JP7638154B2 - PLANT STATE PREDICTION APPARATUS AND PLANT STATE PREDICTION METHOD - Google Patents

Description

This disclosure relates to a plant state prediction device and a plant state prediction method.

In various fields, future events are predicted using a predictive model that represents the relationship between explanatory variables and objective variables that are causally related to each other. For example, in the field of plant control, a predictive model is constructed in which the measurement results of sensors installed in the plant (hereinafter referred to as "state quantity measurements") are used as explanatory variables, and parameters related to the operating state of the plant are used as objective variables. The future behavior of the plant can be verified by calculating predicted values of the parameters using the predictive model.

Such a prediction model can be constructed by a linear regression method such as Lasso regression or multiple regression, as in Patent Document 1. However, linear regression methods are suitable for prediction targets with relatively simple behavior, and may be difficult to apply to phenomena that exhibit nonlinear behavior, such as asynchronous vibrations in steam turbines, whose occurrence mechanism is unclear. For such targets, a prediction model is constructed using a fully connected layer type neural network, as in Patent Document 2. When using a neural network, the prediction model is constructed by learning parameters such as the weights and biases of the neural network so that the output value of the prediction model when training data is input approaches the value of the teacher data.

JP 2007-2673 A JP 2013-40565 A

Methods using neural networks such as those described in Patent Document 2 can handle objects that cannot be handled by linear regression methods, but the network structure is complex and difficult to interpret.

At least one embodiment of the present disclosure has been made in consideration of the above circumstances, and aims to provide a plant state prediction device and a plant state prediction method that can accurately predict the state of a plant that has nonlinear behavior with respect to multiple state quantity measurements while ensuring good interpretability.

In order to solve the above problems, a plant state prediction device according to at least one embodiment of the present disclosure includes:
A plant state prediction device for predicting a state of a plant by inputting a plurality of state quantity measurement values related to the plant into a neural network, comprising:
The neural network comprises:
An input layer;
a linear combination layer capable of calculating an intermediate output value by linearly combining the plurality of state quantity measurement values input to each node constituting the input layer;
a fully connected layer in which nodes between layers are connected to each other and a predicted value regarding the state of the plant can be calculated by inputting the intermediate output value into a network including a nonlinear function;
Equipped with.

In order to solve the above problem, a plant state prediction method according to at least one embodiment of the present disclosure includes:
A plant state prediction method for predicting a state of a plant by inputting a plurality of state quantity measurement values related to the plant into a neural network, comprising:
a step of inputting the plurality of state quantity measurement values input to each node into a linear combination layer of the neural network and linearly combining the input state quantity measurement values to calculate an intermediate output value;
A step of calculating a predicted value regarding a state of the plant by inputting the intermediate output value into a fully connected layer of the neural network in which nodes between layers are connected to each other and include a nonlinear function;
Equipped with.

At least one embodiment of the present disclosure provides a plant state prediction device and a plant state prediction method that can accurately predict the state of a plant that has nonlinear behavior with respect to multiple state quantity measurements while ensuring good interpretability.

1 is a configuration diagram illustrating a plant state prediction device according to an embodiment. FIG. 1 is a diagram illustrating the configuration of a conventional fully connected neural network. FIG. 2 is a configuration diagram of the neural network of FIG. 1. 1. FIG. 4 is a diagram showing a comparison of prediction results obtained in a reference example in which linear regression is used instead of a neural network in the plant state prediction device of FIG. 1 with actual measurement results. 2 is a diagram showing a comparison of prediction results obtained by the plant state prediction device of FIG. 1 with actual measurement results. FIG. 2 is a verification result in which a correlation coefficient between a predicted value and an actual measurement value obtained by the plant state prediction device of FIG. 1 is compared with a reference example. 2 is a diagram showing a change over time in a predicted value when the number of node layers in a fully connected layer is changed in the plant state prediction device of FIG. 1 . FIG.

