JP2003084805A - Plant load predicting method, steady plant simulator, optimum operating method of plant, and optimum designing method of plant - Google Patents

Abstract

PROBLEM TO BE SOLVED: To predict various plant loads by an analyzable neural network, and to provide a steady simulator, an optimum operating method and an optimum designing method of a plant. SOLUTION: Various kinds of loads such as power load, thermal load, air load and the like are predicated by using the neutral network of an analyzable hierarchical structure wherein a fully-coupling part having plural input layer elements and plural intermediate layer elements are provided, and the intermediate layer elements are coupled to all of the input layer elements, and a loosely- coupling part wherein the intermediate layer elements are coupled to a part of plural input layer elements. The neutral network is built by the learning using result values of input factors such as a weather condition, a day of the week, the distinction of business day and holiday, an operating condition of the plant, an operating pattern of a plant elements and the like, and result values of various loads, and the weather condition of the predicted date as a predictor is inputted with other input factors to predict the objective plant load.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a plant load predicting method for predicting various loads such as power load, heat load and air load in a plant by a neural network in order to enable efficient operation of the plant. A steady-state plant simulator for simulating the behavior of the entire plant by calculating the steady input / output state of plant components,
And formulate optimal operation that minimizes plant operating costs and exhaust gas as a nonlinear mixed integer programming problem,
Optimal operation method of a plant that solves this optimization problem by passing information to and from the stationary plant simulator, and evaluation of plant operation costs and exhaust gas within a certain period by using this optimal operation method The present invention relates to an optimum design method of a plant that formulates an optimum problem of combined capacity of plant constituent devices so as to optimize the value and solves it.

[0002]

2. Prior Art and Problems to be Solved by the Invention (1)
Regarding various load predictions of the plant Here, among various loads in the plant, the heat load will be described as an example. For example, in a heat storage system having a heat source device and a heat storage plant, part or all of the heat load that will be required on the next day is stored in the form of cold water, ice, hot water, etc. at night. Here, if the heat load of the next day and the current day can be accurately predicted, the necessary amount of heat can be stored without excess or deficiency, and the heat source device and the heat storage plant can be efficiently operated.

As a conventional heat load prediction method, a method using a reference heat load, a method using a time-series transition of the heat load, and a method using a correlation model between an external factor such as temperature and the heat load are used. It has been known.

Among the above-mentioned prior arts, the method of utilizing a reference heat load sets a reference heat load on a monthly basis and makes a prediction by correcting the reference heat load using the temperature and the actual heat load. ing. However, even in the same month, the actual heat load varies greatly depending on weekdays, Saturdays, holidays, operating conditions, weather conditions, etc., and therefore, there is a problem that highly accurate prediction cannot be performed. Further, as a method of utilizing the time series transition of the heat load, for example, an autoregressive model of the past three heat loads is known. However, this method cannot be applied when there are large fluctuations with time or when there are sudden fluctuations.

As a method of using a correlation model between external factors such as temperature and heat load, a method of constructing a linear model using multiple regression equations, or recently, a nonlinear model using a neural network is used. There is a way to build. However, the method using the linear multiple regression equation cannot consider the non-linearity of the target heat load.

On the other hand, although a non-linear modeling is possible by using a general conventional neural network, various input factors and the number of hidden layer neurons must be set by trial and error, so that the optimum neural network is not always available. The network is not always available. In addition, since the inside of the neural network is a black box, it is not possible to explain the reason why the predicted value was obtained when a certain input was given, which gives the operator anxiety. There are cases. Such problems are common not only to heat loads but also to other plant loads such as electric power loads and air loads.

As described above, each of the conventional load predicting methods has a problem in explaining the prediction accuracy and the reason for the prediction. In other words, the method that uses a linear model such as multiple regression cannot consider the nonlinearity of the prediction target, and the method that uses a nonlinear model cannot guarantee the prediction accuracy because the optimal nonlinear model may not be obtained. . Further, since the reason for the prediction cannot be explained accurately, there is a problem that the operator feels uneasy.

As described above, in constructing a predictive model of various plant loads such as electric power load, heat load, air load, etc., it is possible to model a non-linear relationship from the viewpoint of prediction accuracy, and to give it to the operator. It is desirable to be able to explain the reason for prediction from the perspective of security and reliability. However, conventionally, a neural network is often used to improve the prediction accuracy, and a multiple regression formula is often applied to satisfy the requirement of explaining the reason for prediction. At present, there is no realization of a prediction method that can satisfy one requirement at a time.

Therefore, the invention according to claim 1 enables highly accurate load prediction by a non-linear prediction model by a neural network, and further, the causal relationship and correlation between the load prediction value output from the prediction model and the input factor. An object of the present invention is to provide a method for predicting a plant load that facilitates the explanation of the reason for prediction based on relationships and the like.

(2) Steady-state plant simulator and optimum plant operation method Conventionally, various methods have been proposed for obtaining optimum plant operation using an optimization method in consideration of starting and stopping of various plant components. Treats the input / output characteristics and operation patterns of the plant components as a general linear equation and formulates them as a mixed integer programming problem, or solves the problem, or the nonlinearity of the input / output characteristics of the plant components ( In general, it is a method to obtain the optimum operation of the plant by using heuristic methods that can handle non-linearities such as expert systems and fuzzy inference, after taking into consideration non-continuous, non-differentiable, etc.) and operation patterns.

However, since the input / output characteristics of various plant components have non-linearity and have different operation patterns depending on the time of day, it is not possible to formulate them as general equations that can be directly handled by the optimization method. Usually difficult. Therefore, the optimum operation of the plant can be formulated as a nonlinear mixed integer programming problem that is the most difficult to solve among the optimization problems. However, after comprehensively considering discrete values such as the start / stop of plant components and continuous values such as the fuel injection amount of components, the constraints on the characteristics of each plant component and the supply / demand balance for each load type. Up to now, there has not been a method that can mathematically optimize a problem having non-linearity together with constraints such as and objective functions such as plant operation cost and minimization of exhaust gas.

In view of this, the invention described in claim 2 simulates the behavior of the entire plant in consideration of the nonlinear input / output characteristics and complex operation patterns of the plant component equipment, and the steady input / output of each component equipment. An attempt is made to provide a stationary plant simulator that outputs a state.
In addition, the invention described in claims 3, 4, and 5 formulates the optimum operation of the plant as a nonlinear mixed integer programming problem, and uses the output of the stationary plant simulator to change the nonlinear characteristics and operating patterns of the plant. After taking into consideration the solution that is close to the global optimum, Particle Swarm Optimizatio
It is intended to provide an optimum operation method of a plant in which the start / stop states of plant components and the amount of fuel injection are determined by using an optimization method such as n.

(3) Optimum design method for plant Conventionally, as a method for designing a plant, an optimum operation method constructed by a formulation based on the mixed integer programming problem described in (2) above or a heuristic method is used. There is known a method of determining an appropriate capacity of each plant constituent device while changing a load forecast value of a plant according to various scenarios and performing optimization many times.

In the conventional optimum design method as described above, when the optimum operation method formulated as the mixed integer programming problem is used, it is not possible to consider the nonlinear characteristics of the plant and the change of the operation pattern. It is difficult to generate a realistic solution. Also, when the heuristic method is used, although the nonlinear characteristics of the plant and the change of the operation pattern can be taken into consideration, it is possible to generate a solution that satisfies the constraint condition, and it is guaranteed that a solution closer to the global optimal solution can be generated. Not at all. Furthermore, since the accuracy of load prediction itself cannot be said to be satisfactory, there is a limit to the optimum design based on this load prediction value.

Therefore, in the invention described in claim 6, in consideration of the non-linear characteristics of the plant and the change of the operation pattern, the optimum operation method is used to evaluate the plant operation cost and the exhaust gas within a certain period. The optimum capacity of plant components is formulated as a combined optimization problem so as to optimize the value, and an optimum design method for a plant that seeks a solution close to a global optimum solution is provided.

