Short-term power load prediction method based on EMD-ISSA-LSTM
Through the EMD-ISSA-LSTM method, the original load data is decomposed and predicted in combination with the LSTM model, which solves the problem of difficult to improve the prediction accuracy of power load in the existing technology, and achieves higher prediction accuracy and model generalization capabilities.
Patent Information
- Application Number
- CN202411078077.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-07
- Publication Date
- 2024-11-26
AI Technical Summary
In the prior art, the prediction accuracy of power load is difficult to improve, mainly due to the neglect of data preprocessing and model hyperparameter optimization.
Empirical modal decomposition (EMD) is used to decompose the original load data into eigenmodal functions and residual terms, and combine long and short-term memory neural network (LSTM) and improved sparrow optimization algorithm (ISSA) to construct small-component and large-component LSTMs The model performs fusion of prediction results and optimizes hyperparameters to improve prediction accuracy.
Through the EMD-ISSA-LSTM method, the nonlinearity and instability of the original data are reduced, the load prediction accuracy and generalization ability of the model are improved, and the prediction speed and accuracy are improved.
Smart Images

Figure CN119029846A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of power load forecasting, and in particular to a short-term power load forecasting method based on EMD-ISSA-LSTM. Background Art
[0002] Accurate short-term power load forecasting is crucial to the economic, safe and stable operation of the power system. In recent years, with the continuous development of clean energy, distributed energy in the power grid has increased, resulting in many serious problems in the power grid, such as new energy consumption, transformer overload and reverse power transmission of main transformers. To solve the above problems, the relevant departments of the State Grid should formulate an orderly power consumption plan and reasonably arrange power generation, transmission and supply, and these plans are highly dependent on accurate power load forecasting. Therefore, it is crucial to improve the accuracy of power load forecasting. However, the current load forecasting model considers the deep learning algorithm itself more, while the research on data preprocessing and the optimization of model hyperparameters are less, which makes it difficult to improve the accuracy of load forecasting. Summary of the invention
[0003] In view of this, the present invention provides a short-term power load forecasting method based on EMD-ISSA-LSTM, which is used to at least solve the problem that it is difficult to improve the load forecasting accuracy in the prior art.
[0004] In order to achieve the above object, the present invention adopts the following technical solution:
[0005] A short-term power load forecasting method based on EMD-ISSA-LSTM includes the following steps:
[0006] The data to be predicted is decomposed into an intrinsic mode function to be predicted and a residual term to be predicted by EMD, and the intrinsic mode function to be predicted is input into a small component LSTM load forecasting model for load forecasting, and the residual term to be predicted is input into a large component LSTM load forecasting model for load forecasting, and the two load forecasting results are fused to obtain a final forecasting result; wherein, the construction method of the small component LSTM load forecasting model and the large component LSTM load forecasting model includes the following steps:
[0007] S1. The original load data is decomposed into m intrinsic mode functions and a residual term through EMD. According to the order of magnitude, the intrinsic mode function is used as a small component signal and the residual term is used as a large component signal to build an initial LSTM load forecasting model.
[0008] S2. Using the data set composed of intrinsic mode functions to train the initial LSTM load forecasting model to obtain a small component LSTM load forecasting model;
[0009] S3. The hyperparameters of the initial LSTM load forecasting model are optimized by the ISSA algorithm until the ISSA algorithm converges to obtain the final optimization result of the hyperparameters, and the large-component LSTM load forecasting model is obtained according to the final optimization result.
[0010] Preferably, the specific content of S1 includes:
[0011] Decompose the original load data into m intrinsic mode functions and a residual term:
[0012] A=IMF1+IMF2+···+IMF m +Residual
[0013] Where A is the original load data, IMFm is the mth intrinsic mode function, and Residual is the residual term.
[0014] Preferably, the specific process of decomposition includes:
[0015] 1) Obtain the maximum and minimum points of the original load data s1 to form the maximum and minimum envelopes s2 and s3;
[0016] 2) Averaging the maximum and minimum envelopes to obtain the average line s4;
[0017] 3) Subtract s4 from s1 to obtain s5. If s5 meets the requirements of the intrinsic mode function, the decomposition is completed and s5 is the first intrinsic mode function; otherwise, repeat steps 1)-3) for s5 until the first intrinsic mode function is obtained;
[0018] 4) Subtract the first intrinsic mode function obtained in 3) from s1 to form a new signal, use the new signal to update s1 and then perform steps 1)-4) to obtain the next intrinsic mode function;
[0019] 5) And so on, finally we get m intrinsic mode functions and one residual term.