Below, several embodiments of the present disclosure will be described with reference to the attached drawings. However, the dimensions, materials, shapes, relative positions, etc. of the components described as the embodiments or shown in the drawings are not intended to limit the scope of the present disclosure and are merely illustrative examples.

Figure 1 is a schematic diagram of a plant state prediction device 1 according to an embodiment. The plant state prediction device 1 is a device for predicting the state of a plant, and is composed of, for example, a CPU (Central Processing Unit), a RAM (Random Access Memory), a ROM (Read Only Memory), and a computer-readable storage medium. A series of processes for realizing various functions is stored in a storage medium or the like in the form of a program, for example, and the CPU reads this program into the RAM or the like and executes information processing and arithmetic processing to realize various functions. The program may be installed in a ROM or other storage medium in advance, provided in a state stored in a computer-readable storage medium, or distributed via a wired or wireless communication means. The computer-readable storage medium may be a magnetic disk, a magneto-optical disk, a CD-ROM, a DVD-ROM, a semiconductor memory, or the like.

Such a plant state prediction device 1 receives as input a plurality of state quantities (state quantity measurements) x1, x2, ... related to the plant, which is the prediction target, measured by a measuring means (not shown) such as a sensor installed in the plant. These plurality of state quantity measurements x1, x2, ... are input to a neural network NN included in the plant state prediction device 1, which outputs a predicted value Vp of a parameter related to the plant state.

In the following embodiment, a case will be described in which a predicted value Vp is obtained for the blade vibration value of a turbine equipped in a plant as an example of a parameter related to the plant state, and the occurrence of asynchronous vibration of the turbine is predicted based on the predicted value Vp. The mechanism by which asynchronous vibration of a turbine occurs is unclear and it exhibits nonlinear behavior, but it can be suitably predicted by the plant state prediction device 1 and the plant state prediction method implemented by said device. Note that such a prediction target is merely an example, and can broadly include parameters related to the plant state, such as the temperature and pressure inside the plant.

Here, we will explain a conventional fully connected neural network NN' as the underlying technology for the neural network NN in Figure 1. Figure 2 is a configuration diagram of a conventional fully connected neural network NN'.

The fully connected neural network NN' comprises an input layer 2, a hidden layer 4', and an output layer 6. The aforementioned multiple state quantity measurements x1, x2, ... are input to the input layer 2. The multiple state quantity measurements x1, x2, ... input to the input layer 2 are configured so that, starting from each node N (neuron) of the input layer 2, the results are output from the output layer 6 as a predicted value Vp via the nodes N of each layer constituting the hidden layer 4'.

The hidden layer 4' has at least one node layer. In a fully connected neural network NN', the hidden layer 4' has multiple node layers, and the nodes N in each node layer are fully connected to each other. The connection strength between each node N is set by weights. The weights between each node N are set by tuning through learning so that the error between the network output value (output value from the output layer) and the correct answer is small.

Figure 3 is a diagram showing the configuration of the neural network NN in Figure 1. The neural network NN has an input layer 2 and an output layer 6, similar to the fully connected neural network NN' described above, but the configuration of the hidden layer 4 is different. Specifically, the hidden layer 4 of the neural network NN is configured to include a linear connection layer 8 and a fully connected layer 10.

The linear combination layer 8 is configured to be able to calculate an intermediate output value z by linearly combining a plurality of state quantity measurement values x1, x2, ... input to each node N of the neural network NN. Specifically, the intermediate output value z is calculated by linearly combining a plurality of state quantity measurement values x1, x2, ..., xn input to each node N of the input layer 2 using weights w1, w2, ..., according to the following equation.
z=w1・x1+w2・x2+...wn・xn+w0 (1)

The fully connected layer 10 is configured such that each node N between adjacent node layers is connected to each other, and a predicted value regarding the state of the plant can be calculated by inputting an intermediate output value z to a network including a nonlinear function nf. The nonlinear function nf included in the fully connected layer 10 can be any nonlinear function nf such as a ReLU function, a sigmoid function, or a tanh function. The selection of these nonlinear functions nf may be based on the method of preprocessing performed on the data (normalization processing (processing to convert data so that the maximum value is 1 and the minimum value is 0), standardization processing (processing to convert data so that the average value is 0 and the variance is 1)). In this embodiment, the ReLU function is used as an example of a nonlinear function. Since the ReLU function has a simple form compared to other nonlinear functions, it can effectively reduce the computational load in the fully connected layer 10. Another advantage is that there is no possibility of gradient disappearance (a state in which the network learning does not progress).