[0016]

In order to solve the above-mentioned problems, the plant load predicting method according to claim 1 has a plurality of input layer elements and a plurality of intermediate layer elements, and all input layer elements are provided. A total connection part in which the intermediate layer elements are connected,
Using a neural network of a hierarchical structure that can be analyzed and has a loosely coupled part in which an intermediate layer element is connected to a part of a plurality of input layer elements, power load, heat load, air load, etc. of the target plant It is a method of predicting the various loads of the plant, which is the actual values of input factors such as weather conditions, the distinction between days of the week, weekdays and holidays, the operating state of the plant, the operation patterns of the plant components, and the actual results of the various loads. A neural network is constructed by learning using values and, and the weather condition of the forecast target day is input to the learned neural network together with other input factors according to the forecast value to predict the target plant load.

A steady-state plant simulator according to a second aspect of the invention is such that when the starting / stopping state of plant constituent equipment and the fuel injection amount at each control time are given as planned values,
The steady input / output state of each plant component is controlled for each control time by using the connection status of each plant component, the linear or nonlinear input / output characteristics of each plant component, and the operation pattern of each plant component. From the device on the fuel injection side to the end, it is sequentially calculated and output. Here, the non-linear input / output characteristics of each plant constituent device are expressed using, for example, the least squares method, the response surface method, a neural network, or the like.

The optimum operation method of the plant described in claim 3 satisfies the demand-supply balance for each load type with respect to the predicted load value obtained by the invention described in claim 1, and has restrictions on the characteristics of each plant constituent device. The objective function is to minimize the operation cost of the plant for at least a predetermined period, while taking into account the start / stop state of each plant constituent device for each control time that is a discrete value, and each for each control time that is a continuous value. The steady plant according to claim 2, wherein the amount of fuel injected into the plant components is used as a state variable, the optimization problem of the plant is formulated as a nonlinear mixed integer programming problem, and the initial value of the state variable is generated as a planned value. Input the start / stop status and fuel injection amount of each plant component in the search process to the simulator, and input / output state of each plant component to the objective function Arabic to be output, it is what determines the start and stop condition and the fuel injection amount of each plant component equipment near the final global optimal solution while successively corrected by the optimization technique the planned value.

An optimum operation method for a plant according to claim 4 is the same as the optimum operation method for a plant according to claim 3, wherein the optimization problem is solved by Particle Swarm Optimization.
Is the solution.

The optimum operation method of the plant described in claim 5 is the operation plan of the day created in the past (for example, the day before) using the optimum operation method of the plant described in claim 3 or 4, The operating state of the plant components at this point is set as the fixed initial state of the plant, and this initial state is considered as a constraint on the characteristics of each device, and the supply and demand balance based on the load forecast value for the rest of the day is constrained The operation plan of the day is corrected while considering it as a condition.

According to a sixth aspect of the optimum design method of the plant, in the design stage of the plant, the optimum value of the optimum capacity of each plant constituent device is determined from the device capacity values as the settable discrete values. 4. Formulating as a combination optimization problem to be determined, setting load patterns such as power load, heat load, and air load of the target plant within a predetermined period and design values of the capacity of each plant constituent device, and Using the start / stop status of each plant component and the fuel injection amount determined by the optimum operation method of the plant described in 4, the combination optimization method is used so that at least the evaluation value regarding the plant operation cost becomes the best. The optimum capacity of each plant constituent device is determined by sequentially changing the design value.

[0022]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below with reference to the drawings. 1. First, an embodiment relating to the plant load prediction method according to claim 1 will be described. In the present embodiment, when predicting various loads such as power load, heat load, and air load of the target plant, a linear model using a multiple regression equation that has been conventionally used and a non-linear model using a general neural network are used. Instead of using, for example, Japanese Patent Application No. 2000-166528 "Neural network and its learning method, analysis method and abnormality determination method", which is the prior application of the present applicant, and 2000-230665 "Neural network learning method", 2000-227057
We decided to use a neural network that is capable of the analysis described in "Structure of neural network" (causal relationship between arbitrary input factor and output, analysis of correlation).

Hereinafter, a general structure of a hierarchical neural network used as a prediction model in this embodiment and a learning method thereof will be described based on Japanese Patent Application Nos. 2000-166528 and 2000-230665.

First, the neural network structure will be described. In a normal hierarchical neural network, the input layer elements and the intermediate layer elements are all connected (referred to as fully connected portion), but the hierarchical neural network used in this embodiment has an arbitrary input as shown in FIG. Only the layer elements and any intermediate layer elements are coupled.

That is, this neural network has a fully coupled portion 11 made up of intermediate layer elements connected to all input layer elements and a loosely coupled portion made up of intermediate layer elements connected to some input layer elements. It consists of 12.
In this way, by providing the loosely coupled portion 12 in which the value of the weight (weighting coefficient or coupling coefficient) with some of the input layer elements is set to 0, it is completely compatible with the conventional hierarchical neural network. be able to. That is, since the neural network as shown in FIG. 1 can be realized by setting the value of the weight of a part of the conventional neural network consisting only of all connected parts to 0, the conventional neural network should be used as it is in the program. You can

For example, y = ax ₁ + bx ₂ + cx ₁ x ₂ + d (x ₁ , x ₂ are input variables, y is an output variable, a, b, c,
In the conventional regression equation in which d is expressed as a coefficient, the output factor is clear (y depends on the factors x ₁ and x ₂ ).
It is clear that the change of each of the factors), the function of each factor is clear (independent components of x ₁ and x ₂ (the first term and the second term on the right side of the regression equation) and their mutual components (the third term). )), The degree of influence of each factor on input / output is clear (x
_The degree of influence of ₁ , x ₂ , x ₁ x ₂ is a, b, c, respectively.
And there is d as a fixed component), and these features facilitate the internal analysis.

Therefore, as shown in FIG. 1, if a hierarchical neural network composed of a loosely coupled portion 12 that can be analyzed and a fully coupled portion 11 that guarantees accuracy is used, the ease of analysis of the regression equation can be improved by using the existing neural network. It can be realized by a network, and can be easily applied to an existing system due to compatibility with the conventional hierarchical neural network as described above.

There are various methods for learning the neural network shown in FIG. 1, but one example is as follows.
The learning method described in Japanese Patent Application No. 2000-166528 will be described with reference to the flowchart of FIG. First, the first step A1 is a normal neural network weight initialization process. Specifically, an initial value is given by a small number of random numbers for all weights between elements of each layer of a normal neural network. Here, one may programmatically define a neural network structure without any coupling between the input and intermediate layers. In this case, the processing of the following second and fourth steps is unnecessary.

The second step A2 is processing for changing the weight-initialized neural network to the neural network structure shown in FIG. That is, the coupling between any input layer element and the intermediate layer element is eliminated. Here, the simplest method for deleting a bond is to replace an arbitrary weight value with 0.

The third step A3 is the calculation of the weight correction amount of a normal neural network. The correction amount of the weight between the input layer element and the intermediate layer element is calculated so that the evaluation function for evaluating the learning error becomes small. An example of the evaluation function in this case is shown in Expression 1 below.

[0031]

[Equation 1] J = 1/2 · (o−t) ²

In the equation 1, J: evaluation function,
o: neuro output, t: teacher value (learning target value).

The fourth step A4 is a step of calculating a weight correction amount for the neural network structure shown in FIG. By the calculation of the third step A3, the correction amount is calculated for the combination whose weight is originally 0, and the correction amount Δw is added to the original weight (w = 0), so that the weight is reconstructed. Sometimes. In order to prevent this, in the fourth step A4, the correction amount of such weight is forcibly set to zero.

The fifth step A5 is a weight correction process. The weight between the input layer element and the intermediate layer element is modified according to the final modification amount calculated through the third step and the fourth step. The correction amount of weight is Δw _ij , and the weight is w
_{If ij is} the learning coefficient and α is the learning coefficient, the weight can be corrected by the equation 2.

[0035]

(2) w _ij = w _ij + αΔw _ij

The process from the third step A3 onward is repeated until the learning error is equal to or less than the specified value and it is confirmed that the learning is completed (step A6). Here, the learning end can be determined by whether or not the error with respect to the evaluation function or all the learning data is equal to or less than a specified value, or whether or not the number of times of learning reaches a predetermined number.

