Improved neural network-based target power output prediction method for hydraulic power plant

Abstract

The invention discloses a hydraulic power plant target power output prediction method based on an improved neural network. The method comprises the following steps: 1, acquiring basic information and input and output data of a hydraulic power plant; 2, generating an efficiency analysis result by adopting a data envelope analysis model; 3, analyzing electric power output projection data required by the hydraulic power plant with insufficient efficiency when the efficiency reaches a target level by adopting a reverse data envelope analysis model; 4, processing the power output projection data and the original power input data; 5, constructing a target power output prediction network based on the improved neural network; and 6, constructing an MSE loss function and training the model to obtain an optimal prediction model. According to the method, efficiency optimization is introduced into output prediction, the input and output projection data of each hydraulic power plant when the target efficiency is reached can be obtained, and the model of the neural network is improved based on data feature design, so that the perspectiveness and accuracy of power output prediction are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of hydropower generation plan management, and in particular to a method for predicting the target power output of a hydropower plant based on an improved neural network. Background Art

[0002] As an important part of clean energy, hydropower occupies a key position in the global energy structure. As the core facility for hydropower energy supply, the power output of a hydropower plant directly affects the energy utilization efficiency and the economic return of the enterprise. However, in the comparison with the same industry, many hydropower plants still face the problem of low efficiency. This low-efficiency phenomenon mainly stems from two aspects: First, under the traditional management mode, hydropower plants lack the ability to quantitatively analyze the synergistic effects of core parameters such as equipment operation status, cost input, and personnel scheduling strategies, resulting in difficulty in accurately matching input and output; Second, existing power output prediction methods are mostly based on historical data and do not consider the data after the efficiency optimization of hydropower plants. In actual scenarios, in order to promote the achievement of expected goals, the prediction of power output not only involves the formulation of production plans for hydropower enterprises, but also involves the formulation of performance incentive plans for enterprises. Therefore, with the advancement of power market reform and energy structure transformation, hydropower enterprises urgently need to formulate more forward-looking production plans through more scientific and accurate prediction means.

[0003] Currently, there are still limitations in the power output prediction methods of hydropower plants at home and abroad. First, some traditional hydropower output prediction methods are mostly based on statistical analysis of historical panel data or empirical formulas, which not only make it difficult to capture the system characteristics of nonlinearity and high coupling, resulting in insufficient accuracy and objectivity of prediction results, but also such predictions generally assume that all hydropower plants always maintain the status quo, making it difficult to promote the formulation of production plans, optimization of resource allocation, and improvement of economic benefits for hydropower enterprises; Second, some prediction methods are based on statistical regression or time series models. These methods mostly rely on linear assumptions and data stationarity requirements, and are difficult to be used to handle complex nonlinear relationships in actual situations; Third, some methods focus on the application of traditional machine learning algorithms such as support vector machines and random forests in power output prediction, but in specific use, most of these algorithms rely on obtaining historical panel data for training to output prediction values. On the one hand, they cannot obtain the data after efficiency optimization and cannot provide an efficiency improvement path for inefficient hydropower plants. On the other hand, when the panel data is insufficient or missing, the effect is not good, making it difficult to assist managers in formulating targeted production plans and optimization strategies. Especially in the context of "dual carbon", how to conduct output prediction under the goal of efficiency optimization and achieve a power output plan that takes into account both actual operating rules and theoretical optimal values has become a technical problem for hydropower plants to improve comprehensive benefits. If these technical bottlenecks cannot be broken through, it may be difficult to help inefficiently operating hydropower plants plan improvements, which will cause energy waste and economic benefit losses, and further restrict the sustainable development of the clean energy system. Summary of the Invention

[0004] The present invention aims to solve the deficiencies existing in the above-mentioned prior art, and proposes a method for predicting the target power output of a hydropower plant based on an improved neural network, in order to solve the problem that it is difficult for a hydropower plant to match the efficiency optimization goal with the power output when formulating future production plans, so as to accurately and prospectively help the hydropower plant predict the power output data at the target efficiency.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for predicting the target power output of a hydropower plant based on an improved neural network according to the present invention is characterized by including the following steps:

[0007] S1. Obtain the basic information of the hydropower plant, the power output data matrix of the hydropower plant and the power input data matrix , where represents the power input data of the j-th hydropower plant, represents the power output data of the j-th hydropower plant, n represents the number of hydropower plants, m represents the feature dimension in the power input data, s represents the feature dimension in the power output data, and T represents the transpose;

[0008] S2. Analyze the power output data matrix and the power input data matrix by using the data envelopment analysis model to obtain the efficiency frontier and generate the efficiency analysis result;

[0009] S3. Based on the efficiency frontier and the efficiency analysis result, use the reverse data envelopment analysis model to calculate the power output projection data matrix required for all hydropower plants with insufficient efficiency to reach the target level when the power input remains unchanged , where represents the power output projection data of the j-th hydropower plant;

[0010] S4. Use the min-max normalization method to perform data normalization processing on and respectively to obtain the normalized power input data matrix and the normalized power output projection data matrix , where represents the normalized power input data of the j-th hydropower plant, represents the normalized power output projection data of the j-th hydropower plant;

[0011] S5. Construct a target power output prediction network based on the improved neural network, and input it into the target power output prediction network for processing, and output the target power output prediction data matrix as , where represents the target power output prediction data of the j-th hydropower plant;

[0012] S6. Based on and construct an MSE loss function, and thus use the gradient descent method to train the target power output prediction network, and calculate the MSE loss function to update the model parameters. Stop training until the MSE loss converges or reaches the maximum number of iterations, so as to obtain the optimal target power output prediction model corresponding to the optimal parameters, which is used to obtain the power output prediction data matrix required for the hydropower plant to reach the target efficiency.

[0013] The characteristics of the method for predicting the target power output of a hydropower plant based on the improved neural network according to the present invention also lie in that the efficiency analysis results include the efficiency, efficiency ranking, and linear combination coefficient of each hydropower plant.

[0014] Furthermore, the improved neural network model in S5 includes: an input layer, a fully connected feature extraction layer, a gated residual layer, a multi-head attention layer, a fully connected linear mapping layer, a fully connected dimensionality reduction layer, and an output layer;

[0015] S5.1. The input layer receives and passes it to the fully connected feature extraction layer for feature extraction, so as to obtain high-dimensional non-linear hydropower input features by using Equation (1) , where d represents the dimension of the high-dimensional non-linear hydropower input features;

[0016] (1)

[0017] In Equation (1), represents the activation function, represents the weight matrix of the fully connected feature extraction layer, represents the bias vector of the fully connected feature extraction layer;

[0018] S5.2. The gated residual layer linearly transforms by using Equation (2) to obtain the linearly transformed hydropower input features , and then calculates the update gate value of the hydropower input data by using Equation (3) , so as to obtain the updated high-dimensional mapping hydropower input features by using Equation (4) ;

[0019] (2)

[0020] (3)

[0021] (4)

[0022] In Equations (2) - (4), represents the sigmoid activation function, , represent two weight matrices of the gated residual layer, , represent two bias vectors of the gated residual layer, represents the element-wise multiplication of the matrices;

[0023] S5.3. The multi-head self-attention layer increases the dimension of to obtain the multi-dimensional hydropower input features , and then uses Equation (5) to obtain the multi-dimensional hydropower input features after feature expansion , so as to generate the query matrix of the -th head, the key matrix and the value matrix by using Equations (6), (7), and (8), and finally calculates the hydropower input self-attention features of the -th head through Equation (9), and obtains the synergy matrix of the high-dimensional hydropower input features through Equation (10);

[0024] (5)

[0025] (6)

[0026] (7)

[0027] (8)

[0028] (9)

[0029] (10)

[0030] In equations (5) - (10), is the weight matrix of the multi - head self - attention layer, represents the number of dimensions of the multi - dimensional hydropower input features after each feature expansion, , , respectively represent the mapping weight matrices of the query matrix, key matrix, and value matrix corresponding to the th head, represents the scaling coefficient, and , h represents the number of attention heads, , is the activation function, represents concatenation;

[0031] S5.4. The fully - connected linear mapping layer uses equation (11) to perform feature fusion to obtain the synergy matrix of the fused high - dimensional hydropower input features;

[0032] (11)