In addition, in the above-mentioned fully connected neural network NN', learning is unstable and the expected prediction accuracy may not be obtained. In contrast, in the neural network NN of the present embodiment, a part of the fully connected neural network NN' is essentially configured as the linear connection layer 8, which is advantageous in that learning becomes more stable.

In the neural network NN having such a configuration, the weights w1, w2, ..., wn of the linear combination layer 8 are set by learning the neural network NN and tuning it so that the error of the output predicted value Vp is small. In the linear combination layer 8, the intermediate output value z is calculated by a linear combination using the weights w1, w2, ..., wn tuned in this way, so that it has good interpretability. Specifically, by comparing the weights w1, w2, ..., wn obtained by tuning, the contribution (or importance) of each state quantity measurement value x1, x2, ..., xn to the plant state to be evaluated can be evaluated. For example, when the weights corresponding to the state quantity measurements x1, x2, x3, ..., xn are tuned to w1 = 3.5, w2 = 0.05, w3 = 10, ..., wn = -1, it can be seen that the state measurement value w3 with a relatively large weight is a factor with a large contribution to the predicted value Vp, while the state quantity measurement value w2 with the smallest weight is a factor with a small contribution to the predicted value Vp. It can also be seen that the state quantity measurement value wn with a negative weight is a factor that contributes in the opposite direction to the predicted value Vp compared to other state quantity measurement values with positive weights.

For example, if the vibration amplitude of the turbine blade is the predicted value Vp, various parameters such as the degree of vacuum, output, rotor thermal expansion difference, turbine valve opening, and inlet steam temperature of the turbine equipped in the plant are set as multiple state quantity measurements x1, x2, x3, ..., xn, and the weights w1, w2, ..., wn corresponding to each state quantity measurement value x1, x2, x3, ..., xn are compared to evaluate the contribution of each state quantity measurement value x1, x2, ..., xn to the asynchronous vibration, which is the phenomenon to be evaluated. For example, the plant may be evaluated by classifying the weights w1, w2, ..., wn into a group with a large contribution degree having a weight greater than the threshold value and a group with a small contribution degree having a weight equal to or less than the threshold value by comparing them with a preset threshold value. In this case, the state quantity measurement value with the largest weight may be identified as the main factor of the asynchronous vibration.

Here, the prediction results obtained by the plant state prediction device 1 having the above configuration will be explained in comparison with a reference example using linear regression. In the following explanation, the prediction target is asynchronous vibration, which is a nonlinear phenomenon that can occur in a turbine equipped in a plant, and the vibration amplitude of the turbine blades is treated as the predicted value Vp, but the prediction target is not limited to this.

Figure 4 is a diagram showing a comparison of prediction results obtained in a reference example in which linear regression (Lasso regression) is used instead of the neural network NN in the plant state prediction device 1 of Figure 1 with actual measurement results. Figure 4 shows the temporal changes in the predicted value Vp and the actual measurement value for the vibration amplitude of the turbine blade, and shows that the predicted value Vp is underestimated compared to the actual measurement value in the area surrounded by circle C1. This shows that it is difficult to accurately predict nonlinear phenomena such as asynchronous turbine blade vibration using linear regression.

On the other hand, Figure 5 is a diagram comparing the prediction results obtained by the plant state prediction device 1 in Figure 1 with the actual measurement results. In Figure 5, as shown by the circle C2, in the region where the vibration amplitude increases in the actual measurement values, indicating the occurrence of asynchronous vibration, the predicted value Vp also has an equivalent value. This shows that phenomena with nonlinear behavior, such as asynchronous vibration of turbine blades, can be predicted with high accuracy.