In the learning method described above, the loosely coupled portion 12 and the fully coupled portion 11 in FIG. As a result, the loosely coupled part 1 is originally
The portion (input layer element) that should be connected to the intermediate layer element 2 is coupled to the total coupling portion 11, or the loose coupling portion 1
There are cases where the connection of 2 is small and the connection of all connected portions 11 becomes too large, which makes it difficult to analyze the structure of the neural network.

Japanese Patent Application No. 2000-proposed in view of the above points
The learning method described in 230665 will be described below. That is, as shown in FIG. 3, as a first step B1, learning is independently performed on one loosely coupled portion for all loosely coupled portions. Next, as a second step B2, all loosely coupled parts that have been learned in the first step B1 are combined to generate a first network.

Then, as a third step B3, the first network generated in the second step B2 is learned to generate a second network. Then, as a fourth step B4, all the connected parts are combined with the second network generated by the third step B3 to generate a third network, and as a fifth step B5, a fourth step B4.
Learning is performed on the third network generated in 4 to reduce a learning error, and a final neural network is generated.

FIG. 4 shows the neural network which has undergone the first step B1, and 12A, 12B and 12C are loosely coupled portions. These loosely coupled parts 12
A, 12B, and 12C are defined so that input factors having strong association are included in one loosely coupled portion, and in the illustrated example, inputs 1 and 2 having strong association are included in one loosely coupled portion 12A. . The loosely coupled portions 12A, 12B, and 12C each give predetermined learning data and each independently learns, and the relationship between each input factor and the output is strongly learned.
Here, since the loosely coupled portions 12A, 12B, and 12C are relatively small-scale networks, the learning speed is fast and the risk of falling into a local solution is small.

FIG. 5 shows a first network 12D formed by combining the loosely coupled portions 12A, 12B and 12C of FIG. 4 into one by the second step B2. In addition,
If this network is learned by a third step B3 described later, the structure of FIG.
It can be said that it also shows the network.

In this case, the neural network does not operate normally by simply connecting a plurality of loosely coupled parts. For example, it is assumed that there are three loosely coupled portions and the range of output values of the learning data is 0.1 to 0.9.

Assuming that each loosely coupled part can be satisfactorily learned in the first step B1, the sigmoid function of FIG. 6 showing the input / output characteristic (the input / output characteristic of each element (neuron) constituting the neural network is usually According to (the input / output characteristic of the sigmoid function is used), the range of the input value of each loosely coupled part is -2.2 to +2.2. In this case, if the three loosely coupled parts are simply combined, the range of input values will be -6.6 to +6.6, and as a result, the range of output values will also change to 0 to 1. In order to avoid this, it is necessary to keep the range of input values at -2.2 to +2.2 even after combining a plurality of loosely coupled parts.

There are various conceivable methods of connection,
Here, the following method is used. (1) The value of the coupling coefficient between the intermediate layer and the output layer is divided by the number n of loosely coupled portions. (2) The loosely coupled portions are joined together. (3) Recall all patterns, and if the output value is not within the predetermined range, modify the coupling coefficient so that it falls within that range. For example, the output value range is 0.1-0.9
If the output value range is 0.4 to 0.6, the input value becomes -0.4 to +0.4. To be These correspondences are determined by the input / output characteristics of the sigmoid function shown in FIG. However, if this error is small, the correction can be made only by learning in the next third step B3.

The first network generated by the second step B2 in FIG. 3 (network 12D in FIG. 5)
Corresponds to one in which a plurality of loosely coupled parts are mechanically coupled, so that the learning error is relatively large. Therefore, in the third step B3, in order to adjust the balance of each loosely coupled portion, the first network is made to learn again to generate the second network. In addition, in the fifth step B5, learning is performed again in order to further reduce the learning error.

The second network generated in the third step B3 is composed of only loosely coupled portions 12A, 12B and 12C. That is, there is a limit to the reduction of the error because the interaction between the inputs is missing. Therefore, in the fourth step B4, the total connection part 11 reflecting the interaction between the inputs is added to form the third network (neural network in FIG. 1), and the fifth network is further added.
In step B5, learning is performed again for this third network. As a result, a learning error is finally reduced, and a neural network that is easy to analyze is generated.

Next, an embodiment of the present invention will be described by taking the case of predicting the heat load of a plant as an example. (1) Construction of Prediction Model First, in order to predict a heat load having non-linearity, FIG.
Build a prediction model as shown in. This prediction model is a modification of the prediction model of FIG. 1 and is a hierarchical neural network composed of fully connected parts 11 and loosely connected parts 12 as shown in FIG.

As the input factors input to the prediction model during learning, factors related to the heat load, which is the output, such as hourly temperature, maximum temperature, minimum temperature, minimum humidity, weather, solar radiation amount, etc. Factors related to meteorological conditions, distinction between days of the week and weekdays / holidays (including Saturdays on weekdays / holidays), presence / absence and types of events, plant operating status, plant components (air conditioners, etc.) All or part of the factors that roughly determine the heat load pattern, such as the operation pattern, are used. Actual values are used for these input data. Further, the actual value of the heat load is used for the output, the input factor and the output are given to the neural network and the learning is performed by the learning method described above, and the prediction model is constructed.

In the example of FIG. 7, the input factors given to the input layer are the maximum temperature, the minimum humidity and the day of the week, and the heat load of the next day is predicted and output. Also, the intermediate layer is an element that outputs only the input layer element to which the maximum temperature is input and outputs the temperature component, and an element that is connected to only the input layer element to which the minimum humidity is input and outputs the humidity component, An element that outputs only the day component by combining only the input layer elements that input the day of the week,
It is composed of elements that are coupled to all the input layer elements and output an interaction component. Here, the output layer element is single.

When predicting the hourly heat load of the next day, the output layer elements of the neural network may be assigned to the heat load of each time, and a neural network having 24 output layer elements may be used. in this case,
At the time of learning, for example, the temperature and humidity at fixed time intervals are input to the input layer element, the actual load value at 0 o'clock is given to the first output layer element for learning, and at the time of actual prediction, prediction of the input factor is performed. A value may be given to the input layer element to output the predicted heat load value at 0:00 from the first output layer element. In this way, if the number of output layer elements is 24, the daily heat load pattern can be modeled at all 24 points, so a neural network is constructed every hour (use 24 prediction models).
The number of prediction models can be reduced compared to the method, and the load on the computer can be reduced.

(2) Prediction execution Using the prediction model constructed in (1) above, the prediction calculation of the heat load of the next day is executed. At this time, the forecast value of the next day is used as the input data relating to the weather conditions such as the maximum temperature and the minimum humidity as the input factors. Further, when inputting the operating state of the plant, the operation pattern of the plant constituent equipment, etc., those planned values may be used.

(3) Explanation of Reason for Prediction The reason why the prediction result is obtained will be described. To explain the reason for prediction, it is general to use partial regression coefficient in multiple regression model and input / output sensitivity in neural network.
In the present embodiment, by using the neural network having the fully connected portion 11 and the loosely connected portion 12, it is possible to explain the reason for the prediction that the predicted value was obtained. In the example of FIG. 2, the heat load prediction value is a component related to the maximum temperature, a component related to the minimum humidity, and a component related to the day of the week, and an interaction component in which the maximum temperature, the minimum humidity, and the day of the week are related to each other. It can be divided into and presented.

In explaining the reason for prediction, a method of analyzing the influence of the input factor on the output from the amount of information transmitted from the intermediate layer element to the output layer element when arbitrary data is input will be described below. This analysis method is, for example, Japanese Patent Application 20
It is described in claims 11 and 12 of 00-166528.

First, an analysis method described in claim 11 of Japanese Patent Application No. 2000-166528 will be described. Figure 8
In the neural network shown in (1), the information transmitted from the intermediate layer to the output layer is the sum of products of the intermediate layer output O and the weight v. That is, this information is represented by Equation 3.

