Substation load prediction method based on CEEMDAN and improved Kepler algorithm optimization model

Abstract

The invention discloses a transformer substation load prediction method based on a CEEMDAN and an improved Kepler algorithm optimization model. The method comprises the following steps: firstly, processing volatility and aperiodicity of an original load of a transformer substation by adopting the CEEMDAN; secondly, by introducing a Niederreiter sequence and a self-adaptive chaos variation mechanism, the improved Kepler algorithm effectively avoids the problem of a local optimal solution while keeping population diversity, and the convergence speed and the global search ability of the algorithm are improved. Meanwhile, by combining the advantages of the CNN and the LSTM network, the designed hybrid combined neural network model can fully extract the spatial-temporal characteristics in the load data, and the accuracy and stability of load prediction are improved. Meanwhile, by improving the Kepler optimization algorithm, the number of convolutional layers and the learning rate of the convolutional neural network and the number of hidden layers and the learning rate of the long-short-term memory neural network are optimized, and the defect of manually determining parameters is avoided. The prediction performance of the model is superior to that of a traditional method, and the effectiveness and practicability of the method are verified.

Description

Technical Field

[0001] The invention relates to a substation load prediction method, in particular to a substation load prediction method based on CEEMDAN and an improved Kepler algorithm optimization model, and belongs to the field of power substations. Background Art

[0002] With the development of power grids and the growth of energy demand, substations are key nodes for power transmission and distribution, and their load forecasting is of great significance in the operation and planning of power systems. The accuracy of load forecasting directly affects the dispatching and operation optimization of power grids, equipment maintenance and overhaul, capacity planning and expansion, as well as emergency plans and risk management. Especially in the construction of smart grids, accurate substation load forecasting can provide important data support for intelligent dispatching, intelligent operation and intelligent maintenance, and improve the overall efficiency and reliability of power systems. Traditional load forecasting methods mainly rely on statistical models, such as autoregressive moving average models, which are widely used due to their solid theoretical foundation and low computational complexity. At the same time, the strategy of combining fuzzy membership and support vector quantile regression and the method of combining multiple regression models with dimensionality reduction optimization have shown effectiveness in short-term load forecasting. However, with the increasing complexity of power systems and the surge in data volume, the ability of traditional models in mining nonlinear and high-dimensional data features has gradually revealed limitations, and the prediction accuracy is often difficult to meet actual needs.

[0003] In recent years, the development of machine learning and deep learning technologies has provided new solutions for substation load forecasting. These technologies can capture complex nonlinear relationships by learning and analyzing a large amount of historical data, thereby significantly improving the prediction accuracy. For example, the short-term load forecasting method based on deep belief network uses Gauss-Bernoulli restricted Boltzmann machine and some supervised training algorithms for pre-training to improve the prediction accuracy. In addition, improved convolutional neural networks, deep recurrent neural networks, gated neural networks and their combined models have made significant progress in the field of load forecasting. Nevertheless, these deep learning models still face some challenges, including local optimality that may occur when setting parameters, high computational complexity, overfitting and underfitting, difficulty in hyperparameter tuning, unstable training, gradient disappearance or explosion, and strong data dependence.

[0004] In order to solve the above problems, swarm intelligence optimization algorithms are widely used in the optimization of neural network parameters. These algorithms simulate group behavior in nature and can effectively search in parameter space, thereby improving the prediction performance of the model. For example, the optimization of random forest parameters by the cross-cutting algorithm and the application of the sparrow search algorithm in the long short-term memory neural network (LSTM) have proved the effectiveness of swarm intelligence optimization algorithms in improving load forecasting accuracy.

[0005] Among many swarm intelligence optimization algorithms, the Kepler Optimization Algorithm (KOA) has attracted attention due to its unique advantages. As an emerging optimization algorithm, KOA has efficient search capabilities and global search capabilities, and can find better solutions in complex parameter spaces. However, KOA is still in the development stage and has problems such as being prone to falling into local optimality and slow convergence. Therefore, combining advanced deep learning technology with efficient optimization algorithms will provide a more accurate and stable solution for substation load forecasting. Summary of the invention