Next, FIG. 6 shows the results of a verification comparing the correlation coefficient between the predicted value and the actual measured value obtained by the plant state prediction device 1 in FIG. 1 with a reference example. In FIG. 6, the prediction target is asynchronous vibration, which is a nonlinear phenomenon that can occur in a turbine equipped in a plant, and the verification result using the fully connected neural network NN' shown in FIG. 2 is also shown as a reference example. FIG. 6 also shows the results of a similar verification performed multiple times. This correlation coefficient is an index that approaches "1" as the error between the predicted value and the actual measured value becomes smaller, and the correlation coefficient is shown in the learning section in which the neural network NN (or the fully connected neural network NN' in the reference example) is made to learn, and in the verification section in which predictions are actually made using the trained neural network NN (or the fully connected neural network NN' in the reference example).

As shown in FIG. 6, in all verification results, regardless of the learning interval or the verification interval, the prediction results by the plant state prediction device 1 according to this embodiment have correlation coefficients equivalent to those of the reference example using the fully connected neural network NN'. This shows that the plant state prediction device 1 can obtain sufficient prediction accuracy even if part of the neural network NN is a linear combination section.

Next, the verification results regarding the number of node layers of the neural network NN of the plant state prediction device 1 will be described. FIG. 7 is a diagram showing the time change of the predicted value Vp when the number of node layers m of the fully connected layer is changed in the plant state prediction device of FIG. 1. FIG. 7 shows the cases where the number of node layers of the fully connected layer is m=5 and m=50. In both cases, good reproducibility is obtained with respect to the actual measured value, but in the case of m=5, the mean squared error (MSE) is 0.0046 and the R value is 0.83, whereas in the case of m=50, the mean squared error is 0.0040 and the R value is 0.85. This shows that in the neural network NN, the more the number of node layers m of the fully connected layer, the more the expressive power of the neural network NN increases, and the more the prediction accuracy improves.

As described above, according to the above embodiment, the state of the plant can be predicted from a plurality of state quantity measurements using a neural network NN having a linear coupling layer 8 and a fully coupled layer 10. In such a neural network NN, the intermediate output value z calculated in the linear coupling layer 8 is input to a network including a nonlinear function nf in the fully coupled layer 10 to calculate a predicted value Vp, so that a state prediction can be performed with high accuracy, taking into account nonlinear phenomena that cannot be expressed by conventional linear regression. In addition, when calculating the predicted value Vp, the structure is simplified compared to the conventional neural network NN', which is a fully coupled type overall, by compressing the plurality of state quantity measurements x1, x2, ..., xn into the intermediate output value z, and the stability of learning is improved. On the other hand, since the intermediate output value z calculated in the linear coupling layer 8 is obtained as a linear combination of a plurality of state quantity measurements x1, x2, ..., xn, good interpretability can be obtained by checking the weights w1, w2, ..., wn of each state quantity measurement value x1, x2, ..., xn with respect to the intermediate output value z. This makes it possible to evaluate the contribution and sensitivity direction of the predicted value Vp, for example, by comparing the weights w1, w2, ..., wn of each state quantity measurement value x1, x2, ..., xn.

In addition, the components in the above-described embodiments may be replaced with well-known components as appropriate without departing from the spirit of this disclosure, and the above-described embodiments may be combined as appropriate.

The contents described in each of the above embodiments can be understood, for example, as follows:

(1) A plant state prediction device according to one aspect includes:
A plant state prediction device (e.g., the plant state prediction device 1 of the above embodiment) for predicting a state of a plant by inputting a plurality of state quantity measurement values (e.g., state quantity measurement values x1, x2, ..., xn of the above embodiment) related to a plant to a neural network (e.g., the neural network NN of the above embodiment),
The neural network comprises:
An input layer;
a linear combination layer (e.g., the linear combination layer 8 in the above embodiment) capable of calculating an intermediate output value (e.g., the intermediate output value z in the above embodiment) by linearly combining the plurality of state quantity measurement values input to each node constituting the input layer;
a fully connected layer (e.g., the fully connected layer 10 in the above embodiment) capable of calculating a predicted value regarding the state of the plant (e.g., the predicted value Vp in the above embodiment) by inputting the intermediate output value into a network in which nodes between layers are connected to each other and which includes a nonlinear function (e.g., the nonlinear function nf in the above embodiment);
Equipped with.