[0055]

[Equation 3]

Here, it can be said that the intermediate layer element that outputs the largest value of | _vi O _i | has the strongest influence on the output, and the influence of the input factor coupled to the intermediate layer element is also strong. . For example, referring to FIG.
When v ₁ O ₁ | is the largest, the influence of input 1 is strong,
It can be said that the interaction between the inputs 1 and 2 is strong when | v ₂ O ₂ | by the intermediate layer element 2 is the largest. Therefore, by detecting the amount of information transmitted from each intermediate layer element to the output layer element, it is possible to know the strength of the coupling between the intermediate layer element, the output layer element and the input layer element. We can analyze the effect of input factors on the output in the network.

In the example of FIG. 7, for example, when the product of the output of the temperature component of the intermediate layer and the weight between the component and the output layer is larger than the other components, the predicted value ( Since the influence of the temperature component (that is, the maximum temperature) on the heat load of the next day is considered to be large, it is possible to explain that the maximum temperature has a large influence on the reason that the predicted value was obtained. In addition, when the product of the output of the interaction component of the intermediate layer and the weight between the component and the output layer is larger than the other components, the interaction component (that is, the maximum temperature, Since it is considered that the influence of the minimum humidity, day of the week) is large, it is not possible to unambiguously determine the reason why the predicted value was obtained, and it is explained that it is due to the interaction between the maximum temperature, the minimum humidity and the day of the week. It will be possible.

Next, an analysis method described in claim 12 of Japanese Patent Application No. 2000-166528 will be described. This analysis method is a method of analyzing the influence of an input factor on the output from the correlation between the input data and the information transmitted from the intermediate layer element to the output layer element when arbitrary data is input.

That is, in the neural network of FIG. 8, the values of input 1 and input 2 are always the same (input 1, input 2) = (0,0), (0.2,0.2), ...
,, (0.8, 0.8), (1, 1) in a plurality of 0.2 increments are input, and at that time, the amount of information transmitted from the intermediate layer elements 1 to 3 to the output layer element is Suppose that the result is as shown in FIG. That is, regarding the intermediate layer element 1, the amount of information transmitted to the output layer element is gradually increased, the intermediate layer element 2 is almost constant, and the intermediate layer element 3 is gradually decreased. From this, the following can be understood.

Intermediate layer element 1, which is coupled only to input 1, has a positive correlation, ie input 1 has a positive correlation with the output. The intermediate layer element 2 has almost no effect on the output. That is, there is almost no interaction between the inputs 1 and 2 coupled to the intermediate layer element 2. The intermediate layer element 3, which is coupled only to the input 2, has a negative correlation, ie the input 2 has a negative correlation with the output. The following items are derived from the above items. To increase the output, a large value may be input to input 1 and a small value may be input to input 2.

As described above, unknown data (unlearned inputs x ₁ , x) is detected by detecting the behavior of the intermediate layer element with respect to the input value (the amount of information transmitted from each intermediate layer element to the output layer). ₂ ), the output values can be easily estimated because the input factors and the functions of the intermediate layer elements are known.

According to the example of FIG. 7, the input factors such as the maximum temperature and the minimum humidity have the correlation as shown in FIG. 9 regarding the information transmitted from the temperature component and the humidity component of the intermediate layer to the output layer. If you know in advance, even if you have given data such as maximum temperature and minimum humidity that you have not learned in the past, you can predict to some extent what the heat load of the next day will be, The reason why the predicted value is obtained can also be explained.

As described above, according to this embodiment, while making the prediction model non-linear and explaining the prediction reason,
It is possible to accurately predict a load having a non-linearity such as a heat load. Needless to say, the present invention is also applicable to prediction of power load, air load, etc. of plants other than heat load.

2. Next, an embodiment relating to the steady-state plant simulator described in claim 2 will be described. here,
The steady-state plant simulator calculates the state of the plant, such as the connection status of plant components, linear or nonlinear input / output characteristics, and operation patterns that change according to the time of day (input / output characteristics change due to changes in the operation patterns). Using the database that stores the necessary characteristics, the amount of fuel injected into the plant components for each control time with a fixed time interval, and the start / stop state of the plant components are planned values from the optimization unit described later. At a given time, the steady input / output state of each component device for each control time (here, the transient input / output state is not considered, only the steady state when the output of the device is stable is considered). calculate. Then, the output of a certain component device is sequentially calculated as the input of the component device of the next stage to finally simulate the steady input / output state of the entire plant. The input and output states of the plant components simulated in this way are used to calculate the objective function for determining the optimum operation of the plant.
It will be sent to the optimization unit described later.

For example, as shown in FIG. 10, a gas turbine 21, an exhaust gas boiler 22, a steam turbine 23, a compressor 24, a boiler 25, an absorption refrigerating machine 26, a gas-fired refrigerating machine 27, and a heat storage layer 28 are used. It is assumed that an energy plant for obtaining a load, an electric load, a steam load, and a heat (air conditioning) load is assumed. In the steady-state plant simulator, the above-mentioned configuration states, input / output characteristics, and operation patterns of each of the devices 21 to 28 are given, and further, fuel injection into the gas turbine 21, the boiler 25, and the gas-fired refrigerator 27 at each control time is performed. When the amount and the start / stop state of each device 21 to 28 are given, for example, the gas turbine 21
The amount of exhaust gas output from the boiler and the amount of steam output from the boiler 25 are calculated. Then, these are used as inputs to the exhaust gas boiler 22 and the absorption chiller 26 in the next stage, and the outputs of the devices 22, 26, ... Simulate the state.

At this time, the non-linear input / output characteristic of each plant constituent device is expressed by the following method. General nonlinear equation neural network whose coefficients are adjusted by the polynomial response surface method up to the second order whose coefficients are adjusted by the least squares method

The response surface method is, for example, Andr
e I. Khuri, and John A. Cornell, Response Surfaces
Designs and Analyzes, Second Edition, Marcel Dekke
r Inc., 1996 etc. The neural network is also DE Rumelhart and JL McCelland.
(Eds.), Parallel Distributed Processing, MIT Pres
s, 1986., etc., including all neural networks that automatically generate input-output mapping functions for neural networks, such as the model using a multilayer perceptron. In some cases, the structured neural network described as the embodiment of the invention of claim 1 may be used. In that case,
The advantage that the reason for the input / output relationship of each plant component can be easily explained is obtained.

Here, like the Taguchi method, the response surface method is a method that has been put to practical use in the United States, particularly in the field of quality engineering in which product development is performed in consideration of product variations. The response surface method is a method for obtaining a limited number of responses to an appropriate combination of actual measurement values and creating an approximate function of the response predicted from the predictor variable, using the approximate function created by this response surface method. Input / output characteristics of each device can be expressed. Here, as a method of obtaining the response surface, software simulation may be used, an experiment based on hardware actually manufactured, or a plant actual measurement value may be used.

Although the function form of the approximate function is not limited, the following form can be considered as the approximate function. This is a case of using the linear function of the linear function formula 4, and corresponds to the multiple regression used in statistics.

[0070]

## EQU4 ## y = β ₀ + β ₁ x ₁ + β ₂ x ₂ + ...

Here, β _i : coefficient x _i : prediction variable (design variable).

Quadratic polynomial This is a case of using the quadratic polynomial of Expression 5.

[0073]

Y = β ₀ + β ₁ x ₁ + β ₂ x ₂ + ... + β ₁₁
x ₁ ² + β _{22 22} x ₂ ² + ... + β ₁₂ x ₁ x ₂ + β ₁₃
x ₁ x ₃ + ...

Similarly, it is possible to use a polynomial of third order or higher. In addition, an orthogonal polynomial may be used for a polynomial of second order or higher in order to make an orthogonal experiment as shown below.

The approximate function is created by selecting an experimental point and using the experimental point.

There are the following methods for selecting the experimental points. This is a case of using an orthogonal table used in the orthogonal plan Taguchi method, and is usually used for approximation by a first-order polynomial (linear multiple regression). However, in a polynomial regression model of second order or higher, a multilevel orthogonal table such as L27 is used. Further, at this time, by using the orthogonal function as the approximate function, it becomes possible to independently select the coefficients required for creating the approximate function, which is effective.