[0033] In equation (11), is the weight matrix of the fully - connected linear mapping layer;

[0034] S5.5. The fully - connected dimensionality - reduction layer uses equation (12) to perform feature reduction to obtain the hydropower information data matrix with reduced features. Then, after is dimension - reduced, the two - dimensional hydropower output information data matrix is obtained;

[0035] (12)

[0036] In equation (12), is the weight matrix of the fully - connected dimensionality - reduction layer, is the bias vector of the fully - connected dimensionality - reduction layer, represents the activation function;

[0037] S5.6. The output layer uses equation (13) to convert into the prediction value matrix ;

[0038] (13)

[0039] In formula (13), , are respectively the weight matrix and bias vector of the output layer.

[0040] An electronic device according to the present invention includes a memory and a processor, characterized in that the memory is used to store a program that supports the processor to execute the target power output prediction method of the hydropower plant, and the processor is configured to execute the program stored in the memory.

[0041] A computer-readable storage medium according to the present invention, characterized in that a computer program is stored on the computer-readable storage medium, and when the computer program is run by a processor, it executes the steps of the target power output prediction method of the hydropower plant.

[0042] Compared with the existing technology, the beneficial effects of the present invention are as follows:

[0043] 1. The present invention innovatively combines data envelopment analysis with an improved neural network model, embeds the hydropower plant efficiency optimization goal into the prediction method, analyzes the power output projection data of the hydropower plant when reaching the target efficiency in advance through reverse data packets, and uses it as the training benchmark for the improved neural network, so that the prediction result not only reflects the actual operation trend, but also fits the theoretical optimal production potential better. This fusion mechanism can provide a clear and more forward-looking production plan reference for hydropower enterprises, helping them to take into account historical laws and efficiency optimization goals when formulating production plans, thus avoiding the problems of resource waste or production capacity underestimation caused by traditional predictions that do not consider dynamic optimization as the goal.

[0044] 2. The method used in the present invention realizes the improvement of the model structure. Compared with the traditional prediction method, the present invention uses the cross-sectional data of the hydropower plant for prediction, does not rely on the use of multi-year panel data compulsorily, and reduces the demand for data. And based on the data characteristics of the cross-sectional data of the hydropower plant, the present invention uses a multi-head self-attention layer to learn the interaction between different high-dimensional input features of the hydropower plant and how their synergistic effects affect the output, and introduces a gating mechanism and residual connections to alleviate the problem of gradient disappearance in deep networks. Finally, the applicability, accuracy and robustness of the method used in the present invention for predicting the target power output of the hydropower plant are improved. The popularization and application of the present invention can enable the hydropower plant managers to intuitively obtain the prediction results under the target efficiency, so as to shift the management focus from passive problem response to active resource allocation optimization, providing certain technical support for building a smart hydropower system.

[0045] 3. The target power output prediction method of the present invention does not limit the types and quantities of inputs and outputs of the hydropower plant. The hydropower plant can incorporate new input parameters affecting the output or delete inappropriate input parameters according to the actual industrial scenario to fit the changes in the input-output model of the hydropower industry. Further, in the actual scenario, through the method of the present invention, the hydropower plant can also set specific different levels of target efficiency according to its actual situation and complete the prediction under corresponding circumstances. Therefore, the method of the present invention has good flexibility and scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flowchart of the target power output prediction method for a hydropower plant based on an improved neural network according to the present invention;

[0047] Figure 2 is a structural diagram of the improved neural network model according to the present invention;

[0048] Figure 3 is the internal schematic diagram of the gated residual connection according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0050] In this embodiment, a target power output prediction method for a hydropower plant based on an improved neural network, as Figure 1 shown, includes the following steps:

[0051] S1. Obtain the basic information of the hydropower plant, the power output data matrix of the hydropower plant and the power input data matrix , where represents the power input data of the j-th hydropower plant, represents the power output data of the j-th hydropower plant, n represents the number of hydropower plants, m represents the feature dimension in the power input data, s represents the feature dimension in the power output data, T represents the transpose. In this embodiment, the number of workers, the book value of physical assets, the operating cost, and the investment amount within a period are set as input parameters, and the power output within a period is set as the output parameter, that is, m = 3 and s = 1.

