Prediction evaluation method for short-term power load

Abstract

The invention relates to the technical field of power load prediction, in particular to a short-term power load prediction evaluation method, which comprises the following steps: preprocessing and decomposing historical power load data; analyzing the power load influence factors and selecting the influence factors with strong correlation as strong correlation characteristics; constructing a power load prediction model; inputting the decomposed historical power load data and the strong correlation characteristics into a power load prediction model to obtain a prediction result of each component; and superposing the obtained component prediction results to obtain a final prediction result. And evaluating the prediction result. According to the method, the original load signal is decomposed into a plurality of IMFSs by using the SGMD decomposition algorithm, the nonlinearity and instability of the load sequence can be better processed, weighting is performed according to the correlation coefficient of each modal component, key information of the original power load sequence is highlighted, and the accuracy of load prediction is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power load forecasting, and in particular to a forecasting and evaluation method for short-term power load. Background Art

[0002] Power load forecasting plays a vital role in power system operation and power market transactions. Power load forecasting plays a key role in system operation, dispatching, daily transactions in the power market and resource allocation. For power system operation, load forecasting can effectively guide the power supply and heating plan of the power grid and ensure the stability and reliability of the power system. For power market transactions, accurate load forecasting can help market participants determine power purchase and sale strategies more accurately, reduce transaction risks and improve economic benefits. Power load is determined by many complex factors, such as weather conditions, holidays, economic development level and user behavior. These factors make power load highly nonlinear and unstable. In addition, power load is also affected by random factors, such as emergencies and the randomness of user behavior. All of these make power load forecasting challenging.

[0003] Therefore, how to construct a prediction model that can handle the non-stationarity and complexity of power load data and achieve efficient prediction of power load data is an issue that technical personnel in this field urgently need to solve. Summary of the invention

[0004] The purpose of the present invention is to solve at least one of the technical problems existing in the prior art and to provide a forecasting and evaluation method for short-term power load.

[0005] To achieve the above object, the technical solution adopted by the present invention is as follows: a method for predicting and evaluating short-term power load, specifically comprising the following steps: Step 1, obtaining relevant historical power load data and corresponding power load influencing factors, and preprocessing and decomposing the historical power load data; Step 2, analyzing the power load influencing factors and selecting influencing factors with strong correlation as strong correlation features; Step 3: Introduce the attention mechanism, use the IVYA algorithm to iteratively optimize the hyperparameters in GRU, and build the IVYA-GRU-Attention power load forecasting model based on the best hyperparameter combination of GRU; Step 4, inputting the decomposed historical power load data and the strong correlation features into the power load forecasting model to obtain the forecast results of each component; superimposing the obtained forecast results of each component to obtain the final forecast result; Step 5: Evaluate the prediction results.

[0006] Furthermore, in the step 1, it specifically includes: Step 1.1, preprocessing the historical power load data; The historical power load data and the power load influencing factor data are normalized: (1); in, are the original load value and the influencing factor value, are the normalized load values ​​and influencing factor values, and are the maximum and minimum values ​​of the original load data and influencing factor data respectively; Step 1.2, decomposing the preprocessed historical power load data; SGMD decomposition is used to decompose the load sequence into multiple modal components IMF. The specific algorithm is as follows: Phase space reconstruction: Assuming the timing signal , according to the embedded idea of ​​Takens theorem: (2); in, is the length of the timing signal, is the difference dimension, is the delay factor, for The number of dimensional vectors, ; Symplectic geometry matrix transformation: Assumptions Construct Hamiltonian matrix : (3); make , construct symplectic geometry matrix : (4); in, is an upper triangular matrix, the elements of which are ; matrix The characteristic value of ;matrix The characteristic value of , and the corresponding eigenvector ; Original phase space reconstruction matrix , each single-component matrix composition; (5); in, ; ;

[0007] Diagonal average: For any , Be the elements in the matrix; let For one The matrix of The elements in Indicates that , ,if , ,otherwise , using the diagonal average formula to convert the matrix Y into a sequence , ; (6); By diagonal averaging, the original load sequence is decomposed into The group length is The sequence of , i.e., the symplectic geometric modal components; Calculate the correlation coefficient of each modal component decomposed by SGMD, assign weights according to the coefficients, output each modal component, and superimpose the prediction results of each component with weights to output the final prediction result.