[0006] In order to solve the shortcomings of the above-mentioned technology, the present invention provides a substation load forecasting method based on CEEMDAN and an improved Kepler algorithm optimization model. The purpose of the present invention is to improve the accuracy of substation load forecasting and provide an accurate basis for the operation and maintenance of the power grid. In view of the non-stationarity of the substation load, CEEMDAN (Complete Set Empirical Mode Decomposition Adaptive Noise) is used to decompose to improve the periodicity of the subsequence. Then, on the basis of the traditional Kepler algorithm, the Niederreiter sequence and the adaptive chaos mutation mechanism are introduced to enrich the initial population diversity while being able to quickly jump out of the local optimum through the mutation mechanism, thereby improving the convergence accuracy and convergence speed of the algorithm. Secondly, combining the advantages of convolutional neural network (CNN) and long short-term memory network (LSTM), a combined neural network model is designed, which uses convolutional neural network to mine the local change mode in the substation load data in a short time, extracts spatial features of the load data in the local time period through the convolution kernel, and then mines the time-varying features of the substation load data through the long short-term memory neural network, including long-term trends, periodic effects and time dependencies in the data, and effectively captures these time features using the memory unit and gating mechanism of the long short-term memory neural network. In substation load forecasting, the combination of spatial and temporal features can enable the model to capture both short-term fluctuation patterns and long-term change trends, providing more accurate forecasts.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is: a substation load forecasting method based on CEEMDAN and improved Kepler algorithm optimization model, including the following three stages:

[0008] Step S1: Data preprocessing stage: CEEMDAN decomposition method is used to deal with the volatility and non-periodicity of the original load of the substation;

[0009] Step S2: Parameter optimization stage: By introducing the Niederreiter sequence and adaptive chaotic mutation mechanism, the Kepler optimization algorithm is improved to optimize the hyperparameters of the CNN-LSTM neural network;

[0010] Step S3: Model prediction stage: The spatiotemporal features in the load data are extracted through the CNN and LSTM combined neural network, and finally the load prediction value of the substation is obtained.

[0011] Preferably, it includes the following three stages:

[0012] Step S1: Data preprocessing stage: The CEEMDAN decomposition method is used to decompose the substation historical load data into several subsequences to reduce the volatility and non-periodicity of the original sequence. Considering that the dimensions and magnitudes of historical data such as temperature, current and load in the original data are different, the normalization method is used to eliminate the dimensions contained in the historical data to speed up the algorithm optimization process. Finally, the original data is divided into a training set and a test set in a ratio of 7:3;

[0013] The original data is divided into 70% of the data as a training set and 30% of the data as a test set according to the ratio of 7:3.

[0014] Step S2: parameter optimization stage: input the training set data of each subsequence obtained in step S1 into the CNN-LSTM combined neural network to train the model, and at the same time optimize the hyperparameters in the CNN network and the LSTM network through the improved Kepler algorithm IKOA to obtain the optimal neural network hyperparameters corresponding to each subsequence;

[0015] Step S3: Model prediction stage: The optimal neural network hyperparameters of each subsequence obtained in step S2 are brought into the CNN-LSTM combined neural network, and the test set is brought into the trained optimal CNN-LSTM combined neural network to predict each subsequence. Finally, the prediction results of each subsequence are summed to obtain the final load prediction value of the substation.

[0016] Preferably, in step S1, the steps of the CEEMDAN decomposition method are:

[0017] S11. For a given signal x(t), the purpose of EMD is to decompose it into a set of IMFs and a residual term. Each IFMc i (t) is obtained through the following iterative process:

[0018]

[0019] Where N is the number of IFMs and r(t) is the residual;

[0020] S12. In CEEMDAN, white noise n(t) based on Gaussian distribution is added to the original signal:

[0021] x n (t) = x(t) + n(t)

[0022] S13, after adding noise each time, the signal x n (t) Perform EMD decomposition to obtain a set of IFMs:

[0023]

[0024] In the formula, represents the IMF after j times of noise addition;

[0025] S14. The final IMFs are obtained by averaging the IMFs in all noise addition processes:

[0026]

[0027] S15. A key feature of CEEMDAN is that after each decomposition, the residual term is updated;

[0028] r(t)=r n (t)-c i (t)

[0029] Finally, the original substation load sequence x(t) is decomposed by CEEMDAN into N components and a residual r(t):

[0030]

[0031] The normalization formula used is:

[0032] x nk =(c k -c k,min ) / (c k,max -c k,min )

[0033] Where: x nk is the normalized value, c k is the kth sequence of substation load after decomposition by CEEMDAN, c k,min is the minimum value of the kth sequence, c k,max is the maximum value of the kth sequence.

