A multi-model fusion power daily load online prediction method and system

Abstract

The application relates to a multi-model fusion power daily load online prediction method, which adopts a self-encoder and a moving average to perfect a data set, adopts a unified information coefficient to extract high correlation features, adopts a hiking optimization algorithm to realize KELM model parameter selection, determines single-day load prediction by using a KELM model, obtains next-day prediction results based on a rolling strategy, and finally uses a multi-index evaluation model to predict results.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting, and more specifically to an online method for forecasting daily power load using a multi-model fusion approach. Background Technology

[0002] With the development of the social economy and the improvement of living standards, the demand for electricity is showing an increasing trend. Especially during peak hours, the fluctuation of power load places higher demands on the stability of the power grid. Daily load forecasting is a key task in power system management. It involves predicting the power load for the next day to ensure the reliability and stability of power supply. Accurate daily load forecasting can not only help power companies effectively dispatch power resources, but also optimize power grid operation, improve energy efficiency, and reduce operating costs. Summary of the Invention

[0003] The technical problem of this invention is: With the rapid development of machine learning and deep learning technologies, more and more prediction models are being introduced into the field of load forecasting. However, the actual application of these models has affected the reliability and stability of power load forecasting due to factors such as data quality, model parameter selection, and computational complexity.

[0004] The purpose of this invention is to solve the above-mentioned problems and proposes a multi-model fusion method for online forecasting of daily power load. Specifically, it includes: using an autoencoder and moving average to improve the dataset, using a unified information coefficient to extract highly correlated features, using a walking optimization algorithm to select KELM model parameters, using the KELM model to determine the daily load forecast, obtaining the next day's forecast results based on a rolling strategy, and finally using multiple indicators to evaluate the model forecast results.

[0005] The technical solution of this invention is as follows: A multi-model fusion method for online forecasting of daily electricity load includes the following steps: S1: An autoencoder is used to detect outliers in the collected historical daily load data, and moving average is used to fill in the outliers to obtain the initial dataset D1; S2: Use the unified information coefficient to extract the feature information of the power load in dataset D1 to obtain the input dataset D2; S3: Divide the input dataset D2 into training set, test set and validation set in a fixed ratio, use the training set and hiking optimization algorithm to select KELM model parameters, and use KELM model to determine daily load forecast; S4: Keep the sample size unchanged, update the input dataset D2 using a rolling strategy, and repeat steps S1~S3 until the prediction results for the next day are obtained; S5: Use multiple indicators to evaluate the prediction results and adjust the initial parameters synchronously based on the prediction results.

[0006] Furthermore, in step S1, due to problems such as acquisition system failure and transmission delay, there are outliers in the database. An autoencoder is used for detection, which specifically includes encoder training, reconstruction error calculation and outlier detection. Preferably, the outlier detection principle of the autoencoder is as follows: 1) The encoder is used to map the input data to a low-dimensional representation, and the decoder is used to restore the low-dimensional representation back to the original data space; 2) The training process is completed by minimizing the difference between the input data and the reconstructed data, using mean squared error as the loss function; 3) Use an autoencoder to encode and decode the new data, and calculate the reconstruction error for each data point; 4) Compare the reconstruction error with the threshold to identify outliers.

[0007] Preferably, in step S1, a moving average is used to fill in the detected outliers, as shown in the following formula:

[0008] In the formula, o t For data to be interpolated, r For the size of the time window data, o i This refers to the known data within the window.

[0009] Furthermore, in step S2, characteristic information of the power load is extracted using a unified information coefficient, including the following sub-steps: S21: Mutual information calculation:

[0010] In the formula, For mutual information, p ( x , y ) is the joint probability distribution function; p ( x )and p ( y ) are respectively X and Y The marginal probability distribution function, X and Y These are two sets of feature vectors. x and y For feature vectors X and Y The corresponding element.