[0020] Preferably, the specific content of S2 includes:
[0021] S21. Divide the data set composed of the first intrinsic mode function into training set data and test set data, input the training set data into the initial LSTM load forecasting model for training, compare the training results with the actual load data to obtain the error, and use the Adam algorithm to update the weight matrix and bias matrix of each neuron in the initial LSTM load forecasting model. When the number of training times reaches the preset threshold, the training will be stopped and all parameters of the current LSTM load forecasting model will be saved;
[0022] S22. Input the test set data of the first eigenmode into the current LSTM load forecasting model to obtain the forecast data of the test set;
[0023] S23. Repeat S21-S22 for the remaining eigenmode functions in sequence to obtain corresponding prediction data;
[0024] S24. Select the parameter with the smallest error to obtain a small component LSTM load forecasting model.
[0025] Preferably, the specific content of S3 includes:
[0026] The hyperparameters to be optimized in the initial LSTM load forecasting model are taken as sparrows, and the set of hyperparameters is taken as the population. The ISSA algorithm is used to optimize the hyperparameters of the initial LSTM load forecasting model.
[0027] Use matrix X to represent the sparrow population:
[0028]
[0029] Where n represents the number of sparrow populations, d represents the number of LSTM hyperparameters that need to be optimized, and x represents each sparrow in the sparrow population;
[0030] Initialize the population X and get:
[0031]
[0032] Where rand is a random number between 0 and 1, lb and ub represent the upper and lower bounds of the search space respectively;
[0033] After initialization Update population X;
[0034] According to each set of LSTM hyperparameters, there are n fitness functions:
[0035]
[0036] Among the n fitness functions, the set of LSTM hyperparameters corresponding to the fitness function with the smallest value is the best fitness function:
[0037] best1=[x n1 ,x n2 ,…x nd ]
[0038] The sparrow population is divided into discoverers and followers, and the guards are randomly assigned to avoid danger. After obtaining the best fitness function, the positions of the discoverers and followers are updated. The discoverer position is updated as follows:
[0039]
[0040] Where i termax represents the maximum number of iterations set in the sparrow optimization algorithm, α represents a uniform random number in [0,1], Q represents a standard normal random number, R2 represents a random number in [0,1], ST represents the warning value of the discoverer discovering the predator, Indicates the updated position. represents the position before the update, i and j are the row and column numbers in the X matrix respectively;
[0041] The follower position update is shown as follows:
[0042]
[0043] Where x t worst represents the sparrow with the worst fitness in the current population, x t best Indicates the sparrow individual with the best fitness in the current population;
[0044] The position update of the sentinel is shown as follows:
[0045]
[0046] Where β is a random number that conforms to the standard normal distribution, K is a random number from -1 to 1, and f g and f w Respectively represent the current best and worst fitness values of the sparrow, f i Indicates the current fitness value of the sparrow; where:
[0047]
[0048]
[0049] Preferably, the specific contents of fusing the two load forecasting results to obtain the final forecasting result include:
[0050] The final prediction result after the load prediction result of the small component LSTM load prediction model is fused with the load prediction result of the large component LSTM load prediction model is:
[0051] Y result =W1X1+W2X2
[0052] Among them, Y result is the final prediction result, W1 and W2 are the weights to be optimized, X1 and X2 are the load prediction results of the small component and the large component respectively;
[0053] The optimization process for weights W1 and W2 to be optimized specifically includes:
[0054] (1) Initialize the weight matrix W and the initial number of iterations i;
[0055] (2) Determine whether i exceeds the maximum number of iterations Maxcycle. If i0 does not exceed Maxcycle, calculate the final prediction result based on W, and then calculate the mean absolute percentage error MAPE based on the final prediction result. Record the W matrix corresponding to the best MAPE value, and use the recorded W matrix to update the initial weight matrix W. At the same time, update the initial number of iterations i, and repeat steps (1)-(2); if i exceeds Maxcycle, use the currently recorded W matrix to calculate the final prediction result.