Central Composite Desig
n: CCD) CCD is a method in which the points shown in FIG. 11 are used as experimental points in the case of, for example, two variables, and is known to be effective in approximation by a quadratic polynomial.

As a D-optimal level computer-supported experiment plan, p experimental points that maximize the D-optimality index expressed by the following equation 6 are used.

[0079]

[Equation 6]

Where X: matrix of experimental points (n × k) n: total number of sets of experimental points k: number of design variables p: number of experimental points required for approximation Is.

As shown below, there is a close relationship between the selection of experimental points and the approximation function used. -Selection of experimental points by orthogonal design and approximation by orthogonal function have the advantages that they are easy to calculate and can be applied to higher-order polynomials. However, for example, D _eff of the L27 orthogonal table + quadratic orthogonal function is 0.462, which is not a very good approximation. -CCD is effective only for second-order polynomials. The D _eff of CCD + quadratic polynomial is 0.996, which is a very good approximation. -In cases other than the above, the D-optimal criterion is required.

Since polynomials of second or higher degree can also be converted into first-order polynomials by variable conversion, an approximate function is created using the first-order polynomial, that is, the method used for linear multiple regression in the field of statistics. can do. (1) How to obtain the coefficient of the model Each coefficient β _i of the model of Equations 4 and 5 can be obtained by the least squares method as used in multiple regression. (2) Evaluation of Appropriateness of Model Whether or not the model is suitable can be evaluated by the coefficient of determination (index indicating suitability of regression model) R ² shown in Expression 7. In Equation 7, SSR is the regression sum of squares, S _yy
Is the variation around the mean value of the response y. When the regression equation completely matches the response, SSR and _Syy are completely the same, and the coefficient of determination R ² is 1. This coefficient of determination R ² is 0
A value of ˜1 is taken, and generally, the closer to 1, the more appropriate the regression model can be. The effectiveness of each coefficient of the model can be confirmed by t-test.

[0083]

[Equation 7]

As described above, the steady-state plant simulator according to this embodiment expresses the linear or nonlinear input / output characteristics of each device as shown in FIG. 10 by the least square method, response surface method, neural network, or the like. The input / output characteristics, the operation pattern of each device according to the time zone (including characteristics such as the time until the steady state is reached after starting), and the like are provided in advance as a database. And
The input / output state of each device (input / output gas, steam, water, etc.) from the fuel injection side toward the end of the plant using the fuel injection amount for each control time and the start / stop state of each device as input information. The steady-state input / output state of the whole plant is simulated by sequentially calculating the amount of these and the input / output times of these.

3. Next, an embodiment of the optimum operation method of the plant described in claims 3, 4 and 5 will be described. In this embodiment, the optimization unit passes the start / stop state of each plant constituent device for each control time and the fuel injection amount to the steady plant simulator as a plan, and the steady plant simulator side uses these input information. The input / output state of each device obtained by the above-mentioned operation is returned to the optimization unit, and the optimization unit uses the input / output state of each device to determine a predetermined objective function (minimize plant operation cost or reduce gas emissions). Various optimization methods are used to search for the starting / stopping state of each device and the fuel injection amount that satisfy the above conditions (minimization, etc.).

FIG. 12 shows a conceptual diagram when the optimum operation method of the plant is determined using the stationary plant simulator. The optimizing unit 30 and the stationary plant simulator 40 in FIG. 12 are both realized by computer hardware and software. First, when the predicted values of various plant loads are obtained by the invention according to claim 1, the energy supply (supply / demand balance) with respect to the predicted load value of each load type is satisfied, and the constraints on the characteristics of each device are considered. (These are constraints),
The objective function is to minimize the plant operating cost (including fuel cost) and the gas emission amount (eg CO ₂ emission amount) for a predetermined period, for example, one day (the supply-demand imbalance amount as a penalty and deviation of the device characteristic constraint). Set the quantity), and formulate the plant optimization problem as a nonlinear mixed integer programming problem.

Then, an initial value of the starting / stopping state (discrete value of starting or stopping) and the fuel injection amount (continuous value) of each plant constituent device for each control time is generated, and is generated for the steady plant simulator. , By sequentially giving the start / stop state of the plant component equipment in each search process and the fuel injection amount, while obtaining the input / output state of each plant component equipment from the steady plant simulator as an output,
Optimal operation of the plant by deciding the start / stop state and fuel injection amount of each device that is close to the global optimum solution while sequentially correcting the planned values of start / stop state and fuel injection amount by the optimization method. Decide how.

The formulation of optimization is shown below. (1) State variables State variables are the following plant quantities. Start / stop status of each plant component device (eg gas turbine, boiler, refrigerator, cogeneration, etc.) for each control time Fuel injection amount for each control time (fuel injection amount for each device)

(2) Objective function The objective function is assumed to have the following terms. Minimize plant operation cost (set to f ₁ ) Minimize CO ₂ emissions (set to f ₂ ) Penalty (supply / demand balance imbalance amount, deviation of equipment characteristic constraint) (set to f ₃ ) Actually The following objective function consisting of each term is used. f = w ₁ f ₁ + w ₂ f ₂ + w ₃ f ₃ where w _i is a weight for each term and is set arbitrarily.

(3) Constraint conditions Supply / demand balance for each load type This is a constraint concerning the balance between supply and demand for each load type such as electric power system, thermal system, and air system. The value is taken into account. Restrictions on the characteristics of each device These are the restrictions on the input / output limits and start / stop times of each plant component.

(4) Solving Algorithm The optimum solution of the starting / stopping state and the fuel injection amount is searched while using the calculation result by the steady plant simulator (the input / output state of each plant constituent device at each control time). A modern heuristic optimization method is used as the optimization method. Specifically, Particle Swarm Optim
ization (hereinafter referred to as PSO) and its improvement method, genetic algorithm (hereinafter referred to as GA) and its improvement method,
Tabu Search (TS) and its improvement method, An
t Colony Optimization (hereinafter referred to as ACO) and its improvement method are used. Here, PSO and its improved method are J. Kennedy and R. Eberhart, SwarmIntellige.
nce, Morgan Kaufmann Publishers, 2001, PSO's Gbest and Lbest models based on the herd theory developed by Eberhart et al., or Hybrid Particle Swar developed by Angeline.
Refers to techniques that include various variations of PSO, such as m Optimization. What is GA and its improved method?
ldberg, Genetic Algorithms in Search, Optimizatio
n, and Machine Learning, Addison-Wesley, 1989, Simple Genetic Algorithm (hereinafter, S
GA) and its improved method, and TS and its improved method are F. Glover, "Tabu Search Part I", ORSA.
Means the tabu search and its improved methods described in Journal of Computing, Vol. 1, NO. 3, Summer 1989,
ACO and its improved method are A. Colorni, M. Dorigo,
and V. Maniezzo, "Distributed Optimization by Ant
Colonies ", Proc. OfFirst European Conference on A
rtificial Life, pp.134-142, Cambridge, MA: MIT Pre
Ant Colony Optimization (A
CO) and its improvement method.

In the following, an algorithm using PSO, SGA and TS is shown as an example. (1) PSO State Representation In PSO, the state is represented by the amount of fuel injected into each plant component for each control time and the data string consisting of the start / stop state of each component whose start / stop state can be changed. . Fukuyama et al., "Participity Considering Voltage Reliability," describes a method that enables the solution algorithm PSO to be applied to nonlinear mixed integer programming problems.
Study on Voltage Reactive Power Control Method by le Swarm Optimization ”IEEJ Transactions on B 119, No. 12 (1999, 1
The method proposed in (February) etc. is used.

PSO is one of the modern heuristic methods developed through the simulation of a simplified social model, and was developed by expressing the movement of a flock of birds in a two-dimensional space of continuous variables. In the PSO, the position (state amount) of each agent (the above-mentioned bird) is expressed by x and y coordinates, and the velocity (vector) corresponding to the change in the position (state amount) is v _x (velocity in the x direction). , V _y
It is expressed by (velocity in the y direction). The position of each agent at the next time can be updated from these position and velocity information. Based on this concept, considering that the entire flock of birds behaves to optimize some objective function, the following optimization is possible.