[0011] S22: Variable partitioning:

[0012]

[0013] In the formula, L x and L y They are respectively X and Y The length of the partition unit; x max and x min Variables X The maximum and minimum values; y max and y min Variables Y The maximum and minimum values; m and n They are respectively X and Y The number of segments; This refers to the size of the partition grid. k Indicates the sample size.

[0014] S23: Calculation of Unified Information Coefficient:

[0015] In the formula, To represent unified information, m and n They are respectively X and Y The number of segments, min represents taking the minimum value.

[0016] Furthermore, in step S3, the training set is used for the model to learn patterns and features from the data and to adjust the model's internal parameters; the validation set is used to test the model to prevent overfitting; and the test set is used to finally evaluate the model's performance and test its generalization ability on new, unseen data.

[0017] Preferably, the hiking optimization method is based on the Tobler hiking function, which determines the hiker's walking speed by considering the terrain elevation and walking distance, including: 1) Initialize the hiker positions, evaluate the best hiker position, and define it as the leader; 2) Calculate the hiker's initial speed based on the terrain slope, including the following calculation formula:

[0018]

[0019] In the formula, dh and dx These represent the altitude and horizontal distance traveled by the hiker, respectively. The slope angle of the terrain. Hiker i In the t The speed of each iteration; The slope of the terrain. e is the base of the natural logarithm function.

[0020] 3) Update hiker speeds by leveraging the social thinking of the traveler community and the personal cognition of individual travelers, as detailed below:

[0021] In the formula, A random number in the range [0,1]. Let i be the velocity of traveler i in the t-th iteration; It is the position of team leader; Let i be the position of traveler i in the t-th iteration; It is the scan factor of traveler i, and the scan factor is in [1,2]. The scan factor ensures that the traveler does not stray too far from the leader, so that they can see the leader's direction and receive the leader's signals.

[0022] 4) Update hiker locations as follows:

[0023] In the formula, For travelers i In the t The position of +1 iteration; 5) Update the fitness value of the hiking group and perform iterative convergence judgment. Once the current iteration number reaches the maximum iteration number, the iteration process ends.

[0024] Preferably, the KELM model includes an ELM model and a kernel function, including: 1) ELM Prediction Output F ( x The following is represented:

[0025] In the formula, q For the input vector, H ( q ) represents the output matrix of the hidden nodes. Indicates the output weights; 2) Introducing regularization coefficients and the identity matrix, and transforming the solution process, the least squares solution for the output weights is:

[0026] In the formula, G For the target vector, for The least squares solution, I It is the identity matrix. C This represents the regularization coefficient.

[0027] 3) Change the random mapping to a kernel mapping. The kernel matrix expression is as follows:

[0028] In the formula, Nuclear element, For kernel function, q i and q j This represents the experimental input vector.

[0029] 4) Choose the Gaussian kernel function, as shown in the formula below:

[0030] In the formula, For kernel parameters, q i and q j Represents the experimental input vector. exp It is a natural exponential function.

[0031] 5) The final output is:

[0032] In the formula, n For the sample size, I It is the identity matrix. C The regularization coefficient is . K For kernel function, Nuclear element, T It is a transpose operator. G For the target vector, for KELM Output.

[0033] Furthermore, in step S4, the time range of the input dataset is the [missing information]. d From the 1st to the 1st d + N Daily load data, forecast target is the first d + N +1 day load data. As the dataset continues to grow, a rolling strategy is adopted to update the input dataset over a time range of [number missing]. d+ 1st d + N+ The load data for day 1, the corresponding forecast target is the first day. d +N+ Two days of load data, to balance model training time and maintain model generalization, are represented as follows:

[0034] In the formula, d A day contains multiple sampling times. N For the sample size, Y d , Y d+1 , Y d+2 , Y d+N , Y d+N+1 Represented as the first d , No. d +1, No. d +2, No. d + N and the d + N +1 day's load.

[0035] Furthermore, in step S5, a multi-index evaluation method is used to assess the model prediction results, including root mean square error, explained variance score, and mean absolute percentage error, including: 1) Root mean square error:

[0036] In the formula, The root mean square error, N Indicates sample size; The actual value; This is a predicted value; 2) Explained variance score

[0037] In the formula, For the explainable variance score, var(·) represents the variance; 3) Mean absolute percentage error

[0038] In the formula: The mean absolute percentage error, N Indicates sample size; The actual value; These are predicted values.