[0052] S2. Analyze the power output data matrix using the data envelopment analysis model and the power input data matrix , the efficiency frontier and the generation efficiency analysis results are obtained. The efficiency analysis results include the efficiency, efficiency ranking, and linear combination coefficient of each hydropower plant. In this embodiment, the specific steps for generating the efficiency analysis results are as follows:

[0053] S21. Based on the basic information, input, and output data of each hydropower plant, use the pandas library in Python for data preprocessing. Generally, there are three methods for handling missing values: directly using the features with missing values; deleting the features with missing values; and filling in the missing values. Since the present invention uses the data envelopment analysis method to calculate the power generation efficiency of each hydropower plant, which requires the data of each hydropower plant to conform to reality, in this embodiment, in order to ensure that the efficiency analysis strictly conforms to reality, the method of deleting the features with missing values is used to process the data, including using the dropna() method to delete the rows with missing values and using the drop_duplicates() method to remove the duplicate rows.

[0054] S22. Based on the preprocessed data set, use the data envelopment analysis model. It should be noted that the data envelopment analysis is a type of linear programming model. Therefore, the addConstrs() method in the gurobipy library of Python is used to add constraints, and the setObjective() method of gurobipy.Model is used to set the objective function. The efficiency is calculated by constructing the mathematical model of the linear programming problem, and the efficiency analysis results of the data envelopment analysis are obtained.

[0055] In one implementation, it is assumed that all hydropower plants satisfy variable returns to scale, and the goal is to obtain the maximum output. Therefore, in this embodiment, based on the BCC model in the data envelopment analysis, the efficiency of the k-th hydropower plant is calculated by Equation (1);

[0056] (1)

[0057]

[0058] where , , represents a non-Archimedean infinitesimal quantity, represents the linear combination coefficient of the k-th hydropower plant, and represents the linear coefficient of the k-th hydropower plant relative to the j-th hydropower plant, , , the two represent the slack variable combination in Equation (1), and represents the slack variable for the i-th input, Denote the slack variable for the \(r\)th output, , and Denote the \(i\)th input data of the \(j\)th hydropower plant, , and Denote the \(r\)th output data of the \(j\)th hydropower plant, Denote the efficiency in the calculation of the \(k\)th hydropower plant, Denote the proportion by which the output of the \(k\)th hydropower plant can increase, using and Denote the optimal solution of the model, then Denote the efficiency calculated for the \(k\)th hydropower plant and satisfy , .

[0059] S3. Based on the efficiency frontier and the efficiency analysis results, use the reverse data envelopment analysis model to calculate the power output projection data matrix required for all hydropower plants with insufficient efficiency to reach the target level when the power input remains unchanged , where Denote the power output projection data of the \(j\)th hydropower plant. In this embodiment, the specific steps for generating the power output projection data are as follows:

[0060] S31. Based on the basic information, input, output data, efficiency analysis results, and linear combination coefficients of each hydropower plant, use the gurobipy.Model.addConstrs() method in the gurobipy library of Python to add constraints, and the gurobipy.Model.setObjective() method to set the objective function, and calculate the output by constructing a linear programming problem. For each hydropower plant, calculate the power output projection data required when it reaches the target efficiency respectively;

[0061] S32. Based on the input data of the hydropower plant and the target efficiency, perform \(n\) times of reverse data envelopment analysis model calculations to obtain the power output projection data of \(n\) hydropower plants.

[0062] In one implementation, based on Equation (1) and various data in the constraint conditions of Equation (1), construct a reverse data envelopment analysis model through Equation (2) and calculate the power output projection data required for each hydropower plant to reach the target efficiency respectively;

[0063] (2)

[0064]

[0065] where It represents the increased output value of the r-th output of the k-th hydropower plant after reaching the target efficiency. It represents the weight set vector of all outputs for the k-th hydropower plant. It represents the weight value corresponding to the r-th output in the k-th hydropower plant, and , It represents the target efficiency of the k-th hydropower plant. By The value of controls the target efficiency of each hydropower plant , by Control Under different values of The increase amplitude of It represents the minimum effective value that the r-th output of the k-th hydropower plant should reach. Let It represents when the model reaches the optimal solution The corresponding optimal value, then It represents the power output projection data of the k-th hydropower plant when reaching the target efficiency. In this embodiment, it is set that , = 1, = 1, = 0.8, = 8×10 8 kW·h.