According to the above aspect (1), the state of the plant can be predicted from a plurality of state quantity measurements using a neural network having a linear combination layer and a fully connected layer. In such a neural network, the intermediate output value calculated in the linear combination layer is input to a network including a nonlinear function in the fully connected layer to calculate a predicted value, so that a state prediction can be performed with high accuracy, taking into account nonlinear phenomena that cannot be expressed by conventional linear regression. In addition, when calculating a predicted value, the plurality of state quantity measurements are compressed into an intermediate output value, so that the structure is simplified compared to conventional neural networks that are fully connected as a whole, and the stability of learning is improved. On the other hand, since the intermediate output value calculated in the linear combination layer is obtained as a linear combination of a plurality of state quantity measurements, good interpretability can be obtained by checking the weight of each state quantity measurement value relative to the intermediate output value. As a result, for example, by comparing the weight of each state quantity measurement value, it is possible to evaluate the contribution to the predicted value, the sensitivity direction, and the like.

(2) In another embodiment, in the above embodiment (1),
The importance of each of the plurality of state quantity measurement values is evaluated by calculating a weight (for example, weights w1, w2, ..., wn in the above embodiment) in the intermediate output value for each of the plurality of state quantity measurement values.

According to the above aspect (2), good interpretability can be obtained by evaluating the importance based on the weight of each state quantity measurement value with respect to the intermediate output value calculated by the linear combination unit through the linear combination of multiple state quantity measurement values.

(3) In another embodiment, in the above embodiment (2),
The system is configured to identify a major factor in the state of the plant based on whether the weight is greater than a preset threshold.

According to the above aspect (3), the weight of each state quantity measurement value obtained from an intermediate output value calculated by a linear combination of multiple state quantity measurement values is compared with a threshold value, whereby the main factors that cause the state of the plant to be predicted can be identified from the multiple state quantity measurement values.

(4) In another aspect, in any one of the above (1) to (3),
The linear combination layer is provided immediately after the input layer.

According to the above aspect (4), by providing a linear combination layer immediately after the input layer, the linear combination layer linearly combines and outputs the state quantity measurements input to each node constituting the input layer, thereby achieving good interpretability.

(5) In another aspect, in any one of the above (1) to (4),
The fully connected layer has more neurons per layer than the linearly connected layer.

According to the above aspect (5), by making the number of neurons per layer of the fully connected layer sufficiently greater than that of the linearly connected layer, the expressive power of the network is increased, and prediction accuracy can be improved.

(6) In another aspect, in any one of the above (1) to (5),
The linearly connected layer has one neuron per layer.

According to the above aspect (6), better interpretability can be obtained by configuring the linear combination layer with one neuron per layer.

(7) In another aspect, in any one of the above (1) to (6),
The nonlinear function is the ReLU function.

According to the above aspect (7), the ReLU function has a simpler form than other nonlinear functions, so the computational load can be reduced. Another advantage is that there is no possibility of gradient loss.

(8) In another aspect, in any one of the above (1) to (7),
The predicted value is a parameter related to asynchronous vibration of a turbine included in the plant.

According to the above aspect (8), it is possible to suitably predict parameters related to asynchronous vibrations occurring in a turbine installed in a plant as a prediction target. Such asynchronous vibrations have nonlinear behavior with respect to measurable state quantity values, and therefore may not be explained by state quantity measurements (e.g., output, vacuum level, etc.) that are said to have empirical correlations. However, by using a neural network having the above-mentioned configuration, it is possible to make accurate predictions.

