sale price of the object when buying and selling the object based on the learning data ;
An information processing method for predicting the market price based on the supply and demand curve . On the computer,
classifying historical data on the market price of the object into first data, the data in which the market price is adopted in the trading market of the object more frequently than a predetermined threshold, and into second data, the data in which the market price is equal to or less than the threshold ;
acquiring the first data and the second data within a period according to a frequency;
generating learning data for predicting the performance data based on the acquired first data and the acquired second data;
estimating a supply and demand curve indicating a relationship between a bid volume and a purchase and sale price of the object when buying and selling the object based on the learning data ;
and predicting the market price based on the supply and demand curve .