That is, each agent has the individual best value (pbest) of the objective function up to that point in each search,
The x and y coordinates indicating the position (state quantity) are remembered.
Further, each agent shares the best pbest among the groups, that is, the best value (gbest) information of the objective function of the group up to that point. Each agent, x current self, y coordinates and velocity _v x, _{v y,} and, depending on the distance between pbest and gbest, pbest, gbest
Attempts to change the direction to the existing position. This changing behavior is expressed by modifying the speed. Using the current speed and pbest and gbest, the speed of each agent is modified by Eq.

[0095]

[Equation 8]

[0096] In Equation 8, _{v i:} speed of agent i, rand (): uniform random numbers of up to _{0~1, s} ^{i k:} search the k-th position of agent i (search point), pbest _i: agent i Pbest, w: a weighting function for the speed of the agent, c _i : a weighting coefficient for each term.

By using the above equation 8, a velocity that stochastically approaches the best solution of each agent and the best solution of the group is obtained, and the current position (search point) of each agent is calculated by the equation. Correct by 9. Here, according to the present invention, each agent corresponds to a plant constituent device, and the position s _i of each agent in Expressions 8 and 9 is the start / stop state (discrete value) or fuel injection of each plant constituent device. Equivalent to the quantity (continuous value),
The velocity v _{i of} each agent corresponds to their change.

[0098]

S _i ^{k + 1} = s _i ^k + v _i ^{k + 1}

PSO is a genetic algorithm (Genetic Al
gorithm: In the same way as GA), etc., it is a multi-point search with multiple search points, and by globally changing each search point by using pbest of each search point and gbest of the group, This is a method of obtaining a solution (best solution). In addition, the global search that tries to maintain the speed so far (the first side of the right side of Equation 8)
This is a search method having a mechanism for performing a well-balanced search between the local search term and the local search that tries to approach them using pbest and gbest (the second and third terms on the right side of Expression 8, respectively). Furthermore, P
In SO, it is necessary to evaluate the objective function value at each step of the search, but the number of evaluations does not depend on the scale of the problem, and only the number of agents is required, so application to large-scale problems such as optimum plant operation Is possible.

The overall algorithm of the plant optimum operation method when the PSO of the gbest model is used is shown below. Note that FIG. 13 is a flowchart thereof.

Step.1 Data Input (C1 in FIG. 13) Input of Load Prediction Value-Input the load prediction value for each control time for each load type. This predicted load value is a predicted value of the power load, heat load, air load, etc. of the plant obtained by the invention of claim 1. Input of plant information (these are stored in advance as a database) ・ Input / output characteristics (linear, non-linear) for each plant component
The operation pattern for each control time, the connection relationship between each component and the corresponding characteristic formula, the fuel cost for each fuel type, and the like are input as plant information. Input information related to Particle Swarm Optimization ・ Enter the number of agents, each optimization parameter value, and the maximum number of searches.

Step.2 Initial Value Generation (C2 in FIG. 13) The following values for each control time are randomly calculated for each agent.・ Starting / stopping state (discrete value) for each plant component ・ Fuel injection amount of fuel injecting device among the above plant components (continuous value) Calculation of plant state for each agent and calculation of evaluation value The input / output state of each plant constituent device for each control time is obtained by the steady-state plant simulator according to claim 2 using the started / stopped state and the fuel injection amount of each plant constituent device. -Furthermore, an evaluation value is calculated for each agent using the objective function and constraint conditions shown in FIG. Here, the evaluation value is the sum of plant operation costs and CO ₂ emissions. The evaluation value of each agent calculated in the initial setting of pbest and gbest is set as the current pbest value of each agent. -The best value (minimum value) of the above pbest is gbest. As a result, the initial values of pbest and gbest are set.

Step.3 Position of each agent (search point)
(C3 in FIG. 13) The position (search point) of each agent is corrected with respect to the discrete value and the continuous value by the above-described Expressions 8 and 9. FIG. 14 shows a conceptual diagram of the correction at this time. In FIG. 14, v ^{k + 1} corresponds to the second term on the right side of Expression 9.

Step.4 Evaluation of each agent (C4 in FIG. 13) Calculation of plant state for each agent and calculation of evaluation value-Start / stop state and fuel injection amount for each plant constituent device calculated in Step.2. Using, the input / output state of each plant constituent device for each control time is obtained by a steady plant simulator. -Evaluation value is calculated for each agent using the objective function and the constraint condition of FIG. If the evaluation value for each agent calculated by modifying pbest and gbest is better than the current pbest value for each agent, the evaluation value is set as the pbest for each agent. If the best value of the above pbest is better than the current gbest value, the best value is set as gbest. Through these processes, the optimum solution for the starting / stopping state of each plant component and the fuel injection amount is obtained.

Step.5 Check the end condition (C in FIG. 13)
5) When the number of searches reaches the maximum number of inputs, the process ends. If not, return to Step.3.

(2) SGA State Representation In SGA, since all discrete values have to be discrete values, the fuel injection amount, which is originally a continuous value, uses discrete values at the discrete minimum fuel cost increments. The minimum value to the maximum value of the digitized numbers are associated with integers. It is internally expressed as an integer, and the actual fuel injection amount that is a continuous value is converted from the correspondence table with the integer. On the other hand, the starting / stopping state of the device for each control time, which is originally a discrete value, is expressed by an integer with 1 being start and 0 being stop. That is, the fuel injection amount and the start / stop state are all represented by integers to create a correspondence table, and a device for examining the fuel injection amount (fuel injection device)
The state is represented by a gene having the total length of the number of devices and the number of devices considering the start / stop state. Therefore, each locus is the fuel injection amount or the start / stop state of the device.

Solving Algorithm Step.1 Preconditions and Initial Condition Settings (E1 in FIG. 15) Input the load prediction value obtained by the load prediction method of claim 1.・ Set the number of strings, crossing probability, mutation probability, and maximum number of generations.

Step.2 Generation of initial value (E2 in FIG. 15) -For the loci of each string, an integer value is randomly selected and used as the initial value. -Set the current number of generations to 1.

Step.3 Evaluation and selection of each string (Fig. 1
E3 of 5) -The fuel injection amount or the start / stop state of the device is determined using the integer value of the locus of each string and the correspondence table (the correspondence table in the above-mentioned state expression). -Calculate the evaluation value for each gene using the plant model and the predicted load value. Here, the evaluation value is the sum of plant operation costs and CO ₂ emissions. -Using the evaluation value of each string, select the string by roulette wheel selection.

Step.4 String operation (E4 in FIG. 15) -By performing crossover and mutation on the string set using the crossover probability and mutation probability, the above evaluation value is optimized. Determine the string. Then, using the integer value of the loci of this string and the correspondence table, the optimal solution for the fuel injection amount and the starting / stopping state of the device is determined.

Step.5 End determination (E5 and E6 in FIG. 15) When the generation reaches the preset maximum number of generations set value, the process ends.・ If the set value is not reached, add 1 to the number of generations and Ste
Return to p.3.

(2) TS In the state expression TS, all of the fuel injection amounts, which are originally continuous amounts, should be discretized by the minimum fuel cost step, because all of them must be discrete values. Correspond to the integer from the minimum value to the maximum value of the discretized numerical value. It is expressed internally as an integer, and the actual fuel injection amount that is a continuous amount is converted from the correspondence table with the integer. In addition, the starting / stopping state of the device for each control time, which is originally a discrete value, is expressed by an integer with 1 as starting and 0 as stopping. That is, like SGA, the correspondence table is created by expressing the fuel injection amount and the start / stop state with integers, and at the same time, the number of fuel injection devices and start / stop
It is represented by an integer array that has the total length of the number of devices for which the stop status is considered. Therefore, each element of the array is the fuel injection amount or the start / stop state of the device.

Solving Algorithm Step.1 Precondition and Initial Condition Setting (F1 in FIG. 16) Input the load prediction value obtained by the load prediction method of claim 1.・ Set taboo length and maximum number of searches.