[0034] Preferably, in step S2, the hyperparameters in the CNN network are: the number of convolutional layers and the learning rate, and the hyperparameters in the LSTM network are: the number of hidden layers and the learning rate;

[0035] Among them, the calculation steps of improving the Kepler optimization algorithm to optimize the hyperparameters of the CNN-LSTM neural network are:

[0036] S21, CNN-LSTM hyperparameter initialization;

[0037] S22, planetary velocity calculation;

[0038] S23, updating the position of the target planet;

[0039] S24, update the distance to the sun;

[0040] S25. Optimization of the positions of the sun and planets.

[0041] Preferably, in step S21, during the initialization process, a group of nP planets will be randomly dispersed in dm dimensions. In order to improve the diversity of the initial planet population, the first improvement of the Kepler algorithm in this solution is: using the Niederreiter sequence to improve the initialization process of the Kepler algorithm. The specific calculation formula is:

[0042]

[0043] Where: represents the i-th candidate solution, i.e., the planet, nP represents the number of candidate solutions in the search space, and They represent the lower and upper limits of the jth decision parameter, dm represents the problem dimension, and d k,i To generate the matrix C j The binary number generated by and i, b k+1 Is the base number.

[0044] Preferably, in step S22, the KOA algorithm determines the speed of the planet according to the distance between the planet and the sun. When the distance is close, the movement speed is faster, and when the distance is far, the movement speed is slower. The gravitational influence of the sun increases the proximity of the planet, forcing the planet to speed up to avoid being attracted; on the contrary, when the distance is far, the weak gravity will slow down the speed of the planet, and its mathematical expression is:

[0045] In the following formula, Vc i (t) represents the speed of planet i at time t, Rn i-nrm (t) represents the distance between the sun and the target planet K at time t i The normalized value of the Euclidean distance between i (t) represents the distance between the sun and the target planet K at time t i The distance between them, F is a randomly selected integer from the set {-1,1}, represents a vector randomly selected from the set and {-1,1}; r 1 ,r 2 ,r 3 ,r 4 ,r 5 is a random number uniformly distributed in the range [0,1], μ(t) represents the universal gravitational constant; and represents a solution randomly selected from the population, and M ns and mni Respectively represent K s and K i The quality of R ni (t) represents the optimal solution K at time t s and K i The distance between them, ε is the minimum value to prevent division by zero error;

[0046]

[0047] a i (t) is the semi-major axis of the elliptical orbit of planet i at time t. According to Kepler's third law, its calculation formula is:

[0048]

[0049] Among them, T i is randomly generated using a normal distribution to represent the orbital period of planet i. The KOA algorithm proposes that as the solution gradually approaches the region of the optimal solution, the semi-major axis of the elliptical orbit of planet i should be systematically reduced. The optimal solution K s and K i The normalized Euclidean distance between is expressed as:

[0050]

[0051] Preferably, in step S23, a local search is performed using the solution close to the sun to find the solution with the best position, and the state of each planet at a position far from the sun is updated:

[0052]

[0053] in, represents the new position of planet i at time t+1, Indicates its current position, Vc i (t) represents the velocity of planet i at time t, The sun position representing the solution with the best fitness value;

[0054] represents the attraction between solutions, based on the law of universal gravitation, its expression is:

[0055]

[0056] Among them, e i represents the eccentricity of the planetary orbit, Mn s and mn i The expressions are:

[0057]

[0058] in, represents the candidate solution corresponding to the worst fitness value,

[0059] Indicates the candidate solution corresponding to the optimal fitness value, fit k (t) represents the fitness value of planet k at time t;

[0060] In order to improve the global search capability of the Kepler algorithm and enhance the robustness of the algorithm, the second improvement of the Kepler algorithm in this scheme is to further update the position of each planet using the adaptive chaotic mutation mechanism:

[0061]

[0062] Among them, K i (t+1) represents the latest position of planet i at time t+1, α min and α max are the minimum and maximum variation intensity factors, t and T max Represent the current number of iterations and the maximum number of iterations respectively. represents the new position of planet i at time t+1;

[0063] Tent(K i (t+1)) is the chaotic sequence generated by Tent mapping, and its expression is:

[0064]

[0065] Preferably, in step S24, the distance between the planet and the sun changes with time, so the KOA algorithm uses simulation to improve planetary operations. When the planet is close to the sun, its position is preferentially used; when the planet is far from the sun, exploration is preferentially performed. The mathematical expression is:

[0066]

[0067] in, and represents the solution randomly selected from the population at time t; h is the adaptive control factor of the distance between the sun and the planet at a given time, and its expression is:

[0068]

[0069] Among them, γ is a number randomly generated according to normal distribution, α 2 During the optimization process, it gradually decreases from -1 to -2, and its expression is:

[0070]

[0071] Step S25 uses an elite strategy to optimize the positions of the sun and planets to ensure the best positions of the sun and planets and obtain the optimal hyperparameters of the CNN-LSTM network:

[0072]

[0073] Preferably, in step S3, the optimal neural network hyperparameters include the optimal number of convolutional layers l, CNN learning rate, number of hidden layers s, and LSTM learning rate;

[0074] Among them, the mathematical expressions of the convolution layer and pooling layer of the CNN network are:

[0075]

[0076] In the formula, X is the input data, K is the convolution kernel, X*K represents the convolution operation, and b i is the bias term, f represents the activation function, select the ReLU function, V i It is the feature map obtained by the convolution layer after convolution operation on the input data through a set of convolution kernels; P i,j represents the output of the pooling layer, s is the step size of the pooling window, and l represents the number of convolutional layers; max m，n For the maximum pooling operation, select V l The maximum value of the coverage area (i·s+m,j·s+n) in ; m,n are the indices of the pooling windows.

[0077] Convolutional layer: Each convolutional layer uses multiple convolution kernels to convolve the input data to extract local spatial information. By continuously adding convolutional layers, the model can gradually capture more complex patterns in the input data and generate high-level feature representations.

[0078] Pooling layer: In order to reduce the dimension and computational complexity of the feature map while retaining key feature information, the pooling layer downsamples the feature map after the convolution layer. Commonly used pooling methods include maximum pooling and average pooling, which help extract the most important features and improve computational efficiency.

[0079] Output feature map: The final output of CNN is the extracted high-dimensional feature map, which retains the key patterns and local features in the input data.

[0080] In the present invention, CNN is used to extract features from the input substation load data, including potential periodicity and spatial correlation. These features are not limited to the local structure of the time series data, but also include the possible trends, periodic fluctuations and abnormal changes in the substation load;

[0081] Preferably, h in the LSTM neural network t-1 is the hidden state of the previous moment, and the current input x tDetermine the information to be forgotten through the forget gate and output the corresponding weight matrix f t ; Determine the information that needs to be updated through the input gate and output i t and candidate memory cells f t and the memory cell state C at the previous moment t-1 Multiply, add i t and The product is the current state of the memory unit C. t ; Through the output gate o t Determine the final output information and combine it with the memory cell state C t Calculate the hidden state h at the current moment t , the specific calculation process is as follows:

[0082] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0083] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0084]

[0085] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0086] h t =o t tanh(C t )

[0087] Where σ is the Sigmoid activation function, W f and b f are the weight matrix and bias term of the forget gate respectively; W i and b i are the weight matrix and bias term of the input gate respectively; tanh is the activation function, W C and b C are the weight matrix and bias term of the candidate memory unit respectively; W o and b o They are the weight matrix and bias term of the output gate respectively;

[0088] For the kth sequence, the hidden state calculated at each time step is It is passed to the output layer, which is based on the hidden state of the last time step. To get the prediction result of the kth sequence:

[0089]

[0090] In the formula, is the weight matrix of the k-th sequence output layer, is the bias vector of the kth sequence output layer;

[0091] The prediction results of k sequences are summed to obtain the final prediction result:

[0092]

[0093] in, Represents the final substation load forecast value.

[0094] LSTM solves the gradient vanishing and gradient exploding problems by introducing memory units and gating mechanisms (including input gate, forget gate, and output gate), so that the model can effectively process long-span sequence data.

[0095] Input gate: controls the extent to which new information enters the memory cell, thereby retaining or updating key information.

[0096] Forget gate: decides whether to discard the information in the memory unit of the previous moment, so as to avoid the interference of irrelevant or outdated information on the prediction.

[0097] Output gate: determines the output value at the current moment, that is, the state output by the memory unit.

[0098] Memory unit: retains important timing information through a gating mechanism to model long-term dependencies.

[0099] In the present invention, the LSTM part receives the high-dimensional feature graph extracted by CNN and uses these features as time series input to further capture the time dependency and trend changes in the substation load data. LSTM can effectively handle complex load forecasting tasks, especially when the data contains long-term dependencies, and its gating mechanism can retain key historical information to improve the prediction accuracy.