[0056] It can be seen from the above technical solution that, compared with the prior art, the present invention discloses a short-term power load forecasting method based on EMD-ISSA-LSTM, which has the following beneficial effects:
[0057] Compared with the prior art, the present invention combines EMD with a deep learning algorithm, reduces the nonlinearity and instability of the original data, and can improve the load forecasting accuracy. On this basis, ISSA is added to optimize the hyperparameters of LSTM, which improves the generalization ability of the forecasting method. The small signal model of EMD decomposition uses a single LSTM model, and the ISSA-LSTM model is selected for large component signals, which improves the model training and prediction speed, and has certain application prospects. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0059] Figure 1 A load forecasting process of a short-term power load forecasting method based on EMD-ISSA-LSTM disclosed in the present invention;
[0060] Figure 2 A schematic diagram of the EMD decomposition principle provided by an embodiment of the present invention;
[0061] Figure 3 A flowchart of constructing a small component LSTM load forecasting model and a large component LSTM load forecasting model in a short-term power load forecasting method based on EMD-ISSA-LSTM disclosed by the present invention;
[0062] Figure 4 The LSTM minimum unit provided by the embodiment of the present invention;
[0063] Figure 5 It is a flow chart of the fusion process of two load forecasting results in a short-term power load forecasting method based on EMD-ISSA-LSTM disclosed by the present invention;
[0064] Figure 6 A heat map of correlation analysis provided by an embodiment of the present invention;
[0065] Figure 7 The fitness value of the two-layer neuron provided by the embodiment of the present invention;
[0066] Figure 8 The original load data waveform diagram provided by the embodiment of the present invention;
[0067] Fig. 9 A waveform diagram after EMD decomposition provided by an embodiment of the present invention; Fig. 9 (a)-(d) are the waveforms of the decomposed intrinsic mode functions IMF1~IMF4 respectively. Fig. 9 (f) is the waveform of the residual term IMF5 after decomposition;
[0068] Fig.10 A load forecast result diagram provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0069] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0070] The present invention provides a short-term power load forecasting method based on the combination of Empirical Mode Decomposition (EMD), Long Short-Term Memory (LSTM) neural network and Improved Sparrow Search Algorithm (ISSA) to improve the generalization ability of the load forecasting model and the load forecasting accuracy. Figure 1 As shown, the following steps are included:
[0071] The data to be predicted is decomposed into the intrinsic mode function to be predicted and the residual term to be predicted by EMD, and the intrinsic mode function to be predicted is input into the small component LSTM load forecasting model for load forecasting, and the residual term to be predicted is input into the large component LSTM load forecasting model for load forecasting, and the two load forecasting results are fused to obtain the final forecast result; the EMD decomposition principle is as follows Figure 2 As shown, the construction method of the small component LSTM load forecasting model and the large component LSTM load forecasting model includes the following steps: Figure 3 As shown:
[0072] S1. The original load data is decomposed into m intrinsic mode functions and a residual term through EMD. According to the order of magnitude, the intrinsic mode function is used as a small component signal and the residual term is used as a large component signal to build an initial LSTM load forecasting model.
[0073] S2. Using the data set composed of intrinsic mode functions to train the initial LSTM load forecasting model to obtain a small component LSTM load forecasting model;
[0074] S3. The hyperparameters of the initial LSTM load forecasting model are optimized by the ISSA algorithm until the ISSA algorithm converges to obtain the final optimization result of the hyperparameters, and the large-component LSTM load forecasting model is obtained according to the final optimization result.
[0075] In order to further implement the above technical solution, the specific contents of S1 include:
[0076] Decompose the original load data into m intrinsic mode functions and a residual term:
[0077] A=IMF1+IMF2+···+IMF m +Residual
[0078] Where A is the original load data, IMFm is the mth intrinsic mode function, and Residual is the residual term.