[0039] Preferably, the root mean square error is used to measure the deviation between the predicted value and the true value, and its value range is [value range missing]. The smaller the value, the smaller the prediction error and the stronger the predictive ability of the model. The Explained Variance Score (EVS) ranges from 0 to 1. When EVS is 1, it means the model perfectly predicted the data; conversely, it means the model explains less of the data. The Mean Absolute Percentage Error (MAPE) ranges from [value missing]. A MAPE greater than 100% indicates a poor model, while a MAPE less than 100% indicates a model with high prediction accuracy.

[0040] Compared with the prior art, the beneficial effects of the present invention include: 1) This invention discloses an online forecasting method for daily power load based on multi-model fusion. It uses an autoencoder and moving average to improve the initial dataset, uses a unified information coefficient to extract highly correlated input features, and uses a walking optimization algorithm to select KELM model parameters. It uses the KELM model to determine the daily load forecast, obtains the forecast results for the next day based on a rolling strategy, and finally uses multiple indicators to evaluate the model forecast results. 2) Considering that errors are inevitable during data acquisition and transmission, resulting in a slight deviation between the acquired data and the real data, the data recognition strategy based on the autoencoder of this invention can effectively capture outliers in the data and effectively avoid data omissions and false detections. 3) Different daily load sequences have similar variation patterns, that is, there is daily periodicity. The moving average interpolation technique of this invention can make the interpolated sequence also have similar periodicity. 4) Unlike common correlation coefficients, the unified information coefficient of this invention can measure both linear and nonlinear correlations between variables, and has a certain robustness to noisy data, which is beneficial for feature variable extraction. 5) This invention uses a walking optimization algorithm to select parameters for the prediction model, which improves the prediction accuracy and ability to handle nonlinear problems of the KELM model, and has high adaptability and stability. 6) Existing prediction models often achieve the prediction of future load demand through a single training, ignoring the dynamic changes in load sequences with social development. This invention can reduce the computational burden and avoid model overfitting by updating the input dataset in real time while maintaining the sample size. Attached Figure Description

[0041] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0042] Figure 1 This is a flowchart illustrating the power load forecasting process according to an embodiment of the present invention. Figure 2 This is a schematic diagram of an autoencoder according to an embodiment of the present invention; Figure 3 This is a schematic diagram comparing the actual and predicted power load values ​​of each model in the embodiments of the present invention; Figure 4 This is a system module structure diagram of an embodiment of the present invention. Detailed Implementation

[0043] Example 1: like Figure 1 As shown, the power load of New England from January 1, 2024 to June 31, 2024 was simulated and predicted according to the method of the present invention. The total sample size was 4368, the sampling frequency was 1 hour, and the characteristic variables included marginal price, dry bulb temperature and wet bulb temperature.

[0044] A multi-model fusion method for online forecasting of daily electricity load includes the following steps: S1: Anomaly detection is performed on the collected historical daily load data based on an autoencoder, and the outliers are filled using a moving average to obtain the initial dataset D1; S2: Use the unified information coefficient to extract the feature information of the power load in dataset D1 to obtain the feature dataset D2; S3: Divide the feature dataset D2 into training set, test set and validation set according to a fixed ratio, build a load forecasting model using KELM model, select KELM model parameters using training set and walking optimization algorithm, and use KELM model to determine daily load forecast; S4: Keep the sample size unchanged, update the dataset D2 using a rolling strategy, and repeat steps S1~S3 until the prediction results for the next day are obtained; S5: Use multiple indicators to evaluate the prediction results and adjust the initial parameters synchronously based on the prediction results.