[0100] The present invention discloses a substation load forecasting method based on CEEMDAN and improved Kepler algorithm optimization deep learning. First, CEEMDAN is used to process the volatility and non-periodicity of the original load of the substation. Secondly, by introducing the Niederreiter sequence and the adaptive chaotic mutation mechanism, the improved Kepler algorithm effectively avoids the problem of local optimal solution while maintaining population diversity, and improves the convergence speed and global search ability of the algorithm. At the same time, combining the advantages of CNN and LSTM networks, the designed hybrid combination neural network model can fully extract the spatiotemporal characteristics in the load data, and improve the accuracy and stability of load forecasting. At the same time, the number of convolutional layers and learning rate of the convolutional neural network and the number of hidden layers and learning rate of the long short-term memory neural network are optimized by the improved Kepler optimization algorithm to avoid the defect of artificially determining parameters. The prediction performance of the model is better than that of the traditional method, which verifies the effectiveness and practicality of the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0101] Figure 1 is a prediction flow chart of the present invention;

[0102] Figure 2 This is a schematic diagram of original load data in an embodiment of the present invention;

[0103] Figure 3 This is a CEEMDAN decomposition result diagram in an embodiment of the present invention;

[0104] Figure 4 This is a CNN-LSTM combined neural network model diagram in an embodiment of the present invention;

[0105] Figure 5 This is a load forecast result diagram in an embodiment of the present invention;

[0106] Figure 6 It is an error box plot of load prediction in an embodiment of the present invention. DETAILED DESCRIPTION

[0107] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0108] Example

[0109] according to Figure 1 The substation load prediction flow chart based on CEEMDAN-IKOA-CNN-LSTM is shown in Figure 2 The original substation load data in is decomposed into several subsequences using the CEEMDAN decomposition method in step 1, where the standard noise difference is selected as 0.2, the number of added noise sequences is 500, and the maximum number of iterations is 5000.

[0110] X k →(x1 ,x 2 ,...,x k ,...,x K )

[0111] Where: 1≤k≤K, x k is the reconstructed load sequence of the kth substation.

[0112] After CEEMDAN decomposition, we get Figure 3 Results are shown. The original load data of the substation is decomposed into 5 IMF subsequences and a residual component RES. Among them, the IMF1 component shows very high-frequency oscillations with large amplitudes. This means that it captures the instantaneous fluctuations or noise part in the data and reduces the non-periodicity of the main signal; IMF2 to IMF5 gradually show longer periods and lower-frequency fluctuations; RES is a relatively smooth trend line with almost no obvious high-frequency fluctuations. In summary, CEEMDAN decomposition can effectively extract different periodic and trend components from the original substation load data, thereby reducing the volatility and non-periodicity of the data and helping to improve the accuracy of load forecasting.

[0113] Since the input data includes data such as load, temperature and current, in order to prevent the difference in units and magnitudes of different input data from affecting the prediction results and to eliminate the dimension to speed up the optimization process of the algorithm, the data is normalized as follows:

[0114] x nk =(c k -c k,min ) / (c k,max -c k,min );

[0115] Where: x nk is the normalized value, c k is the kth sequence of substation load after decomposition by CEEMDAN, c k,min is the minimum value of the kth sequence, c k,max is the maximum value of the kth sequence.

[0116] The normalized data is divided into training set and test set.

[0117] Using the improved Kepler optimization algorithm in step 2, set the number of planets to 50 and the number of iterations to 100. At the same time, substitute the training set into the CNN-LSTM combined model for training. The improved Kepler optimization algorithm is used to optimize the number of hidden layers and learning rates of the CNN network and LSTM network to obtain the optimal parameters, where the fitness function of the planet is:

[0118]

[0119] In the formula, and x i are the predicted value and the actual value respectively. N represents the prediction step length, and N is 24 in this embodiment.

[0120] Adjust the CNN-LSTM parameters until the fitness requirement or the maximum number of iterations is reached to obtain the optimal CNN-LSTM hyperparameters.

[0121] Substitute the trained optimal parameters into the CNN-LSTM model and substitute the test set into Figure 4 In the CNN-LSTM network model shown in the figure, first, the original sequence can effectively extract spatial features from historical data through the CNN layer. Through convolution and pooling operations, CNN can capture the spatial features of load data at different time scales, which is very important for analyzing the spatial laws in load data. Secondly, based on the feature sequence extracted by CNN, the LSTM layer further captures the long-term dependencies in the substation load data through the hidden layer. The LSTM network can effectively memorize the long-term dependencies in historical data, such as seasonal changes and nonlinear trends in load data, through its gating mechanism, including input gate, forget gate and output gate, and can effectively extract the spatiotemporal features of substation load data. Finally, the substation load forecasting results of K modes are obtained through the fully connected layer, and the weighted sum of the results is obtained to obtain the load forecasting result as follows:

[0122]

[0123] Where: is the predicted value of the kth sample mode, x p The load forecast result.

