Power load multi-model integrated prediction method, device, equipment and medium

Abstract

The application discloses a power load multi-model integrated prediction method and device, equipment and medium, and relates to the technical field of power load prediction, which comprises the following steps: cleaning and normalizing the historical load time series data, then adopting a strict causal sliding window mechanism to construct multi-scale features, determining the optimal parameters of a tree model, a gated recurrent single-rank regression meta-model and a bidirectional recurrent neural network through a two-stage hyperparameter joint optimization strategy and constructing corresponding models, inputting structured features, a first time series matrix and a second time series matrix for orthogonal division prediction, and finally outputting results through enhanced stacked fusion. The method realizes root-cause avoidance of data leakage, does not depend on exogenous variables, reduces the parameter tuning calculation complexity, realizes model differentiation and complementation, and improves the accuracy and robustness of power load prediction.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting technology, and in particular to a method, apparatus, equipment and medium for multi-model integrated forecasting of power load. Background Technology

[0002] To achieve power load forecasting, existing technologies have proposed various forecasting methodologies and integrated optimization schemes, mainly including: 1) Statistical model-based forecasting methods, such as ARIMA, exponential smoothing, and seasonal decomposition, which achieve forecasting by modeling the statistical characteristics of historical load data; 2) Machine learning-based forecasting methods, such as support vector regression, random forest, and gradient boosting trees, which capture the nonlinear relationship of load by manually constructing features input to the model; 3) Deep learning-based forecasting methods, such as long short-term memory networks, gated recurrent units, and temporal convolutional networks, which can learn time-dependent features end-to-end from the original load sequence; and 4) Decomposition-integration-based forecasting methods, which combine temporal decomposition algorithms with forecasting models, first decomposing the load sequence into trend, periodic, and residual components, then modeling and forecasting each component separately before reconstructing the result.

[0003] Existing power load forecasting methods all have significant limitations. First, the feature construction stage is prone to data leakage risks. Some methods use global time-series decomposition algorithms that are non-causal filters, which require the use of the overall statistical characteristics of the load sequence. This leads to the indirect leakage of future information into the training features, resulting in inflated model evaluation results and a significant drop in generalization performance during actual deployment. Second, hyperparameter optimization strategies have shortcomings. If each sub-model is tuned independently, the global optimality of the ensemble cannot be guaranteed, which can easily destroy the differentiated and complementary value of the sub-models in the integration process. If brute-force full-grid joint parameter tuning is used, the number of parameters explodes exponentially, resulting in high computational costs and making it difficult to implement in engineering. Third, feature construction suffers from homogenization. The input feature information sources of each sub-model highly overlap, making it impossible to form effective differentiated division of labor and complementarity, resulting in limited performance gains after integration.

[0004] Therefore, how to construct multi-scale features and achieve high-precision power load forecasting based on pure historical load data while avoiding the risk of data leakage caused by global decomposition has become an urgent problem to be solved. Summary of the Invention

[0005] The main objective of this application is to provide a method, apparatus, equipment, and medium for multi-model integrated forecasting of power load, aiming to solve the technical problem of how to avoid data leakage and improve the accuracy and robustness of load forecasting without relying on exogenous variables.

[0006] To achieve the above objectives, this application proposes a multi-model integrated forecasting method for power load, comprising:

[0007] The collected historical load time series data are cleaned and normalized to obtain preprocessed time series data;

[0008] Based on the preprocessed time-series data, a strict causal sliding window mechanism is used to construct multi-scale features, generating structured features, a first time-series matrix, and a second time-series matrix, respectively.

[0009] The optimal hyperparameter combination of the tree model, the gated recurrent single-ridge regression model, and the bidirectional recurrent neural network model is determined based on the two-stage hyperparameter joint optimization strategy, and the optimal tree model, the optimal gated recurrent single-ridge regression model, and the optimal bidirectional recurrent neural network model are constructed based on the optimal hyperparameter combination.

[0010] The structured features, the first time series matrix, and the second time series matrix are respectively input into the optimal tree model, the optimal gated recurrent single-ridge regression model, and the optimal bidirectional recurrent neural network model for prediction, to obtain the corresponding first prediction result, second prediction result, and third prediction result;

[0011] The first, second, and third prediction results are enhanced by stacking and fusing to obtain the power load prediction result.