(9) A plant state prediction method according to one aspect includes:
A plant state prediction method for predicting a state of a plant by inputting a plurality of state quantity measurement values (e.g., state quantity measurement values x1, x2, ..., xn) related to a plant to a neural network (e.g., the neural network NN of the above embodiment), comprising:
a step of inputting the plurality of state quantity measurement values input to each node into a linear combination layer (e.g., the linear combination layer 8 in the above embodiment) of the neural network and linearly combining the input state quantity measurement values to calculate an intermediate output value (e.g., the intermediate output value z in the above embodiment);
a step of inputting the intermediate output value into a fully connected layer (e.g., the fully connected layer 10 in the above embodiment) of the neural network in which nodes between layers are connected to each other and which includes a nonlinear function (e.g., the nonlinear function nf in the above embodiment), thereby calculating a predicted value regarding the state of the plant (e.g., the predicted value Vp in the above embodiment);
Equipped with.

According to the above aspect (9), the state of the plant can be predicted from a plurality of state quantity measurements using a neural network having a linear combination layer and a fully connected layer. In such a neural network, the intermediate output value calculated in the linear combination layer is input to a network including a nonlinear function in the fully connected layer to calculate a predicted value, so that a state prediction can be performed with high accuracy, taking into account nonlinear phenomena that cannot be expressed by conventional linear regression. In addition, when calculating a predicted value, the structure is simplified compared to conventional neural networks that are fully connected as a whole by compressing a plurality of state quantity measurements into intermediate output values, and the stability of learning is improved. On the other hand, since the intermediate output value calculated in the linear combination layer is obtained as a linear combination of a plurality of state quantity measurements, good interpretability can be obtained by checking the weight of each state quantity measurement value relative to the intermediate output value. As a result, for example, by comparing the weight of each state quantity measurement value, it is possible to evaluate the contribution to the predicted value, the sensitivity direction, and the like.

1 Plant state prediction device 2 Input layer 4 Hidden layer 6 Output layer 8 Linearly connected layer 10 Fully connected layer NN Neural network NN' Fully connected neural network N Node z Intermediate output values x1, x2, ..., xn State quantity measurement values w1, w2, ..., wn Weight

Claims (6)

A plant state prediction device for predicting a state of a plant by inputting a plurality of state quantity measurement values related to the plant into a neural network, comprising:
The neural network comprises:
An input layer;
a linear combination layer provided immediately after the input layer, the linear combination layer having one neuron per layer, and capable of calculating an intermediate output value by linearly combining the plurality of state quantity measurement values input to each node constituting the input layer;
a fully connected layer in which nodes between layers are connected to each other and a predicted value regarding the state of the plant can be calculated by inputting the intermediate output value into a network including a nonlinear function;
Equipped with
a weight in the intermediate output value for each of the plurality of state quantity measurement values, thereby evaluating the importance of each of the plurality of state quantity measurement values . 2. The plant state prediction device according to claim 1 , configured to identify a main factor in the state of the plant based on whether or not the weight is greater than a preset threshold value. 3. The plant state prediction device according to claim 1 , wherein the fully connected layer has a greater number of neurons per layer than the linearly connected layer. The plant state prediction device according to claim 1 , wherein the nonlinear function is a ReLU function. The plant state prediction device according to claim 1 , wherein the predicted value is a parameter related to asynchronous vibration of a turbine included in the plant. A plant state prediction method for predicting a state of a plant by inputting a plurality of state quantity measurement values related to the plant into a neural network, comprising:
a step of inputting the plurality of state quantity measurement values input to each node of an input layer into a linear combination layer provided immediately after the input layer of the neural network, the linear combination layer having one neuron per layer, and linearly combining the input state quantity measurement values to calculate an intermediate output value;
A step of calculating a predicted value regarding a state of the plant by inputting the intermediate output value into a fully connected layer of the neural network in which nodes between layers are connected to each other and include a nonlinear function;
Equipped with
a weight in the intermediate output value is calculated for each of the plurality of state quantity measurement values, thereby evaluating the importance of each of the plurality of state quantity measurement values .

Images