[0066] S4. Use the min-max normalization method to respectively perform data normalization processing on and to obtain the normalized power input data matrix and the normalized power output projection data matrix , where represents the normalized power input data of the j-th hydropower plant, represents the normalized power output projection data of the j-th hydropower plant. Existing prediction models and methods analyze and predict based on existing panel data, but do not perform efficiency optimization, making it difficult to provide a reference for the improvement of hydropower plants. In contrast, the present invention uses and to form a training set, which can help hydropower plants formulate more forward-looking production plans through prediction. In this embodiment, the data is normalized using Equation (3) to convert the data into data in the interval [0, 1];

[0067] (3)

[0068] Among them, represents the minimum value of the data in the β-th column, Represents the maximum value of the data in the β-th column, Represents the value of the data in the α-th row and β-th column after conversion.

[0069] S5. Construct a target power output prediction network based on an improved neural network, and input it into the target power output prediction network for processing, and output the target power output prediction data matrix as , where represents the target power output prediction data of the j-th hydropower plant. As Figure 2 shown, the improved neural network model includes: an input layer, a fully connected feature extraction layer, a gated residual layer, a multi-head attention layer, a fully connected linear mapping layer, a fully connected dimensionality reduction layer, and an output layer. By mainly introducing a gating mechanism, residual connections, and a multi-head self-attention mechanism, it alleviates the problem of gradient disappearance in deep networks and captures and fuses the non-linear synergistic effects between different high-dimensional input features of hydropower plants, ultimately improving the adaptability of the prediction to hydropower data and the accuracy and robustness of the results. In practical applications, the functions of each layer are as follows:

[0070] S5.1. The input layer receives and transmits it to the fully connected feature extraction layer for feature extraction, so as to obtain high-dimensional non-linear hydropower input features using Equation (4) , where d represents the dimension of the high-dimensional non-linear hydropower input features;

[0071] (4)

[0072] In Equation (4), represents the activation function, represents the weight matrix of the fully connected feature extraction layer, represents the bias vector of the fully connected feature extraction layer. In this embodiment, d is set to 64.

[0073] S5.2. The gated residual layer, the principle of which is as Figure 3 shown, uses Equation (5) to perform a linear transformation on to obtain the linearly transformed hydropower input features , and then uses Equation (6) to calculate the update gate value of the hydropower input data , so as to obtain the updated high-dimensional mapped hydropower input features using Equation (7) ;

[0074] (5)

[0075] (6)

[0076] (7)

[0077] In formulas (5)-(7), represents the sigmoid activation function, , represent the two weight matrices of the gated residual layer, , represent the two bias vectors of the gated residual layer, represents the element-wise multiplication of the corresponding positions of the matrices.

[0078] S5.3. Traditional analysis models and prediction methods ignore the potential interaction that may exist between different input features. Therefore, in this embodiment, the multi-head self-attention layer is used to learn the interaction between different high-dimensional input features of the hydropower plant and how their synergy affects the output. The multi-head self-attention layer increases the dimension of , treats the d-dimensional input features of each hydropower plant sample as a feature sequence, with each sequence corresponding to a feature dimension, to obtain the multi-dimensional hydropower input features , and then uses formula (8) to obtain the multi-dimensional hydropower input features after feature expansion , thereby using formulas (9), (10), and (11) to generate the query matrix for the th head, the key matrix , and the value matrix . Finally, the hydropower input self-attention feature for the th head is calculated through formula (12), and the synergy effect matrix of the hydropower high-dimensional input features is obtained through formula (13);

[0079] (8)

[0080] (9)

[0081] (10)

[0082] (11)

[0083] (12)

[0084] (13)

[0085] In equations (5)-(10), is the weight matrix of the multi-head self-attention layer, represents the number of dimensions of the multi-dimensional hydropower input features after each feature expansion, , , represents the mapping weight matrix of the query matrix, key matrix, and value matrix corresponding to the th head, represents the scaling coefficient, and , h represents the number of attention heads, , is the activation function, represents concatenation. In this embodiment, h = 4 is set, .