[0012] In one embodiment, the step of constructing multi-scale features based on the preprocessed time-series data using a strictly causal sliding window mechanism, and generating structured features, a first time-series matrix, and a second time-series matrix respectively, includes:

[0013] Determine the preset neighbor window length, preset historical days, and preset sliding window length;

[0014] Based on the preprocessed time series data, the historical load data of continuous nearby points are extracted using a strict causal sliding window with a preset nearest neighbor window length, with the current prediction time as the endpoint, to obtain the nearest neighbor historical load characteristics.

[0015] Based on the preprocessed time series data, using the current prediction time as a benchmark, a strict causal sliding window is used to extract historical load data at the same time within a preset number of days to obtain historical synchronous periodic characteristics.

[0016] Based on the preprocessed time series data, a strict causal sliding window with a preset sliding window length is used to calculate the rolling mean, and the sliding mean feature is obtained.

[0017] By fusing the nearest-neighbor historical load features and the historical contemporaneous period features, a structured feature in the input form of the adaptive tree model is obtained;

[0018] By integrating the nearest-neighbor historical load characteristics and the moving average characteristics, a first time series matrix adapted to the input form of the gated cyclic single-ridge regression metamodel is obtained;

[0019] By integrating the nearest neighbor historical load features, the moving average features, and the historical periodic features, a second time-series matrix adapted to the input form of a bidirectional recurrent neural network model is obtained.

[0020] In one embodiment, the step of determining the optimal hyperparameter combination of the tree model, the gated recurrent single-ridge regression meta-model, and the bidirectional recurrent neural network model based on a two-stage hyperparameter joint optimization strategy, and constructing the optimal tree model, the optimal gated recurrent single-ridge regression meta-model, and the optimal bidirectional recurrent neural network model based on the optimal hyperparameter combination, includes:

[0021] A large-grid sequential coarse search is performed on the tree model, the gated recurrent single-ridge regression meta-model, and the bidirectional recurrent neural network model. The hyperparameters of the other two models are fixed to the default configuration. The optimal hyperparameter neighborhood of each individual model is searched in the large grid to obtain the optimal hyperparameter neighborhood of each of the tree model, the gated recurrent single-ridge regression meta-model, and the bidirectional recurrent neural network model after the large-grid coarse search.

[0022] Based on the optimal hyperparameter neighborhood of each model, a small grid is constructed for each model to form a small grid space for joint search of hyperparameters of the three models;

[0023] A joint fine search is performed on the three models within the small grid space, traversing all hyperparameter combinations and calculating the prediction error of each combination;

[0024] The hyperparameter combination with the smallest prediction error is selected as the optimal hyperparameter combination for the tree model, the gated recurrent single-ridge regression model, and the bidirectional recurrent neural network model, respectively.

[0025] The optimal hyperparameter combinations of each model are configured into the corresponding base model for model training to obtain the optimized tree model, the optimized gated recurrent single-ridge regression meta-model, and the optimized bidirectional recurrent neural network model.

[0026] The performance of the optimized tree model, the optimized gated recurrent single-ridge regression meta-model, and the optimized bidirectional recurrent neural network model were tested, and the test results were obtained.

[0027] When the test results show that the optimized tree model, optimized gated cyclic single-ridge regression model, and optimized bidirectional recurrent neural network model all meet the preset requirements, the current optimized tree model, optimized gated cyclic single-ridge regression model, and optimized bidirectional recurrent neural network model are determined as the optimal tree model, optimal gated cyclic single-ridge regression model, and optimal bidirectional recurrent neural network model.

[0028] In one embodiment, the step of inputting the structured features, the first time series matrix, and the second time series matrix into the optimal tree model, the optimal gated recurrent single-ridge regression meta-model, and the optimal bidirectional recurrent neural network model for prediction, respectively, to obtain the corresponding first prediction result, second prediction result, and third prediction result, includes:

[0029] The structured features are input into the optimal tree model, and the baseline load regression prediction is performed to obtain the first initial prediction result;

[0030] The first time series matrix is ​​input into the optimal gated cyclic single-ridge regression model to perform short-term load fluctuation fitting prediction and obtain the second initial prediction result.

[0031] The second time series matrix is ​​input into the optimal bidirectional recurrent neural network model to perform long-term load cycle memory prediction and obtain the third initial prediction result.