[0045] Furthermore, in step S1, due to problems such as acquisition system failure and transmission delay, there are outliers in the database. An autoencoder is used for detection, which specifically includes encoder training, reconstruction error calculation and outlier detection. Preferred, Figure 2 This is a schematic diagram of an autoencoder structure. The outlier detection principle is as follows: 1) The encoder is used to map the input data to a low-dimensional representation, and the decoder is used to restore the low-dimensional representation back to the original data space; 2) The training process is completed by minimizing the difference between the input data and the reconstructed data, using mean squared error as the loss function; 3) Use an autoencoder to encode and decode the new data, and calculate the reconstruction error for each data point; 4) Compare the reconstruction error with the threshold to identify outliers.

[0046] Preferably, in step S1, a moving average is used to fill in the detected outliers, as shown in the following formula:

[0047] In the formula, o t For data to be interpolated, r For the size of the time window data, o i This refers to the known data within the window.

[0048] Furthermore, in step S1, the abnormal data is filled using a moving average. The load data of the day before the day in which the abnormal value is located is extracted, and the average load of the day before the day is calculated as the filling value.

[0049] Furthermore, in step S2, characteristic information of the power load is extracted using a unified information coefficient, including the following sub-steps: S21: Mutual information is calculated as follows:

[0050] In the formula, For mutual information, p ( x , y ) is the joint probability distribution function; p ( x )and p ( y ) are respectively X and Y The marginal probability distribution function, X and Y These are two sets of feature vectors. x and y For feature vectors X and Y The corresponding element.

[0051] S22: The results of variable partitioning are as follows:

[0052]

[0053] In the formula, L x and L y They are respectively X and Y The length of the partition unit; x max and x min Variables X The maximum and minimum values; y max and y min Variables YThe maximum and minimum values; m and n They are respectively X and Y The number of segments; This refers to the size of the partition grid. k Indicates the sample size.

[0054] S23: Calculation of Unified Information Coefficient:

[0055] In the formula, To represent unified information, m and n They are respectively X and Y The number of segments, min represents taking the minimum value.

[0056] Furthermore, in step S3, the training set is used for the model to learn patterns and features from the data and to adjust the model's internal parameters; the validation set is used to prevent the model from overfitting; and the test set is used to finally evaluate the model's performance and test its generalization ability on new, unseen data.

[0057] Preferably, the hiking optimization method is based on the Tobler hiking function, which determines the hiker's walking speed by considering the terrain elevation and walking distance, including the following calculation formula: 1) Initialize the hiker positions, evaluate the best hiker position, and define it as the leader; 2) Calculate the hiker's initial speed based on the terrain slope, including the following calculation formula:

[0058]

[0059] In the formula, dh dx and dx represent the elevation and horizontal distance traveled by the hiker, respectively. The slope angle of the terrain. It is the speed of hiker i in the t-th iteration; Let be the slope of the terrain, and e be the base of the natural logarithm function.

[0060] 3) The hiking speed is updated by combining the social thinking of the traveler group with the personal cognition of the individual traveler. The calculation formula is as follows:

[0061] In the formula, A random number in the range [0,1]. Let i be the velocity of traveler i in the t-th iteration; It is the position of team leader; Let i be the position of traveler i in the t-th iteration; It is the scan factor of traveler i, and the scan factor is in [1,2]. The scan factor ensures that the traveler does not stray too far from the leader, so that they can see the leader's direction and receive the leader's signals.

[0062] 4) Furthermore, the formula for updating hiker locations is as follows:

[0063] In the formula, For travelers i In the t The position of +1 iteration; 5) Update the fitness value of the hiking group and perform iterative convergence judgment. Once the current iteration number reaches the maximum iteration number, the iteration process ends.

[0064] Preferably, the KELM model includes an ELM model and a kernel function, and includes the following calculation formulas: 1) ELM Prediction Output F ( q The following is represented:

[0065] In the formula, q For the input vector, H ( x ) represents the output matrix of the hidden nodes. Indicates the output weights; 2) Introducing regularization coefficients and the identity matrix, and transforming the solution process, the least squares solution for the output weights is:

[0066] In the formula, G For the target vector, for The least squares solution, I It is the identity matrix. C This represents the regularization coefficient.