[0124] In the specific application example, CNN-LSTM, VMD (variational mode decomposition)-CNN-LSTM, CEEMDAN-CNN-LSTM, CEEMDAN-KOA-CNN-LSTM, and CEEMDAN-IKOA-CNN-LSTM models are constructed for comparison and verification. Figure 5 and Figure 6 The prediction results and error box plots are shown.

[0125] The experimental results show that the CEEMDAN-CNN-LSTM model shows higher prediction accuracy than the VMD-CNN-LSTM model, which indicates that the CEEMDAN decomposition method is more effective in extracting key features from power load data. In addition, the improved Kepler optimization algorithm is introduced to optimize the number of convolutional layers and learning rate of the convolutional neural network and the number of hidden layers and learning rate of the long short-term memory neural network. The CEEMDAN-IKOA-CNN-LSTM model performs better than the unoptimized model, especially when predicting the peak and trough of the load, which is closer to the actual data. The CEEMDAN-IKOA-CNN-LSTM model not only inherits the advantages of CNN and LSTM, but also reduces the volatility of the original load sequence through data decomposition, and further optimizes the parameters of the neural network by introducing IKOA, which enables the CEEMDAN-IKOA-CNN-LSTM model to more accurately capture the subtle fluctuations of load changes and significantly improve the prediction effect.

[0126] from Figure 6 It can be seen that the CEEMDAN-IKOA-CNN-LSTM model has the lowest median error of -0.01%, which is significantly lower than other models. In contrast, the median error of the CNN-LSTM model is about 3.05%, the VMD-CNN-LSTM model is about 2.83%, and the CEEMDAN-CNN-LSTM model is about 0.99%. This significant difference in medians indicates that the CEEMDAN-IKOA-CNN-LSTM model performs better in terms of overall prediction accuracy.

[0127] In addition, the error range of the proposed model is the smallest and the interquartile range is the narrowest, indicating that its prediction error distribution is more concentrated and has no outliers, while other models have a certain number of outliers, which show extreme deviations in the prediction results in certain periods. The CEEMDAN-IKOA-CNN-LSTM model has no outliers, which means that its prediction results are more stable and reliable, without the influence of extreme errors.

[0128] The purpose of the present invention is to provide a substation load forecasting method based on CEEMDAN and an improved Kepler algorithm optimization model, which includes a data preprocessing stage, a parameter optimization stage and a model prediction stage. In the data preprocessing stage, in order to cope with the influence of the strong volatility of the substation load on the prediction result, CEEMDAN is used to decompose the original load data into several subsequences to reduce the non-stationarity of the sequence. In the parameter optimization stage, in view of the problems that the Kepler algorithm (KOA) is prone to fall into local optimality and slow convergence speed, an improved Kepler optimization algorithm is proposed by introducing the Niederreiter sequence and the adaptive chaotic mutation mechanism. In the model prediction stage, in order to overcome the limitations of a single neural network in learning the characteristics of the historical data of the substation load, a serial convolutional neural network and a long short-term memory neural network (CNN-LSTM) are proposed to fully explore the spatiotemporal characteristics of the substation load. Experimental results show that this method is significantly better than other models in prediction accuracy.

[0129] The above implementation modes are not limitations of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the scope of the technical solution of the present invention also belong to the protection scope of the present invention.

Claims

1. A substation load forecasting method based on CEEMDAN and improved Kepler algorithm optimization model is characterized by: It includes the following three stages: Step S1: Data preprocessing stage: CEEMDAN decomposition method is used to deal with the volatility and non-periodicity of the original load of the substation; Step S2: Parameter optimization stage: By introducing the Niederreiter sequence and adaptive chaotic mutation mechanism, the Kepler optimization algorithm is improved to optimize the hyperparameters of the CNN-LSTM neural network; Step S3: Model prediction stage: The spatiotemporal features in the load data are extracted through the CNN and LSTM combined neural network, and finally the load prediction value of the substation is obtained.