[0032] The first initial prediction result, the second initial prediction result, and the third initial prediction result are format-standardized to obtain the first prediction result, the second prediction result, and the third prediction result.

[0033] In one embodiment, the step of enhancing and stacking the first prediction result, the second prediction result, and the third prediction result to obtain the power load prediction result includes:

[0034] The arithmetic mean of the three prediction results is calculated based on the first prediction result, the second prediction result, and the third prediction result to obtain the prediction mean feature.

[0035] Based on the first prediction result, the second prediction result and the third prediction result, the standard deviation of the three prediction results is calculated to obtain the prediction standard deviation feature that characterizes the model's degree of divergence.

[0036] The first prediction result, the second prediction result, the third prediction result, the predicted mean feature, and the predicted standard deviation feature are concatenated according to a preset dimension to obtain a five-dimensional feature vector;

[0037] The five-dimensional feature vector is input into the ridge regression model and solved using a robust piecewise fusion weight learning strategy to obtain the fusion weight coefficients.

[0038] The fusion weight coefficients are configured into the ridge regression model to perform a weighted fusion calculation on the five-dimensional feature vector to obtain the fusion intermediate prediction result;

[0039] The fused intermediate prediction results are subjected to data inverse normalization to restore them to the original dimensions of the load data, thus obtaining the power load prediction results.

[0040] In one embodiment, the step of inputting the five-dimensional feature vector into the ridge regression model and solving it using a robust piecewise fusion weight learning strategy to obtain the fusion weight coefficients includes:

[0041] The five-dimensional meta-feature vector is retrieved, and the verification set corresponding to the five-dimensional meta-feature vector is divided into sub-verification sets to obtain the first sub-verification set, the second sub-verification set, and the third sub-verification set.

[0042] The first sub-validation set, the second sub-validation set, and the third sub-validation set are input into the ridge regression model and solved based on the minimum error criterion to obtain the corresponding first fusion weight coefficient, second fusion weight coefficient, and third fusion weight coefficient.

[0043] The arithmetic mean of the first fusion weight coefficient, the second fusion weight coefficient, and the third fusion weight coefficient is calculated to obtain the average weight coefficient.

[0044] The average value of the weight coefficients is determined as the fusion weight coefficients that adapt to the five-dimensional feature vector.

[0045] In one embodiment, the step of configuring the fusion weight coefficients to perform weighted fusion calculations on the five-dimensional feature vectors using the ridge regression model to obtain intermediate fusion prediction results includes:

[0046] The fusion weight coefficients are analyzed to obtain weight allocation values ​​that match the dimensions of the five-dimensional feature vector;

[0047] The weight allocation values ​​are configured into the feature weight layer of the ridge regression model to update the weight parameters, resulting in the updated ridge regression model.

[0048] The five-dimensional feature vector is input into the updated ridge regression model to perform a weighted linear calculation of the feature dimensions, and the weighted calculation result is obtained.

[0049] The updated ridge regression model is subjected to regularization constraints to obtain the constrained calculation results;

[0050] The calculation results after the constraints are determined as the fusion intermediate prediction results.

[0051] Furthermore, to achieve the above objectives, this application also proposes a multi-model integrated forecasting device for power load, the multi-model integrated forecasting device for power load comprising:

[0052] The preprocessing module is used to clean and normalize the collected historical load time series data to obtain preprocessed time series data.

[0053] The feature construction module is used to construct multi-scale features based on the preprocessed time-series data using a strict causal sliding window mechanism, generating structured features, a first time-series matrix, and a second time-series matrix, respectively.

[0054] The optimization module is used to determine the optimal hyperparameter combination of the tree model, the gated recurrent unit model, and the bidirectional recurrent neural network model based on a two-stage hyperparameter joint optimization strategy, and to construct the optimal tree model, the optimal gated recurrent unit model, and the optimal bidirectional recurrent neural network model based on the optimal hyperparameter combination.

[0055] The prediction module is used to input the structured features, the first time series matrix, and the second time series matrix into the optimal tree model, the optimal gated recurrent unit model, and the optimal bidirectional recurrent neural network model for prediction, and to obtain the corresponding first prediction result, second prediction result, and third prediction result.

[0056] The results module is used to perform enhanced stacking and fusion of the first prediction result, the second prediction result, and the third prediction result to obtain the power load prediction result.