[0079] In order to further implement the above technical solution, the specific process of decomposition includes:
[0080] 1) Obtain the maximum and minimum points of the original load data s1 to form the maximum and minimum envelopes s2 and s3;
[0081] 2) Averaging the maximum and minimum envelopes to obtain the average line s4;
[0082] 3) Subtract s4 from s1 to obtain s5. If s5 meets the requirements of the intrinsic mode function, the decomposition is completed and s5 is the first intrinsic mode function; otherwise, repeat steps 1)-3) for s5 until the first intrinsic mode function is obtained;
[0083] 4) Subtract the first intrinsic mode function obtained in 3) from s1 to form a new signal, use the new signal to update s1 and then perform steps 1)-4) to obtain the next intrinsic mode function;
[0084] 5) And so on, finally we get m intrinsic mode functions and one residual term.
[0085] In order to further implement the above technical solution, the specific contents of S2 include:
[0086] S21. Divide the data set composed of the first intrinsic mode function into training set data and test set data, input the training set data into the initial LSTM load forecasting model for training, compare the training results with the actual load data to obtain the error, and use the Adam algorithm to update the weight matrix and bias matrix of each neuron in the initial LSTM load forecasting model. When the number of training times reaches the preset threshold, the training will be stopped and all parameters of the current LSTM load forecasting model will be saved;
[0087] S22. Input the test set data of the first eigenmode into the current LSTM load forecasting model to obtain the forecast data of the test set;
[0088] S23. Repeat S21-S22 for the remaining eigenmode functions in sequence to obtain corresponding prediction data;
[0089] S24. Select the parameter with the smallest error to obtain a small component LSTM load forecasting model.
[0090] It should be noted that:
[0091] The smallest unit of LSTM is Figure 4 As shown, where σ is the sigmoid function, C t is the memory information output by this cell, h t is the output of this cell, C t-1 is the memory information output by the previous cell, h t-1 is the output of the previous cell, x t Represents the input data of the cell, and tanh is an activation function.
[0092] The sigmoid function and tanh function are
[0093]
[0094]
[0095] The input information of this cell is superimposed with the output information of the previous cell, and then the useless information is filtered out through the sigmoid function:
[0096] f t =σ(W f ·[h t-1 ,x t ]+b f )
[0097] Where W f and b f are the weight matrix and bias matrix respectively.
[0098] The input gate determines the extent to which the current input data enters the cell state. The data updates the cell state through the input gate.
[0099] i t =σ(W i ·[h t-1 ,x t ]+b i )
[0100]
[0101] The memory information of this cell is updated through the superposition of the forget gate and the input gate, specifically:
[0102]
[0103] The output of the LSTM unit is a stacked integration of the updated memory information and the input information, and its formula is
[0104] o t =σ(W o [h t-1 ,x t ]+b o )
[0105] h t =o t tanh(C t )
[0106] The output of each LSTM unit is linked to the next layer of neural network to finally get the output of the model.
[0107] In order to further implement the above technical solutions, the specific contents of S3 include:
[0108] The hyperparameters to be optimized in the initial LSTM load forecasting model are taken as sparrows, and the set of hyperparameters is taken as the population. The ISSA algorithm is used to optimize the hyperparameters of the initial LSTM load forecasting model.
[0109] Use matrix X to represent the sparrow population:
[0110]
[0111] Where n represents the number of sparrow populations, d represents the number of LSTM hyperparameters that need to be optimized, and x represents each sparrow in the sparrow population;
[0112] Initialize the population X and get:
[0113]
[0114] Where rand is a random number between 0 and 1, lb and ub represent the upper and lower bounds of the search space respectively;
[0115] After initialization Update population X;
[0116] According to each set of LSTM hyperparameters, there are n fitness functions:
[0117]
[0118] Among the n fitness functions, the set of LSTM hyperparameters corresponding to the fitness function with the smallest value is the best fitness function:
[0119] best1=[x n1 ,x n2 ,…x nd ]
[0120] The sparrow population is divided into discoverers and followers, and the guards are randomly assigned to avoid danger. After obtaining the best fitness function, the positions of the discoverers and followers are updated. The discoverer position is updated as follows:
[0121]
[0122] Where i termax represents the maximum number of iterations set in the sparrow optimization algorithm, α represents a uniform random number in [0,1], Q represents a standard normal random number, R2 represents a random number in [0,1], ST represents the warning value of the discoverer discovering the predator, Indicates the updated position. represents the position before the update, i and j are the row and column numbers in the X matrix respectively;
[0123] The follower position update is shown as follows:
[0124]
[0125] Where x t worst represents the sparrow with the worst fitness in the current population, x t best Indicates the sparrow individual with the best fitness in the current population;
[0126] The position update of the sentinel is shown as follows:
[0127]
[0128] Where β is a random number that conforms to the standard normal distribution, K is a random number from -1 to 1, and f g and f w Respectively represent the current best and worst fitness values of the sparrow, f i Indicates the current fitness value of the sparrow; where:
[0129]
[0130]
[0131] Finally, the sparrow individual with the best fitness is found by continuously iteratively updating the position of the sparrow population, that is, a set of LSTM hyperparameters with the best fitness value.