[0008] Furthermore, in the step 2, it specifically includes: The power load influencing factors include temperature, humidity, cloud type, precipitation, dew point, wind speed, wind direction, air pressure, and surface albedo; The strong correlation features are selected by combining the Spearman coefficient method and the Relief-f weight method.

[0009] Furthermore, in step 3, it specifically includes: The gated recurrent unit GRU is used to process the nonlinear relationship of the input sequence and capture the long-term dependency relationship in the historical power load data preprocessed in step 1. The calculation process is as follows: (7); in, , , , , , They are input information, reset gate state, update gate state, network memory variables at the previous moment, network memory variables at the current moment, and candidate set hidden state; is the weight matrix corresponding to the gate and candidate set; The attention mechanism is introduced, and the strongly correlated features selected in step 2 are used as input. Different weights are assigned to the GRU output state to highlight the key features related to the power load. The conversion process of the attention mechanism is as follows: (8); (9); (10); in, and are the attention probability distribution and attention weight respectively; and is the weight matrix; is the bias coefficient; is the summation term of the summation function; For the attention layer Output at the moment; Use IVYA algorithm to iteratively optimize the hyperparameters in GRU; Finally, the IVYA-GRU-Attention power load prediction model is constructed based on the optimal hyperparameter combination of GRU.

[0010] Furthermore, the specific steps of iteratively optimizing the hyperparameters in GRU using the IVYA algorithm are as follows: The optimization parameters are the initial learning rate of GRU and the number of neurons in the hidden layer; Initialize the parameter values ​​of the IVYA algorithm, including the lower limit of the search space , the lower limit of the search space , maximum number of iterations , Dimension and population size ; By Calculate the initial number of ivy plants ; Among them, the vector and They represent the discrete-time system at time and time growth rate; A random number representing a random variable whose probability density function is equal to ; The dimension is A random vector whose components are random numbers from a standard Gaussian distribution; Indicates a population member; Represents the total number of populations in the search space; Represents the element-wise product of two vectors; Substitution , by formula calculate ;in, For new members The growth rate, represents Hadamard division; Calculate the fitness objective function value of all members in the initial population and set ;in, Represents a vector set Sort from best to worst; Calculate random parameters ; By calculate hour The value of like , then the ivy expands the width of its leaves, ;in, is the fitness objective function; For population members Find the nearest neighbor by roaming the search space globally; like , the ivy grows upward, climbing, ; Repeat the iterative optimization of the hyperparameters in GRU until the maximum number of iterations is reached; Extract the optimal solution.

[0011] Furthermore, in step 4, it specifically includes: The power load data set is divided into a training set and a test set. 70% of the data set is used as the training set and 30% of the data set is used as the test set. The power load forecasting model is called to predict the power load for the next month, and the predicted value of each component is output. The prediction results of each component are superimposed to output the final prediction result.

[0012] Furthermore, in step 5, it specifically includes: The mean absolute percentage error, root mean square error and goodness of fit value are selected as evaluation indicators of the prediction accuracy of the power load model; When the evaluation indicator is MAE mean absolute percentage error, the closer the indicator value is to 0, the more accurate the model prediction result is; When the evaluation index is RMSE root mean square error, the closer the error value is to 0, the smaller the error is, and the smaller the dispersion between the predicted value and the true value is; The evaluation index is When the goodness of fit value is close to 1, the better the model prediction effect is.