[0086] S5.4. The fully connected linear mapping layer uses equation (14) to perform feature fusion to obtain the synergy effect matrix of the fused high-dimensional hydropower input features, enabling the model to better learn the information on the synergy effect between the high-dimensional hydropower input features of multiple heads;

[0087] (14)

[0088] In equation (14), is the weight matrix of the fully connected linear mapping layer.

[0089] S5.5. The fully connected dimensionality reduction layer uses equation (15) to perform feature reduction to obtain the hydropower information data matrix with reduced features. Then, through the squeeze() method in the Pytorch framework, is dimensionally reduced, that is, , thus obtaining the two-dimensional hydropower output information data matrix ;

[0090] (15)

[0091] In equation (15), is the weight matrix of the fully connected dimensionality reduction layer, is the bias vector of the fully connected dimensionality reduction layer, represents the activation function.

[0092] S5.6. The output layer uses Equation (16) to convert it into a predicted value matrix of the hydropower plant's power output ;

[0093] (16)

[0094] In Equation (16), , are the weight matrix and bias vector of the output layer, respectively.

[0095] S6. Based on and construct an MSE loss function, and thus use the gradient descent method to train the target power output prediction network, and calculate the MSE loss function to update the model parameters. After the MSE loss converges or reaches the maximum number of iterations, stop the training, so as to obtain the optimal target power output prediction model corresponding to the optimal parameters, which is used to obtain the power output prediction data matrix required for the hydropower plant to reach the target efficiency.

[0096] In this embodiment, an electronic device includes a memory and a processor. The memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory;

[0097] In this embodiment, a computer-readable storage medium stores a computer program, and when the computer program is run by a processor, it executes the steps of the above method.

[0098] In summary, the present invention proposes a method for predicting the target power output of a hydropower plant based on an improved neural network, aiming to more scientifically and accurately predict the target power output of the hydropower plant and support a more forward-looking production plan. In the method of the present invention, the hydropower plant efficiency optimization target is embedded in the prediction model. Based on reverse data envelopment analysis, the power output projection data of the hydropower plant when the efficiency reaches the target level is calculated and used as the training benchmark for the improved neural network, so that the prediction result not only reflects the actual operation trend but also fits the production potential better. First, the data envelopment analysis model is used to process the input and output data of the hydropower plant collected, which can convert the multiple input and output data of the hydropower plant into efficiency without pre-specifying the power output function of the hydropower plant and allowing the power output efficiency of the hydropower plant to change over time. Second, the reverse data envelopment analysis model is used to project and optimize the original input-output data, and its feature is that it can give the resource allocation improvement information of the hydropower plant with lower efficiency, which is beneficial to optimizing the training data of the neural network according to the target efficiency. Finally, the present invention uses a model based on an improved neural network for prediction, which increases the objectivity and efficiency of the prediction. Generally speaking, this invention improves the adaptability to hydropower plant data and the prediction accuracy of the target power output by using a linear programming model and a deep learning algorithm. The prediction with embedded efficiency optimization can help the hydropower plant convert the target orientation into a process orientation to promote better utilization of resources by hydropower enterprises in the future.

[0099] It should be noted that the specific embodiments described above are only the preferred embodiments of the present invention and are not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made by those skilled in the art within the spirit and principle of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for predicting the target power output of a hydropower plant based on an improved neural network, characterized in that, It includes the following steps: S1. Obtain the basic information of the hydropower plant, the power output data matrix of the hydropower plant and the power input data matrix , where represents the power input data of the j-th hydropower plant, represents the power output data of the j-th hydropower plant, n represents the number of hydropower plants, m represents the feature dimension in the power input data, s represents the feature dimension in the power output data, and T represents the transpose; S2. Analyze the power output data matrix using the data envelopment analysis model and the power input data matrix to obtain the efficiency frontier and generate the efficiency analysis results; S3. Based on the efficiency frontier and the efficiency analysis results, use the reverse data envelopment analysis model to calculate the power output projection data matrix required for all hydropower plants with insufficient efficiency to reach the target level when the power input remains unchanged. , where represents the power output projection data of the j-th hydropower plant; S4. Respectively use the min-max normalization method for and to perform data normalization processing, obtaining the normalized power input data matrix and the normalized power output projection data matrix . Among them, represents the normalized power input data of the j-th hydropower plant, and represents the normalized power output projection data of the j-th hydropower plant; S5. Construct a target power output prediction network based on the improved neural network, and input it into the target power output prediction network for processing, and output the target power output prediction data matrix as , where represents the target power output prediction data of the j-th hydropower plant; S6. Based on and construct an MSE loss function, thereby training the target power output prediction network using the gradient descent method, and calculating the MSE loss function to update the model parameters. Stop training until the MSE loss converges or reaches the maximum number of iterations, thereby obtaining the optimal target power output prediction model corresponding to the optimal parameters, which is used to obtain the power output prediction data matrix required for the hydropower plant to reach the target efficiency.