[0132] It should be noted that:
[0133] ISSA is used to optimize the hyperparameters of LSTM. ISSA updates the LSTM parameters that need to be optimized through multiple iterations, and the updated parameters are brought into LSTM to train the LSTM model. The trained LSTM model is used to predict the root mean square error, and a set of LSTM hyperparameters with the smallest root mean square error is found in LSTM through multiple iterations as the result of the optimization algorithm.
[0134] ISSA is a swarm intelligence optimization algorithm that imitates sparrows foraging. The present invention uses an improved sparrow search algorithm to optimize the hyperparameters of LSTM.
[0135] In order to further implement the above technical solution, Figure 5 As shown in the figure, the specific contents of the final prediction result obtained by fusing the two load prediction results include:
[0136] The final prediction result after the load prediction result of the small component LSTM load prediction model is fused with the load prediction result of the large component LSTM load prediction model is:
[0137] Y result =W1X1+W2X2
[0138] Among them, Y result is the final prediction result, W1 and W2 are the weights to be optimized, X1 and X2 are the load prediction results of the small component and the large component respectively;
[0139] The optimization process for weights W1 and W2 to be optimized specifically includes:
[0140] (1) Initialize the weight matrix W and the initial number of iterations i;
[0141] (2) Determine whether i exceeds the maximum number of iterations Maxcycle. If i0 does not exceed Maxcycle, calculate the final prediction result based on W, and then calculate the mean absolute percentage error MAPE based on the final prediction result. Record the W matrix corresponding to the best MAPE value, and use the recorded W matrix to update the initial weight matrix W. At the same time, update the initial number of iterations i, and repeat steps (1)-(2); if i exceeds Maxcycle, use the currently recorded W matrix to calculate the final prediction result.
[0142] The present invention will be further described below by means of specific embodiments:
[0143] First, the model input features are selected. In this embodiment, three years of load data are used, with one data point every 15 minutes, that is, 96 data points per day. To avoid data redundancy, the Pearson correlation coefficient method is used to perform correlation analysis on the input features.
[0144] like Figure 6 The figure shows the correlation analysis of the input features, where Load represents load data, Day represents date data, T_h represents the highest temperature of the day, T_l represents the lowest temperature of the day, T_p represents the average temperature of the day, R_x represents the rainfall of the day, Week represents weekly data, and Quarter represents seasonal data. It can be seen from the heat map that the correlation coefficient between rainfall data and load data is 0.069, which is a low correlation. Therefore, whether rainfall data exists in the input features remains to be determined.
[0145] In the experimental comparison, this embodiment uses four model evaluation indicators, namely, determination coefficient (R2), mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) [23-24], specifically:
[0146] Coefficient of determination:
[0147]
[0148] Mean absolute error:
[0149]
[0150] Root Mean Square Error:
[0151]
[0152] Mean absolute percentage error:
[0153]
[0154] Where n represents the number of predictions, y i " indicates the predicted value, y i Indicates the actual value, y i ” indicates the mean of the true values.
[0155] The evaluation indicators of the load forecasting model under different characteristics are shown in Table 1, where the characteristic items represent other features added in addition to the basic characteristics. The basic characteristics include historical load, temperature and date, and other characteristics include rainfall characteristics, week characteristics, seasonal characteristics and comprehensive all characteristic data. As can be seen from Table 1, compared with the load forecasting error under other characteristics, the determination coefficient of adding the week characteristics to the basic characteristics is 0.96, the average absolute error is 0.023, and the root mean square error is 0.029. Finally, historical load, temperature, date basic characteristics and week characteristics are selected as the input features of the model.