[0013] From the above description of the present invention, it can be seen that, compared with the prior art, the forecasting and evaluation method for short-term power load of the present invention has the following beneficial effects: 1. The present invention uses the SGMD decomposition algorithm to decompose the original load signal into several IMFS, which can better handle the nonlinearity and instability of the load sequence, weight the correlation coefficients of each modal component, highlight the key information of the original power load sequence, and effectively improve the accuracy of load forecasting; 2. The present invention combines two different feature selection algorithms, the Spearman coefficient method and the Relief-f weight method, to screen various features that affect power load; 3. The present invention combines the GRU neural network with the attention mechanism to build a GRU-Attention model, which can not only utilize the GRU's ability to process complex relationships, but also use the attention mechanism to assign different weights to different input features, highlighting the key features related to the power load, and use the IVYA algorithm to optimize the hyperparameters of the GRU network. The optimization is used to obtain the best prediction model for load forecasting, thereby further improving the accuracy of load forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 The present invention is a flowchart of the steps of a method for predicting and evaluating short-term power load in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0015] The technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all of the embodiments. Based on the embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0016] Reference Figure 1 As shown, a preferred embodiment of the present invention is a method for predicting and evaluating short-term power load, which specifically includes the following steps: Step 1, obtaining relevant historical power load data and corresponding power load influencing factors, and preprocessing and decomposing the historical power load data; Step 2, analyzing the power load influencing factors and selecting influencing factors with strong correlation as strong correlation features; Step 3: Introduce the attention mechanism, use the IVYA algorithm to iteratively optimize the hyperparameters in GRU, and build the IVYA-GRU-Attention power load forecasting model based on the best hyperparameter combination of GRU; Step 4, inputting the decomposed historical power load data and the strong correlation features into the power load forecasting model to obtain the forecast results of each component; superimposing the obtained forecast results of each component to obtain the final forecast result; Step 5: Evaluate the prediction results.

[0017] As a preferred embodiment of the present invention, it may also have the following additional technical features: In this embodiment, step 1 specifically includes: Step 1.1, preprocessing the historical power load data; The historical power load data and the power load influencing factor data are normalized: (1); in, are the original load value and the influencing factor value, are the normalized load values ​​and influencing factor values, and are the maximum and minimum values ​​of the original load data and influencing factor data respectively; Step 1.2, decomposing the preprocessed historical power load data; SGMD decomposition is used to decompose the load sequence into multiple modal components IMF. The specific algorithm is as follows: Phase space reconstruction: Assuming the timing signal , according to the embedded idea of ​​Takens theorem: (2); in, is the length of the timing signal, is the difference dimension, is the delay factor, for The number of dimensional vectors, ; Symplectic geometry matrix transformation: Assumptions Construct Hamiltonian matrix : (3); make , construct symplectic geometry matrix : (4); in, is an upper triangular matrix, the elements of which are ; matrix The characteristic value of ;matrix The characteristic value of , and the corresponding eigenvector ; Original phase space reconstruction matrix , each single-component matrix composition; (5); in, ; ;

[0018] Diagonal average: For any , Be the elements in the matrix; let For one The matrix of The elements in Indicates that , ,if , ,otherwise , using the diagonal average formula to convert the matrix Y into a sequence , ; (6); By diagonal averaging, the original load sequence is decomposed into The group length is The sequence of , i.e., the symplectic geometric modal components; Calculate the correlation coefficient of each modal component decomposed by SGMD, assign weights according to the coefficients, output each modal component, and superimpose the prediction results of each component with weights to output the final prediction result.

[0019] In this embodiment, step 2 specifically includes: The power load influencing factors include temperature, humidity, cloud type, precipitation, dew point, wind speed, wind direction, air pressure, and surface albedo; The strong correlation features are selected by combining the Spearman coefficient method and the Relief-f weight method.

[0020] In this embodiment, step 3 specifically includes: The gated recurrent unit GRU is used to process the nonlinear relationship of the input sequence and capture the long-term dependency relationship in the historical power load data preprocessed in step 1. The calculation process is as follows: (7); in, , , , , , They are input information, reset gate state, update gate state, network memory variables at the previous moment, network memory variables at the current moment, and candidate set hidden state; is the weight matrix corresponding to the gate and candidate set; The attention mechanism is introduced, and the strongly correlated features selected in step 2 are used as input. Different weights are assigned to the GRU output state to highlight the key features related to the power load. The conversion process of the attention mechanism is as follows: (8); (9); (10); in, and are the attention probability distribution and attention weight respectively; and is the weight matrix; is the bias coefficient; is the summation term of the summation function; For the attention layer Output at the moment; Use IVYA algorithm to iteratively optimize the hyperparameters in GRU; Finally, the IVYA-GRU-Attention power load prediction model is constructed based on the optimal hyperparameter combination of GRU.