[0067] 3) Change the random mapping to a kernel mapping. The kernel matrix expression is as follows:

[0068] In the formula, Nuclear element, For kernel function, q i and q j This represents the experimental input vector.

[0069] 4) Choose the Gaussian kernel function, as shown in the formula below:

[0070] In the formula, For kernel parameters, q i and q j Represents the experimental input vector. exp It is a natural exponential function.

[0071] 5) The final output calculation formula is:

[0072] In the formula, n For the sample size, I Represents the identity matrix. C Represents the regularization coefficient. K For kernel function, Nuclear element, T It is a transpose operator. G For the target vector, for KELM Output.

[0073] Furthermore, in step S4, the time range of the input dataset is the [missing information]. d From the 1st to the 1st d + N Daily load data, forecast target is the first d + N +1 day load data. As the dataset continues to grow, a rolling strategy is adopted to update the input dataset on the time scale of the +1 day load. d+ From the 1st to the d + N+ Based on the load data for day 1, the forecast target is changed accordingly to day 2. d + N+ Two days of load data, to balance model training time and maintain model generalization, are represented as follows:

[0074] In the formula, d A day contains multiple sampling times. N For the sample size, Y d , Y d+1 , Y d+2 , Y d+N , Y d+N+1 Represented as the first d , No. d +1, No.d +2, No. d + N and the d + N +1 day's load.

[0075] Furthermore, in step S5, a multi-index evaluation method is used to assess the model prediction results, including root mean square error, explained variance score, and mean absolute percentage error, including the following calculation formulas: 1) Root mean square error:

[0076] In the formula, The root mean square error, N For sample size; The actual value; These are predicted values.

[0077] 2) Explained variance score

[0078] In the formula, Explainable variance score var Let Variance be the variance.

[0079] 3) Mean absolute percentage error

[0080] In the formula: The mean absolute percentage error, N Indicates sample size; The actual value; These are predicted values.

[0081] Preferably, the root mean square error is used to measure the deviation between the predicted value and the true value, and its value range is [value range missing]. The smaller the value, the smaller the prediction error and the stronger the predictive ability of the model. The Explained Variance Score (EVS) ranges from 0 to 1. When EVS is 1, it means the model perfectly predicted the data; conversely, it means the model explains less of the data. The Mean Absolute Percentage Error (MAPE) ranges from [value missing]. A MAPE greater than 100% indicates a poor model, while a MAPE less than 100% indicates a model with high prediction accuracy.

[0082] like Figure 3As shown, all different models achieved daily load forecasting, with identical input data and a forecast timescale of 24 sampling points, representing daily load demand. RNNs suffer from vanishing and exploding gradients, resulting in significant prediction errors. LSTM and GRU, further optimized from RNNs, effectively improved load forecasting performance through gate structure improvements. Kernel Extreme Learning Machines (KELMs) introduced a Gaussian function to enhance nonlinear mapping capabilities, demonstrating significant advantages in handling complex nonlinear relationships. The prediction results optimized for KELM using HOA were closest to the true values, exhibiting the highest load forecasting accuracy. This indicates that effective parameter selection plays a crucial role in improving load forecasting performance and further underscores the necessity of parameter selection. As shown in Table 1, compared with the five single prediction models RNN, LSTM, GRU, ELM and KELM, the RMSE was reduced by 36.25%, 35.61%, 30.57%, 40.24% and 28.34% respectively. The smaller the value, the higher the model accuracy. The MAPE was reduced by 33.95%, 32.19%, 21.05%, 39.09% and 18.26% respectively. Similar to RMSE, the smaller the value, the better. The EVS was increased by 0.45%, 0.89%, 1.09%, 1.09% and 0.49% respectively. Unlike the above two indicators, the larger the value, the more data variance the model explains.

[0083] The embodiments of the present invention employ an autoencoder to detect anomalies in load data, a KELM model to construct a load prediction model, a walking optimization algorithm to optimize the parameter selection of the model, a rolling strategy to update the input dataset, and multiple indicators to evaluate the model performance, thereby obtaining online power load prediction results and producing a technical effect greater than the sum of its parts.