[0156] Table 1 Model evaluation indicators with different features added
[0157]
[0158] Input the original load data into EMD, such as Figure 8 is the original load waveform, Fig. 9 is the waveform after empirical mode decomposition, where IMF1 to IMF4 are the decomposed intrinsic mode functions, and IMF5 is the residual.
[0159] Then the large component signal is input into the large component LSTM load forecasting model, and the optimal network layer of LSTM is determined to be two layers by adjusting the parameters. Figure 7 The figure shows the RMSE function of different numbers of neurons in a two-layer network, where the horizontal and vertical axes represent the number of neurons in the first and second layers, and the height represents the RMSE value corresponding to different neuron groups.
[0160] When the fitness function of ISSA is set to RMSE and the optimization parameter of LSTM is the number of neurons in the two layers, by improving the sparrow search algorithm, it can be found that the fitness function has a minimum value when the number of neurons in the first layer is 104 and the number of neurons in the second layer is 134, and the load forecasting error under this parameter is the smallest.
[0161] Finally, the large component signal and the small component signal are fused to obtain the final load forecast result.
[0162] Fig.10 The final prediction results under different methods are shown, among which the centralized method includes directly using the LSTM prediction model to predict all data, directly using the ISSA-LSTM model without decomposing the data, and the EMD-ISSA-LSTM model disclosed in the present invention. Fig.10 It can be seen that the prediction curve generated by the method of the present invention is closest to the true value.
[0163] As shown in Table 2, which are the evaluation indicators of different models, it can be seen that the determination coefficient of the EMD-ISSA-LSTM method is 0.98, the average absolute error is 0.015, the root mean square error is 0.019, and the average absolute percentage error is 2.04%. It has the largest determination coefficient and the smallest error. It verifies that the model can improve the accuracy of load forecasting.
[0164] Table 2 Model evaluation indicators
[0165]
[0166] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.
Claims
1. A short-term power load forecasting method based on EMD-ISSA-LSTM, characterized in that: The following steps are involved: The data to be predicted is decomposed into an intrinsic mode function to be predicted and a residual term to be predicted by EMD, and the intrinsic mode function to be predicted is input into a small component LSTM load forecasting model for load forecasting, and the residual term to be predicted is input into a large component LSTM load forecasting model for load forecasting, and the two load forecasting results are fused to obtain a final forecasting result; wherein, the construction method of the small component LSTM load forecasting model and the large component LSTM load forecasting model includes the following steps: S1. The original load data is decomposed into m intrinsic mode functions and a residual term through EMD. According to the order of magnitude, the intrinsic mode function is used as a small component signal and the residual term is used as a large component signal to build an initial LSTM load forecasting model. S2. Using the data set composed of intrinsic mode functions to train the initial LSTM load forecasting model to obtain a small component LSTM load forecasting model; S3. The hyperparameters of the initial LSTM load forecasting model are optimized by the ISSA algorithm until the ISSA algorithm converges to obtain the final optimization result of the hyperparameters, and the large-component LSTM load forecasting model is obtained according to the final optimization result.
2. The short-term power load forecasting method based on EMD-ISSA-LSTM according to claim 1 is characterized in that: The specific contents of S1 include: Decompose the original load data into m intrinsic mode functions and a residual term: A=IMF1+IMF2+···+IMF m +Residual Where A is the original load data, IMFm is the mth intrinsic mode function, and Residual is the residual term.
3. The short-term power load forecasting method based on EMD-ISSA-LSTM according to claim 2 is characterized in that: The specific process of decomposition includes: 1) Obtain the maximum and minimum points of the original load data s1 to form the maximum and minimum envelopes s2 and s3; 2) Averaging the maximum and minimum envelopes to obtain the average line s4; 3) Subtract s4 from s1 to obtain s5. If s5 meets the requirements of the intrinsic mode function, the decomposition is completed and s5 is the first intrinsic mode function; otherwise, repeat steps 1)-3) for s5 until the first intrinsic mode function is obtained; 4) Subtract the first intrinsic mode function obtained in 3) from s1 to form a new signal, use the new signal to update s1 and then perform steps 1)-4) to obtain the next intrinsic mode function; 5) And so on, finally we get m intrinsic mode functions and one residual term.