[0021] In this embodiment, the specific steps of iteratively optimizing the hyperparameters in the GRU using the IVYA algorithm are: The optimization parameters are the initial learning rate of GRU and the number of neurons in the hidden layer; Initialize the parameter values ​​of the IVYA algorithm, including the lower limit of the search space , the lower limit of the search space , maximum number of iterations , Dimension and population size ; By Calculate the initial number of ivy plants ; Among them, the vector and They represent the discrete-time system at time and time growth rate; A random number representing a random variable whose probability density function is equal to ; The dimension is A random vector whose components are random numbers from a standard Gaussian distribution; Indicates a population member; Represents the total number of populations in the search space; Represents the element-wise product of two vectors; Substitution , by formula calculate ;in, For new members The growth rate, represents Hadamard division; Calculate the fitness objective function value of all members in the initial population and set ;in, Represents a vector set Sort from best to worst; Calculate random parameters ; By calculate hour The value of like , then the ivy expands the width of its leaves, ;in, is the fitness objective function; For population members Find the nearest neighbor by roaming the search space globally; like , the ivy grows upward, climbing, ; Repeat the iterative optimization of the hyperparameters in GRU until the maximum number of iterations is reached; Extract the optimal solution.

[0022] In this embodiment, step 4 specifically includes: The power load data set is divided into a training set and a test set. 70% of the data set is used as the training set and 30% of the data set is used as the test set. The power load forecasting model is called to predict the power load for the next month, and the predicted value of each component is output. The prediction results of each component are superimposed to output the final prediction result.

[0023] In this embodiment, step 5 specifically includes: The mean absolute percentage error, root mean square error and goodness of fit value are selected as evaluation indicators of the prediction accuracy of the power load model; When the evaluation indicator is MAE mean absolute percentage error, the closer the indicator value is to 0, the more accurate the model prediction result is; When the evaluation index is RMSE root mean square error, the closer the error value is to 0, the smaller the error is, and the smaller the dispersion between the predicted value and the true value is; The evaluation index is R 2 When the goodness of fit value is close to 1, the better the model prediction effect is.

[0024] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and improved concepts of the present invention within the technical scope disclosed by the present invention, and they should be covered by the protection scope of the present invention.

Claims

1. A method for predicting and evaluating short-term power load, characterized in that: The specific steps include: Step 1, obtaining relevant historical power load data and corresponding power load influencing factors, and preprocessing and decomposing the historical power load data; Step 2, analyzing the power load influencing factors and selecting influencing factors with strong correlation as strong correlation features; Step 3: Introduce the attention mechanism, use the IVYA algorithm to iteratively optimize the hyperparameters in GRU, and build the IVYA-GRU-Attention power load forecasting model based on the best hyperparameter combination of GRU; Step 4, inputting the decomposed historical power load data and the strong correlation features into the power load forecasting model to obtain the forecast results of each component; The prediction results of each component are superimposed to obtain the final prediction result; Step 5: Evaluate the prediction results.

2. A method for predicting and evaluating short-term power load according to claim 1, characterized in that: In the step 1, it specifically includes: Step 1.1, preprocessing the historical power load data; The historical power load data and the power load influencing factor data are normalized: （1）； in, are the original load value and the influencing factor value, are the normalized load values ​​and influencing factor values, and are the maximum and minimum values ​​of the original load data and influencing factor data respectively; Step 1.2, decomposing the preprocessed historical power load data; SGMD decomposition is used to decompose the load sequence into multiple modal components IMF. The specific algorithm is as follows: Phase space reconstruction: Assuming the timing signal , according to the embedded idea of ​​Takens theorem: （2）； in, is the length of the timing signal, is the difference dimension, is the delay factor, for The number of dimensional vectors, ; Symplectic geometry matrix transformation: Assumptions Construct Hamiltonian matrix : （3）； make , construct symplectic geometry matrix : （4）； in, is an upper triangular matrix, the elements of which are ; matrix The characteristic value of ;matrix The characteristic value of , and the corresponding eigenvector ; Original phase space reconstruction matrix , each single-component matrix composition; （5）； in, ; ; Diagonal average: For any , Be the elements in the matrix; let For one The matrix of The elements in Indicates that , ,if , ,otherwise , using the diagonal average formula to convert the matrix Y into a sequence , ; （6）； By diagonal averaging, the original load sequence is decomposed into The group length is The sequence of , i.e., the symplectic geometric modal components; Calculate the correlation coefficient of each modal component decomposed by SGMD, assign weights according to the coefficients, output each modal component, and superimpose the prediction results of each component with weights to output the final prediction result.