Step.2 Generation of initial value (F2 in FIG. 16) -For each element of the array of state representation, an integer value is randomly selected to be the initial value.・ Enter the current status in the taboo list.・ Set the current search count to 1.

Step.3 Generation of adjacent state and determination of next state (F3 in FIG. 16) -For each element of the array of the current state, integer values +1 and -1 (in the case of upper and lower limit values, , Value is not generated) is generated as an adjacent state. For example, when the arrayed integer values are (2,3,4), +1 and -1 are added to each element (1,3,4), (3,3,4),
(2,2,4), (2,4,4), (2,3,3),
(2, 3, 5) becomes the adjacent state. -For the array of integer values of each adjacent state, determine the fuel injection amount or the start / stop state of the plant component equipment by referring to the correspondence table. -Calculate the evaluation value for each gene using the plant model and the predicted load value. Here, the evaluation value is the sum of plant operation costs and CO ₂ emissions. Among the adjacent states, the adjacent state having the best evaluation value is determined by processing the next state that is not taboo and has the best evaluation. Then, using the integer value of the adjacent state and the correspondence table, the optimum solution of the fuel injection amount and the start / stop state of the device is determined.

Step.4 End judgment (F4 and F5 in FIG. 16) When the number of searches reaches the preset maximum number of searches,
finish. -If the set value is not reached, add 1 to the current search count and return to Step 3.

Incidentally, it is desirable to correct (re-plan) the operation plan of the day by utilizing the above-mentioned optimum operation method. That is, as described above, an objective function (f = w ₁ f ₁ +) is used by using a predetermined state variable (start / stop state of each plant constituent device at each control time, fuel injection amount at each control time).
w ₂ f ₂ + w ₃ f ₃ ) is minimized by taking into account various constraint conditions (supply / demand balance for each load type, constraint on characteristics of each device), and starting / stopping the plant component devices by the optimization method. As a result of determining the state and the fuel injection amount, FIG.
It is assumed that the operation plan of the day as shown in 7 (a) is obtained on the previous day (referred to as the previous day's plan in the sense that it was planned on the previous day).

When the load-demand balance for each load type among the constraint conditions considered in the previous day's plan uses the load predicted value on the previous day, and the load value on the day shifts from the predicted value. There are many. Therefore, on the day, it is necessary to consider the supply and demand balance based on the new load forecast value as a constraint condition. In other words, it is necessary to correct the previous day's plan to create the current day's corrected plan.

Therefore, assuming that the present time is t o'clock (after elapse of t time) on the current day as shown in FIG. 17 (b), the remaining time (24-t hours) on the current day is considered as the consideration time (correction plan). The target operating time) is the operating status of the plant constituent equipment at this point or the operating status of the equipment that has already been commanded and cannot be changed, and is used as the fixed initial state of the plant. In the above, the current day correction plan may be created by the optimal operation method while considering the supply and demand balance based on the load forecast value of the remaining time (24-t hours) of the day as a constraint condition. The above idea corresponds to the invention described in claim 5.

4. Next, an embodiment of the optimum design method for a plant according to claim 6 will be described. This optimum design method formulates the problem of determining the optimum capacity of each plant component in the plant design stage as a combination optimization problem that determines the optimum value from the settable device capacity values (discrete values). At the same time, the load pattern of the representative day of each month of the year is set empirically using actual values, etc.
Alternatively, the optimum operation method described in 4 is used to obtain the capacity of each plant constituent device that optimizes the evaluation value regarding the annual plant operation cost and CO ₂ emission amount. This optimal design method is realized by the following procedure.

Step.1 Precondition and initial condition setting (Fig. 1
D1 of 8) -Determine the system configuration of the plant (connection status with the plant component equipment). A load value is set as a load value for each control time during a predetermined period, for example, a representative day of each month of the year. If the target plant is an existing plant, the load pattern is set using the past actual values. Also, if it is a new plant,
Set by referring to the actual load value of the existing plant.・ Set the initial value of the capacity of each plant component as a planned value.

Step.2 By using the value set in Step.1, the plant optimum operation method according to claim 3 or 4 determines the start / stop state of each device and the fuel injection amount for each representative day, and performs optimum operation. The method is determined (D2 in FIG. 18).

Step.3 Evaluation value related to plant operation cost and CO ₂ emissions throughout the year (sum of evaluation values on each representative day)
Until the best value is obtained, the capacities of the components of each plant are sequentially changed using the combination optimization method (D3 in FIG. 18).

Here, as the combination optimization method, a genetic algorithm and its improved method, taboo search and its improved method, Ant Colony Optimization and its improved method, etc. are used. What is the above genetic algorithm? E. Gol
dberg, Genetic Algorithms in Search, Optimization,
and Machine Learning, Addison-Wesley, 1989, Simple Genetic Algorithm (hereinafter, SGA
And its improved method, Tabu Search is F. Glover,
"Tabu Search Part I", ORSA Journal of Computing, V
ol. 1, NO. 3, Tabu Search (TS) and its improvement method described in Summer 1989, Ant Colony
What is Optimization? A. Colorni, M. Dorigo, and V. Mani
ezzo, "Distributed Optimization by Ant Colonies",
Proc. Of First European Conference on Artificial Li
fe, pp.134-142, Cambridge, MA: Ant Colony Optimization described in MIT Press 1991.
Means) and its improved method. Hereinafter, an algorithm using SGA and TS will be described as an example with reference to FIGS. 15 and 16 described above.

(1) SGA State Representation In SGA, the gene length is as many as the number of plant constituent devices for which the optimum capacity is obtained, and each locus is an integer value expressing the selectable capacity of each constituent device as an integer from the smallest. To do. Therefore, the actual capacity is obtained from the correspondence table between each integer value and the actual capacity. For example, 1 is 100kVA, 2 is 150k
VA, 3 corresponds to 200 kVA, etc.

Solving Algorithm Step.1 Setting Preconditions and Initial Conditions (E1 in FIG. 15) -Determine the system configuration of the plant (the mutual connection state of the components of the plant).・ Set the load value for each control time on the representative day of each month of the year. If it is an existing plant, it will be set using the actual values obtained so far. If it is a new plant, it is set with reference to the actual load value of the existing plant.・ Set the initial value of the capacity of each plant component as a planned value.・ Set the number of strings, crossing probability, mutation probability, and maximum number of generations.

Step.2 Generation of initial value (E2 in FIG. 15) -For the loci of each string, an integer value is randomly selected and used as the initial value. -Set the current number of generations to 1.

Step.3 Evaluation and selection of each string (Fig. 1
E3 of 5) -Set the integer value of the locus of each string using the correspondence table and determine the capacity.・ By determining the start / stop status of each device and the fuel injection amount by using the plant model in which the capacity of each component is determined and the load value for each control time set on the first representative day of each month The optimum operation method is obtained, and the sum of the evaluation values regarding the plant operation cost and the CO ₂ emission amount with respect to the representative day of each month throughout the year is used as the evaluation value of each string. -Using the evaluation value of each string, select the string by roulette wheel selection.

Step.4 String operation (E4 in FIG. 15) -The sum of evaluation values throughout the year by executing crossover and mutation using crossover probability and mutation probability for a string set Determine the capacity of each component so that

Step.5 Termination determination (E5 and E6 in FIG. 15) When the number of generations reaches the preset maximum number of generations set value, the process is terminated.・ If the set value is not reached, add 1 to the number of generations and Ste
Return to p.3.

(2) TS In the state expression TS, the array length is the number of plant constituent devices for which the optimum capacity is obtained, and each element of the array is an integer value that represents the selectable capacity of each constituent device as an integer from the smallest. And Therefore, the actual capacity is obtained from the correspondence table between each integer value and the capacity. For example, 1 is 100 kVA, 2 is 150 kVA, 3
Corresponds to 200 kVA, etc.