2. The substation load forecasting method based on CEEMDAN and improved Kepler algorithm optimization model according to claim 1 is characterized in that: It includes the following three stages: Step S1: Data preprocessing stage: The substation historical load data is decomposed into several subsequences using the CEEMDAN decomposition method to reduce the volatility and non-periodicity of the original sequence. At the same time, the normalization method is used to eliminate the dimensions contained in the historical data to speed up the algorithm optimization process. Finally, the original data is divided into a training set and a test set in a ratio of 7:3; Step S2: parameter optimization stage: input the training set data of each subsequence obtained in step S1 into the CNN-LSTM combined neural network to train the model, and at the same time optimize the hyperparameters in the CNN network and the LSTM network through the improved Kepler algorithm IKOA to obtain the optimal neural network hyperparameters corresponding to each subsequence; Step S3: Model prediction stage: The optimal neural network hyperparameters of each subsequence obtained in step S2 are brought into the CNN-LSTM combined neural network, and the test set is brought into the trained optimal CNN-LSTM combined neural network to predict each subsequence. Finally, the prediction results of each subsequence are summed to obtain the final load prediction value of the substation.

3. The substation load forecasting method based on CEEMDAN and improved Kepler algorithm optimization model according to claim 2 is characterized in that: In step S1, the steps of the CEEMDAN decomposition method are: S11. For a given signal x(t), the purpose of EMD is to decompose it into a set of IMFs and a residual term. Each IFMc i (t) is obtained through the following iterative process: Where N is the number of IFMs and r(t) is the residual; S12. In CEEMDAN, white noise n(t) based on Gaussian distribution is added to the original signal: x n (t)=x(t)+n(t) S13, after adding noise each time, the signal x n (t) Perform EMD decomposition to obtain a set of IFMs: In the formula, represents the IMF after j times of noise addition; S14. The final IMFs are obtained by averaging the IMFs in all noise addition processes: S15. A key feature of CEEMDAN is that after each decomposition, the residual term is updated; r(t)=r n (t)-c i (t) Finally, the original substation load sequence x(t) is decomposed by CEEMDAN into N components and a residual r(t): The normalization formula used is: x nk =(c k -c k,min ) / (c k,max -c k,min ) Where: x nk is the normalized value, c k is the kth sequence of substation load after decomposition by CEEMDAN, c k,min is the minimum value of the kth sequence, c k,max is the maximum value of the kth sequence.

4. The substation load forecasting method based on CEEMDAN and improved Kepler algorithm optimization model according to claim 2 is characterized in that: In step S2, the hyperparameters in the CNN network are: the number of convolutional layers and the learning rate, and the hyperparameters in the LSTM network are: the number of hidden layers and the learning rate; Among them, the calculation steps of improving the Kepler optimization algorithm to optimize the hyperparameters of the CNN-LSTM neural network are: S21, CNN-LSTM hyperparameter initialization; S22, planetary velocity calculation; S23, updating the position of the target planet; S24, update the distance to the sun; S25. Optimization of the positions of the sun and planets.

5. The substation load forecasting method based on CEEMDAN and improved Kepler algorithm optimization model according to claim 4 is characterized in that: In step S21, during the initialization process, a group of nP planets are randomly dispersed in dm dimensions, and the initialization process of the Kepler algorithm is improved using the Niederreiter sequence. The specific calculation formula is: Where: represents the i-th candidate solution, i.e., the planet, nP represents the number of candidate solutions in the search space, and They represent the lower and upper limits of the jth decision parameter, dm represents the problem dimension, and d k,i To generate the matrix C j The binary number generated by and i, b k+1 Is the base number.

6. The substation load forecasting method based on CEEMDAN and improved Kepler algorithm optimization model according to claim 4 is characterized in that: In step S22, the KOA algorithm determines the speed of the planet according to the distance between the planet and the sun, and its mathematical expression is: Where Vc i (t) represents the speed of planet i at time t, Rn i-nrm (t) represents the distance between the sun and the target planet K at time t i The normalized value of the Euclidean distance between i (t) represents the distance between the sun and the target planet K at time t i The distance between them, F is a randomly selected integer from the set {-1,1}, represents a vector randomly selected from the set and {-1,1}; r1, r2, r3, r4, r5 are random numbers uniformly distributed in the range [0,1], μ(t) represents the universal gravitational constant; and represents a solution randomly selected from the population, and Mn s and mn i Respectively represent K s and K i The quality of R ni (t) represents the optimal solution K at time t s and K i The distance between them, ε is the minimum value to prevent division by zero error; a i (t) is the semi-major axis of the elliptical orbit of planet i at time t. According to Kepler's third law, its calculation formula is: Among them, T i is randomly generated using a normal distribution to represent the orbital period of planet i. The optimal solution K s and K i The normalized Euclidean distance between is expressed as:

7. The substation load forecasting method based on CEEMDAN and improved Kepler algorithm optimization model according to claim 4 is characterized in that: In step S23, a local search is performed using the solution close to the sun to find the solution with the best position, and the state of each planet at a position far from the sun is updated: in, represents the new position of planet i at time t+1, Indicates its current position, Vc i (t) represents the velocity of planet i at time t, The sun position representing the solution with the best fitness value; represents the attraction between solutions, based on the law of universal gravitation, its expression is: Among them, e i represents the eccentricity of the planetary orbit, Mn s and mn i The expressions are: in, represents the candidate solution corresponding to the worst fitness value, Indicates the candidate solution corresponding to the optimal fitness value, fit k (t) represents the fitness value of planet k at time t; The position of each planet is further updated using an adaptive chaotic mutation mechanism: Among them, K i (t+1) represents the latest position of planet i at time t+1, α min and α max are the minimum and maximum variation intensity factors, t and T max Represent the current number of iterations and the maximum number of iterations respectively. represents the new position of planet i at time t+1; Tent(K i (t+1)) is the chaotic sequence generated by Tent mapping, and its expression is:

8. The substation load forecasting method based on CEEMDAN and improved Kepler algorithm optimization model according to claim 4 is characterized by: In step S24, simulation is used to improve planetary operations. When a planet is close to the sun, its position is preferentially utilized; when a planet is far from the sun, exploration is preferentially performed. The mathematical expression is: in, and represents the solution randomly selected from the population at time t; h is the adaptive control factor of the distance between the sun and the planet at a given time, and its expression is: Among them, γ′ is a number randomly generated according to the normal distribution; α2 gradually decreases from -1 to -2 during the optimization process, and its expression is: Step S25 uses an elite strategy to optimize the positions of the sun and planets to obtain the optimal hyperparameters of the CNN-LSTM network:

9. The substation load forecasting method based on CEEMDAN and improved Kepler algorithm optimization model according to claim 2 is characterized by: In step S3, the optimal neural network hyperparameters include the optimal number of convolutional layers l, CNN learning rate, number of hidden layers s and LSTM learning rate; Among them, the mathematical expressions of the convolution layer and pooling layer of the CNN network are: In the formula, X is the input data, K is the convolution kernel, X*K represents the convolution operation, and b i is the bias term, f represents the activation function, select the ReLU function, V i It is the feature map obtained by the convolution layer after convolution operation on the input data through a set of convolution kernels; P i,j represents the output of the pooling layer, s is the step size of the pooling window, and l represents the number of convolutional layers; max m，n For the maximum pooling operation, select V l The maximum value of the coverage area (i·s+m,j·s+n) in ; m,n are the indices of the pooling windows.

10. The substation load forecasting method based on CEEMDAN and improved Kepler algorithm optimization model according to claim 2, characterized in that: In step S3, h in the LSTM neural network t-1 is the hidden state of the previous moment, and the current input x t Determine the information to be forgotten through the forget gate and output the corresponding weight matrix f t ; Determine the information that needs to be updated through the input gate and output i t and candidate memory cells f t and the memory cell state C at the previous moment t-1 Multiply, add i t and The product is the current state of the memory unit C. t ; Through the output gate o t Determine the final output information and combine it with the memory cell state C t Calculate the hidden state h at the current moment t , the specific calculation process is as follows: f t =σ(W f ·[h t-1 ,x t ]+b f ) i t =σ(W i ·[h t-1 ,x t ]+b i ) the t =σ(W o ·[h t-1 ,x t ]+b o ) h t =o t ·tanh(C t ) Where σ is the Sigmoid activation function, W f and b f are the weight matrix and bias term of the forget gate respectively; W i and b i are the weight matrix and bias term of the input gate respectively; tanh is the activation function, W C and b C are the weight matrix and bias term of the candidate memory unit respectively; W o and b o They are the weight matrix and bias term of the output gate respectively; For the kth sequence, the hidden state calculated at each time step is It is passed to the output layer, which is based on the hidden state of the last time step. To get the prediction result of the kth sequence: In the formula, is the weight matrix of the k-th sequence output layer, is the bias vector of the kth sequence output layer; The prediction results of k sequences are summed to obtain the final prediction result: in, Represents the final substation load forecast value.

Images