3. A method for predicting and evaluating short-term power load according to claim 1, characterized in that: In the step 2, it specifically includes: The power load influencing factors include temperature, humidity, cloud type, precipitation, dew point, wind speed, wind direction, air pressure, and surface albedo; The strong correlation features are selected by combining the Spearman coefficient method and the Relief-f weight method.

4. A method for predicting and evaluating short-term power load according to claim 1, characterized in that: In the step 3, it specifically includes: The gated recurrent unit GRU is used to process the nonlinear relationship of the input sequence and capture the long-term dependency relationship in the historical power load data preprocessed in step 1. The calculation process is as follows: （7）； in, , , , , , They are input information, reset gate state, update gate state, network memory variables at the previous moment, network memory variables at the current moment, and candidate set hidden state; is the weight matrix corresponding to the gate and candidate set; The attention mechanism is introduced, and the strongly correlated features selected in step 2 are used as input. Different weights are assigned to the GRU output state to highlight the key features related to the power load. The conversion process of the attention mechanism is as follows: （8）； （9）； （10）； in, and are the attention probability distribution and attention weight respectively; and is the weight matrix; is the bias coefficient; is the summation term of the summation function; For the attention layer Output at the moment; Use IVYA algorithm to iteratively optimize the hyperparameters in GRU; Finally, the IVYA-GRU-Attention power load prediction model is constructed based on the optimal hyperparameter combination of GRU.

5. A method for predicting and evaluating short-term power load according to claim 4, characterized in that: The specific steps of iteratively optimizing the hyperparameters in GRU using the IVYA algorithm are as follows: The optimization parameters are the initial learning rate of GRU and the number of neurons in the hidden layer; Initialize the parameter values ​​of the IVYA algorithm, including the lower limit of the search space , the lower limit of the search space , maximum number of iterations , Dimension and population size ; By Calculate the initial number of ivy plants ; Among them, the vector and They represent the discrete-time system at time and time growth rate; A random number representing a random variable whose probability density function is equal to ; The dimension is A random vector whose components are random numbers from a standard Gaussian distribution; Indicates a population member; Represents the total number of populations in the search space; Represents the element-wise product of two vectors; Substitution , by the formula calculate ;in, For new members The growth rate, represents Hadamard division; Calculate the fitness objective function value of all members in the initial population and set ;in, Represents a vector set Sort from best to worst; Calculate random parameters ; By calculate hour The value of like , then the ivy expands the width of its leaves, ;in, is the fitness objective function; For population members Find the nearest neighbor by roaming the search space globally; like , the ivy grows upward, climbing, ; Repeat the iterative optimization of the hyperparameters in GRU until the maximum number of iterations is reached; Extract the optimal solution.

6. A method for predicting and evaluating short-term power load according to claim 1, characterized in that: In the step 4, it specifically includes: The power load data set is divided into a training set and a test set. 70% of the data set is used as the training set and 30% of the data set is used as the test set. The power load forecasting model is called to predict the power load for the next month, and the predicted value of each component is output. The prediction results of each component are superimposed to output the final prediction result.

7. A method for predicting and evaluating short-term power load according to claim 1, characterized in that: In the step 5, it specifically includes: The mean absolute percentage error, root mean square error and goodness of fit value are selected as evaluation indicators of the prediction accuracy of the power load model; When the evaluation indicator is MAE mean absolute percentage error, the closer the indicator value is to 0, the more accurate the model prediction result is; When the evaluation index is RMSE root mean square error, the closer the error value is to 0, the smaller the error is, and the smaller the dispersion between the predicted value and the true value is; The evaluation index is When the goodness of fit value is close to 1, the better the model prediction effect is.

Images