[0084] Table 1 Performance of Power Load Forecasting for Each Model

[0085] Example 2: This invention discloses an online daily power load forecasting system based on multi-model fusion, such as... Figure 4 As shown, it includes: Data detection and imputation module: An autoencoder is used to detect outliers in historical load data and provide the outlier data and its location. The outlier data is imputed by moving average to obtain the initial dataset. Feature extraction module: Calculates the uniform information coefficient between the feature variables in the initial dataset and the loading data, and selects highly correlated features and loading data to construct the input dataset; Model training module: The input dataset is divided into training, test and validation sets in a fixed ratio to train the KELM model. At the same time, the walking optimization algorithm is used to optimize the parameters to obtain the best model parameters. Daily load forecasting module: Based on the optimal model parameters, input the characteristic data of the day to be forecasted, and obtain the load data of the day to be forecasted; Data update module: The input dataset is updated using a rolling strategy, and the data detection, data imputation, feature extraction, model training and prediction processes are executed again to obtain the load prediction results for the next day; Model update module: Evaluates the forecast performance of load forecast results using multiple indicators and adjusts the initial parameter settings based on the evaluation results.

[0086] It should be noted that the above examples are only some preferred examples of the present invention. The present invention has been described in detail with reference to the above embodiments. Those skilled in the art should understand that modifications or equivalent substitutions can be made to the specific embodiments of the present invention without departing from the spirit and scope of the present invention. Any modifications or equivalent substitutions should be covered within the scope of protection of the claims of the present invention.

Claims

1. A multi-model fusion method for online forecasting of daily electricity load, characterized in that, Includes the following steps: S1: An outlier detection is performed on the collected historical daily load data using an autoencoder, and outliers are filled in using a moving average to obtain the initial dataset D1; S2: Use the unified information coefficient to extract the feature information of the power load in the initial dataset D1 to obtain the input dataset D2; The method of extracting characteristic information of power load using a unified information coefficient includes: S21: Mutual information calculation: In the formula, Indicates mutual information, p ( x , y ) is the joint probability distribution function; p ( x )and p ( y ) are respectively X and Y The marginal probability distribution function; X and Y These are two sets of feature vectors. x For feature vectors X The corresponding element, y For feature vectors Y The corresponding element; S22: Variable partitioning: In the formula, L x and L y They are respectively X and Y The length of the partition unit; x max and x min Variables X The maximum and minimum values; y max and y min Variables Y The maximum and minimum values; m and n They are respectively X and Y The number of segments; This refers to the size of the partition grid. k For sample size; S23: Calculation of Unified Information Coefficient: In the formula, To represent unified information, m and n They are respectively X and Y The number of segments, min This indicates taking the minimum value; S3: Divide the input dataset D2 into training set, test set and validation set according to a fixed ratio, build a load forecasting model using KELM model, select KELM model parameters using training set and walking optimization algorithm, and realize daily load forecasting through KELM model; The hiking optimization algorithm includes: 1) Initialize the hiker positions, evaluate the best hiker position, and define it as the leader; 2) Calculate the hiker's initial speed based on the terrain slope, including the following calculation formula: In the formula, dh and dx These represent the altitude and horizontal distance traveled by the hiker, respectively. The slope angle of the terrain. Hiker i In the t The speed of each iteration; The slope of the terrain. e is the base of the natural logarithm function; 3) Update hiker speeds by leveraging the social thinking of the traveler community and the personal cognition of individual travelers, as detailed below: In the formula, A random number in the range [0,1]. For travelers i In the t The speed of each iteration; It is the position of team leader; For travelers i In the t The position of the next iteration; Traveler i The scan factor SF is in the range [1,2]. SF ensures that the traveler does not stray too far from the leader and receives the leader's signal. 4) Further updates to hiker locations are as follows: In the formula, For travelers i In the t The position of +1 iteration; 5) Update the fitness value of the hiking group and perform an iterative convergence check. When the current iteration count reaches the maximum iteration count, the iteration process ends. S4: Keeping the sample size unchanged, update the optimized KELM with a rolling strategy to obtain the load forecast value, and repeat steps S1~S3 until the forecast result for the next day is obtained. S5: Use multiple indicators to evaluate the prediction results and adjust the initial parameters synchronously based on the prediction results.