2. The method for predicting the target power output of a hydropower plant based on an improved neural network according to claim 1, wherein The efficiency analysis results include the efficiency, efficiency ranking, and linear combination coefficient of each hydropower plant.

3. The method for predicting the target power output of a hydropower plant based on an improved neural network according to claim 1, characterized in that The improved neural network model in S5 includes: an input layer, a fully connected feature extraction layer, a gated residual layer, a multi-head attention layer, a fully connected linear mapping layer, a fully connected dimensionality reduction layer, and an output layer; S5.

1. The input layer receives and passes it to the fully-connected feature extraction layer for feature extraction, so as to obtain high-dimensional non-linear hydropower input features by using Equation (1) , where d represents the dimension of the high-dimensional non-linear hydropower input features; (1) In formula (1), represents the activation function, represents the weight matrix of the fully connected feature extraction layer, represents the bias vector of the fully connected feature extraction layer; S5.

2. The gated residual layer performs a linear transformation on to obtain the linearly transformed hydropower input features , and then calculates the update gate value of the hydropower input data using Equation (3) , so as to obtain the updated high-dimensional mapped hydropower input features using Equation (4) ; (2) (3) (4) In formulas (2)-(4), represents the sigmoid activation function, , represent two weight matrices of the gated residual layer, , represent two bias vectors of the gated residual layer, represents the element-wise multiplication of matrices; S5.

3. The multi-head self-attention layer increases the dimension to obtain multi-dimensional hydropower input features , and then uses Equation (5) to obtain the multi-dimensional hydropower input features after feature expansion , thereby generating the query matrix of the -th head, the key matrix , and the value matrix using Equations (6), (7), and (8). Finally, calculate the self-attention features of hydropower input of the -th head through Equation (9) , and obtain the synergy effect matrix of the high-dimensional hydropower input features through Equation (10); (5) (6) (7) (8) (9) (10) In Formula (5) - Formula (10), is the weight matrix of the multi - head self - attention layer, represents the number of dimensions of the multi - dimensional hydropower input features after each feature expansion, , , respectively represent the mapping weight matrices of the query matrix, key matrix, and value matrix corresponding to the th head, represents the scaling coefficient, and , where h represents the number of attention heads, , is the activation function, represents concatenation; S5.

4. The fully connected linear mapping layer uses Equation (11) to perform feature fusion to obtain a collaborative effect matrix of the fused high-dimensional input features ; (11) In formula (11), is the weight matrix of the fully connected linear mapping layer; S5.

5. The fully connected dimensionality reduction layer uses Equation (12) to perform feature reduction to obtain a feature-reduced hydropower information data matrix . Then, after dimensionality reduction on , a two-dimensional hydropower output information data matrix is obtained; (12) In formula (12), is the weight matrix of the fully connected dimensionality reduction layer, is the bias vector of the fully connected dimensionality reduction layer, represents the activation function; S5.

6. The output layer uses Equation (13) to convert it into a predicted value matrix of the power output of the hydropower plant ; (13) In Equation (13), , are the weight matrix and the bias vector of the output layer, respectively.

4. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor to execute the hydropower plant target power output prediction method described in any one of claims 1 to 3, and the processor is configured to execute the program stored in the memory.

5. A computer-readable storage medium, on which a computer program is stored, characterized in that, When the computer program is run by the processor, it executes the steps of the hydropower plant target power output prediction method described in any one of claims 1 to 3.

Images