4. The short-term power load forecasting method based on EMD-ISSA-LSTM according to claim 1 is characterized in that: The specific contents of S2 include: S21. Divide the data set composed of any intrinsic mode function into training set data and test set data, input the training set data into the initial LSTM load forecasting model for training, compare the training results with the actual load data to obtain the error, use the Adam algorithm to update the weight matrix and bias matrix of each neuron in the initial LSTM load forecasting model, stop training when the number of training times reaches the preset threshold, and save all parameters of the current LSTM load forecasting model; S22. Input the test set data of the current intrinsic mode into the current LSTM load forecasting model to obtain the load forecasting result corresponding to the test set; S23. Repeat S21-S22 for the remaining intrinsic mode functions in sequence to obtain the corresponding load prediction results; S24. Select the parameters corresponding to the minimum error to form a small component LSTM load forecasting model.
5. The short-term power load forecasting method based on EMD-ISSA-LSTM according to claim 1 is characterized in that: The specific contents of S3 include: The hyperparameters to be optimized in the initial LSTM load forecasting model are taken as sparrows, and the set of hyperparameters is taken as the population. The ISSA algorithm is used to optimize the hyperparameters of the initial LSTM load forecasting model. Use matrix X to represent the sparrow population: Where n represents the number of sparrow populations, d represents the number of LSTM hyperparameters that need to be optimized, and x represents each sparrow in the sparrow population; Initialize the population X and get: Where rand is a random number between 0 and 1, lb and ub represent the upper and lower bounds of the search space respectively; After initialization Update population X; According to each set of LSTM hyperparameters, there are n fitness functions: Among the n fitness functions, the set of LSTM hyperparameters corresponding to the fitness function with the smallest value is the best fitness function: best1=[x n1 ,x n2 ,…x nd ] The sparrow population is divided into discoverers and followers, and the guards are randomly assigned to avoid danger. After obtaining the best fitness function, the positions of the discoverers and followers are updated. The discoverer position is updated as follows: Where i termax represents the maximum number of iterations set in the sparrow optimization algorithm, α represents a uniform random number in [0,1], Q represents a standard normal random number, R2 represents a random number in [0,1], ST represents the warning value of the discoverer discovering the predator, Indicates the updated position. represents the position before the update, i and j are the row and column numbers in the X matrix respectively; The follower position update is shown as follows: Where x t worst represents the sparrow with the worst fitness in the current population, x t best Indicates the sparrow individual with the best fitness in the current population; The position update of the sentinel is shown as follows: Where β is a random number that conforms to the standard normal distribution, K is a random number from -1 to 1, and f g and f w Respectively represent the current best and worst fitness values of the sparrow, f i Indicates the current fitness value of the sparrow; where:
6. The short-term power load forecasting method based on EMD-ISSA-LSTM according to claim 1 is characterized in that: The specific contents of fusing the two load forecasting results to obtain the final forecasting result include: The final prediction result after the load prediction result of the small component LSTM load prediction model is fused with the load prediction result of the large component LSTM load prediction model is: Y result =W1X1+W2X2 Among them, Y result is the final prediction result, W1 and W2 are the weights to be optimized, X1 and X2 are the load prediction results of the small component and the large component respectively; The optimization process for weights W1 and W2 to be optimized specifically includes: (1) Initialize the weight matrix W and the initial number of iterations i; (2) Determine whether i exceeds the maximum number of iterations Maxcycle. If i0 does not exceed Maxcycle, calculate the final prediction result based on W, and then calculate the mean absolute percentage error MAPE based on the final prediction result. Record the W matrix corresponding to the best MAPE value, and use the recorded W matrix to update the initial weight matrix W. At the same time, update the initial number of iterations i, and repeat steps (1)-(2); if i exceeds Maxcycle, use the currently recorded W matrix to calculate the final prediction result.
Citation Information
Cited By
Railway traction station load prediction method and system
CN120320317A