2. In the online forecasting method for daily power load using multi-model fusion according to claim 1, in step S1, the autoencoder performs outlier detection on the collected historical daily load data, including encoder training, reconstruction error calculation, and outlier detection.

3. The online forecasting method for daily electricity load using multi-model fusion according to claim 1, wherein in step S3, the method is characterized in that... The KELM model includes an ELM model and kernel functions, including: 1) ELM Prediction Output F ( q The following is represented: In the formula, q For the input vector, H Output a matrix for the hidden nodes. For output weights; This is the output function; 2) Introducing regularization coefficients and the identity matrix, and transforming the solution process, the least squares solution for the output weights is: In the formula, G For the target vector, for The least squares solution, I Let C be the identity matrix and C be the regularization coefficient. 3) Change the random mapping to a kernel mapping. The kernel matrix expression is as follows: In the formula, Nuclear element, For kernel function, q i and q j Represents the experimental input vector; 4) Choose the Gaussian kernel function. The formula is: In the formula, For kernel parameters, q i and q j Represents the experimental input vector. exp It is a natural exponential function; 5) The final output is: In the formula, n For the sample size, I It is the identity matrix. C The regularization coefficient is . K For kernel function, Nuclear element, T It is a transpose operator. G For the target vector, for KELM Output.

4. The online forecasting method for daily electricity load using multi-model fusion according to claim 1, wherein in step S4, the method is characterized in that... The method for obtaining the next day's prediction results is as follows: the original sample size is... N Time range d ~ d + N The feature dataset D2 is updated using a rolling strategy. Specifically, the process involves removing data from D2 at time i. d The daily load data, with the time added at the end of D2, is the [number]th [period]. d + N The updated load data for +1 day is calculated using the following formula: In the formula, d A day contains multiple sampling times. N For the sample size, Y d , Y d+1 , Y d+2 , Y d+N , Y d+N+1 Represented as the first d , No. d +1, No. d +2, No. d + N and the d + N +1 day's load.

5. The online forecasting method for daily electricity load using multi-model fusion according to claim 1, wherein in step S5, the method is characterized in that... The load forecasting results from step S4 are used to evaluate the model's forecasting accuracy, specifically including: 1) Root mean square error: In the formula, The root mean square error, N For sample size; The actual value; This is a predicted value; 2) Explained variance score In the formula, Explainable variance score var For variance; 3) Mean absolute percentage error In the formula: The mean absolute percentage error, N For sample size; The actual value; These are predicted values.

6. The online power load forecasting system of the multi-model fusion method for online power load forecasting as described in claim 1, characterized in that, include: Data preprocessing module: An autoencoder is used to detect outliers in historical load data and the outliers and their locations are provided. The outliers are then interpolated using a moving average to obtain the initial dataset. Feature extraction module: Calculates the uniform information coefficient between the feature variables in the initial dataset and the loading data, and selects highly correlated features and loading data to construct the input dataset; Model training and parameter optimization module: The input dataset is divided into training set, test set and validation set according to a fixed ratio to train the KELM model. At the same time, the walking optimization algorithm is used to optimize the parameters to obtain the best model parameters. Based on the best model parameters, the feature data of the day to be predicted is input to obtain the load data of the day to be predicted. Data update module: The input dataset is updated using a rolling strategy, and the data detection, data imputation, feature extraction, model training and prediction processes are executed again to obtain the load prediction results for the next day; Evaluation and parameter tuning module: The load forecast results are evaluated using multiple indicators to assess the forecast performance, and the initial parameter settings are adjusted based on the evaluation results.

Images