Solving Algorithm Step.1 Setting Preconditions and Initial Conditions (F1 in FIG. 16) -Determine the system configuration of the plant (the mutual connection state of the components of the plant).・ Set the load value for each control time on the representative day of each month of the year. If it is an existing plant, it will be set using the actual values obtained so far. If it is a new plant, it is set with reference to the actual load of the existing plant.・ Set the initial value of the capacity of each plant component as a planned value.・ Set taboo length and maximum number of searches.

Step.2 Generation of initial value (F2 in FIG. 16) -For each element of the array of state representation, an integer value is randomly selected to be the initial value.・ Enter the current status in the taboo list.・ Set the current search count to 1.

Step.3 Generation of adjacent state and determination of next state (F3 in FIG. 16) -For each element of the array of the current state, integer values of +1 and -1 (in the case of upper and lower limit values, , Value is not generated) is generated as an adjacent state. For example, when the arrayed integer values are (2,3,4), +1 and -1 are added to each element (1,3,4), (3,3,4),
(2,2,4), (2,4,4), (2,3,3),
(2, 3, 5) becomes the adjacent state. -The capacity of each component is determined by referring to the correspondence table for the array of integer values in each adjacent state. -Determine the start / stop status of each device and the fuel injection amount by using the plant model in which the capacity of each component device is determined and the load value for each control time that was initially set on the representative day of each month. Then, the optimum operation method is obtained by using the sum of evaluation values for plant operation costs and CO ₂ emissions for representative days of each month throughout the year as evaluation values for each adjacent state. -The optimum capacity of each component is determined so that the sum of the evaluation values throughout the year is the best by the process of setting the next evaluation state that is not taboo among the adjacent states.

Step.4 End judgment (F4 and F5 in FIG. 16) When the number of searches reaches the preset maximum number of searches,
finish. -If the set value is not reached, add 1 to the current search count and return to Step 3.

[0136]

As described above, according to the first aspect of the present invention, various loads of a plant having nonlinearity can be predicted with high accuracy, and efficient operation of plant constituent equipment and energy saving are possible. At the same time, it is possible to easily explain the reason for the prediction and dispel the anxiety of the operator.

According to the invention described in claim 2,
It is possible to realize a steady-state plant simulator that accurately simulates the behavior of the entire plant in consideration of the nonlinear input / output characteristics of plant components and complicated operation patterns.

Further, according to the invention described in claims 3 and 4, the optimum operation of the plant is formulated as a nonlinear mixed integer programming problem, and the output of the stationary plant simulator is used to obtain a plant configuration close to the global optimum solution. It is possible to determine the start / stop status of the device and the fuel injection amount. According to the invention of claim 5, it is possible to correct the operation plan in consideration of the difference between the predicted load value and the actual load value of the day, and it is possible to further optimize the operation of the plant. is there.

According to the invention described in claim 6, the optimum capacity of the plant constituent equipment for optimizing the evaluation value concerning the plant operation cost and the exhaust gas within a certain period by using the optimum operation method is combined and optimized. It can be easily determined by solving the problem.

[Brief description of drawings]

FIG. 1 is a diagram showing a general structure of a hierarchical neural network to which an embodiment of the invention of claim 1 is applied.

FIG. 2 is a flowchart showing a learning method of a neural network.

FIG. 3 is a flowchart showing a learning method of a neural network.

FIG. 4 is a diagram showing a network structure of a first step in FIG.

5 is an explanatory diagram of a first network in FIG.

FIG. 6 is a diagram showing an input / output relationship of a sigmoid function.

FIG. 7 is a diagram showing a specific structure of a hierarchical neural network to which the embodiment of the invention of claim 1 is applied.

FIG. 8 is a diagram showing a structure of a neural network for explaining a method of analyzing an influence of an input factor on an output.

FIG. 9 is a diagram showing a correlation between an input value and information transmitted to an output layer.

FIG. 10 is a configuration diagram showing an example of an energy plant.

FIG. 11 is a conceptual diagram showing experimental points of a central composite design in the response surface method.

FIG. 12 is a conceptual diagram when determining an optimum operation method of a plant using a plant simulator.

FIG. 13 is a flowchart showing an embodiment of the invention described in claim 4;

FIG. 14 is a conceptual diagram of correction in PSO.

FIG. 15 is a flowchart showing a solution finding algorithm in SGA.

FIG. 16 is a flowchart showing a solution finding algorithm in TS.

FIG. 17 is a conceptual diagram showing correction of an operation plan.

FIG. 18 is a flowchart showing an embodiment of the invention described in claim 6;

[Explanation of symbols]

11,11A Total connection 12, 12A, 12B, 12C Loosely coupled part 12D first network 21 gas turbine 22 Exhaust gas boiler 23 steam turbine 24 compressor 25 boiler 26 Absorption refrigerator 27 Gas-fired refrigerator 28 Heat storage layer 30 Optimization unit 40 Stationary plant simulator

   ─────────────────────────────────────────────────── ─── Continued front page    F-term (reference) 3L061 BA05                 5H004 GA34 GB04 HA02 HB01 HB02                       HB03 HB04 JA03 KC06 KC14                       KD32 KD67 LA05

Claims (6)

[Claims] 1. A total coupling part having a plurality of input layer elements and a plurality of intermediate layer elements, wherein the intermediate layer elements are coupled to all the input layer elements, and a part of the plurality of input layer elements. A method of predicting a plant load for predicting various loads such as power load, heat load, and air load of a target plant by using a neural network of a hierarchical structure that can be analyzed and that includes a loosely coupled part in which intermediate layer elements are connected However, a neural network is constructed by learning using actual values of input factors such as weather conditions, distinction of days of the week, weekdays / holidays, plant operation status, operation patterns of plant components, and actual values of various loads. Then, the weather condition of the forecast target day is input to the learned neural network together with other input factors by the forecast value, and the target plant load is predicted. Prediction method of cement load. 2. Start-up of plant constituent equipment at each control time
When the stop state and the fuel injection amount are given as planned values, the connection state of each plant component equipment, the linear or nonlinear input / output characteristics of each plant component equipment, and the operation pattern of each plant component equipment are used to A steady-state plant simulator characterized by sequentially calculating and outputting the steady input / output state of each plant constituent device from the device on the fuel injection side toward the end at each control time. 3. A plant satisfying a load-demand balance for each load type with respect to the predicted load value obtained by the invention according to claim 1, and taking into account the restrictions on the characteristics of each plant constituent device, at least for a plant for a predetermined period. With the objective of minimizing operating costs, the start and stop states of each plant component for each control time, which is a discrete value, and the amount of fuel injected into each plant component for each control time, which is a continuous value, are state variables. And formulating an optimization problem of the plant as a nonlinear mixed integer programming problem, and generating initial values of the state variables as planned values.
Input the start / stop state and fuel injection amount of each plant component in the search process to the steady-state plant simulator described and output the input / output state of each plant component for calculation of the objective function, and calculate the planned value. An optimum operation method for a plant, characterized in that the start / stop state of each plant constituent device and the fuel injection amount are finally determined while being sequentially corrected by an optimization method. 4. The optimum operation method for a plant according to claim 3, wherein the optimization problem is solved by using Particle Swarm Optimization. 5. An operation plan of the day created in the past using the optimum operation method of the plant according to claim 3 or 4 is targeted, and the operating state of the plant constituent equipment at the present day of the day is fixed in the plant. The initial state is taken into consideration, and this initial state is considered as a constraint on the characteristics of each device, and the operation plan of the day is corrected while considering the supply and demand balance based on the load forecast value of the remaining time of the day as a constraint condition. Optimal operation method for the plant. 6. In the plant design stage, the problem of determining the optimum capacity of each plant constituent device is formulated as a combination optimization problem of determining the optimum value from the device capacity values as settable discrete values. , A load pattern such as electric load, heat load, and air load of the target plant within a predetermined period and a design value of the capacity of each plant constituent device are set, and determined by the optimum operation method of the plant according to claim 3 or 4. Each plant configuration while sequentially changing the design value by using the combination optimization method so that at least the evaluation value relating to the plant operation cost becomes the best by using the start / stop state of each plant constituent device and the fuel injection amount. An optimum plant design method characterized by determining the optimum capacity of equipment.