[0057] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the power load multi-model integrated prediction method described above.

[0058] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the power load multi-model integrated prediction method described above.

[0059] This application preprocesses historical load time-series data by cleaning and normalizing it, then constructs multi-scale features using a strict causal sliding window mechanism. A two-stage hyperparameter joint optimization strategy determines the optimal parameters for the tree model, the gated cyclic single-ridge regression meta-model, and the bidirectional recurrent neural network, and constructs corresponding models. Structured features, the first time-series matrix, and the second time-series matrix are input separately for orthogonal division of labor prediction. Finally, the results are output through enhanced stacking fusion. This approach avoids data leakage at its source, does not rely on exogenous variables, reduces the computational complexity of parameter tuning, achieves model differentiation and complementarity, and improves the accuracy and robustness of power load forecasting. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a flowchart illustrating the first embodiment of the multi-model integrated prediction method for power load of this application;

[0062] Figure 2 This is a flowchart illustrating the second embodiment of the multi-model integrated prediction method for power load of this application;

[0063] Figure 3 This is a schematic diagram of the module structure of the power load multi-model integrated prediction device of this application;

[0064] Figure 4 This is a schematic diagram of the equipment structure of the hardware operating environment involved in the power load multi-model integrated prediction method in the embodiments of this application.

[0065] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0066] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0067] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0068] To achieve power load forecasting, existing technologies have proposed various forecasting methodologies and integrated optimization schemes, mainly including: 1) Statistical model-based forecasting methods, such as ARIMA, exponential smoothing, and seasonal decomposition, which achieve forecasting by modeling the statistical characteristics of historical load data; 2) Machine learning-based forecasting methods, such as support vector regression, random forest, and gradient boosting trees, which capture the nonlinear relationship of load by manually constructing features input to the model; 3) Deep learning-based forecasting methods, such as long short-term memory networks, gated recurrent units, and temporal convolutional networks, which can learn time-dependent features end-to-end from the original load sequence; and 4) Decomposition-integration-based forecasting methods, which combine temporal decomposition algorithms with forecasting models, first decomposing the load sequence into trend, periodic, and residual components, then modeling and forecasting each component separately before reconstructing the result.

[0069] Existing power load forecasting methods all have significant limitations. First, the feature construction stage is prone to data leakage risks. Some methods use global time-series decomposition algorithms that are non-causal filters, requiring the utilization of the overall statistical characteristics of the load sequence. This leads to the indirect leakage of future information into the training features, resulting in inflated model evaluation results and a significant drop in generalization performance during actual deployment. Second, hyperparameter optimization strategies have shortcomings. If each sub-model is tuned independently, the global optimality of the ensemble cannot be guaranteed, easily destroying the differentiated and complementary value of the sub-models in the integration process. If brute-force full-grid joint parameter tuning is used, the number of parameters explodes exponentially, resulting in high computational costs and making it difficult to implement in engineering. Third, feature construction suffers from homogenization. The input feature information sources of each sub-model highly overlap, failing to form effective differentiated division of labor and complementarity, resulting in limited performance gains after integration. Therefore, how to construct multi-scale features based on pure historical load data and achieve high-precision power load forecasting while avoiding the data leakage risks caused by global decomposition has become an urgent problem to be solved.

[0070] Based on the above, this application also provides a multi-model integrated forecasting method for power load, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the multi-model integrated prediction method for power load of this application.

[0071] In this embodiment, the multi-model integrated prediction method for power load includes steps S10 to S50:

[0072] Step S10: Perform data cleaning and normalization on the collected historical load time series data to obtain preprocessed time series data.

[0073] It should be noted that historical load time series data refers to power load values ​​continuously recorded at fixed time intervals. This data fully reflects the entire pattern of power load changes over time and is the sole original basis for load forecasting. By deploying smart meters and synchronous phasor measurement devices at substations, distribution terminals, and user sides, active and reactive power in the power grid are periodically measured at a preset sampling frequency (e.g., every 15 minutes or hour) to obtain instantaneous values ​​of the original load, ensuring coverage of electricity consumption behaviors across different voltage levels and user types. Secondly, load data measured at each collection point is transmitted in real-time to a data center or edge computing node via a dedicated fiber optic network or wireless communication network. During transmission, clock synchronization calibration between the collection terminal and the main station system is performed based on the Network Time Protocol (NAT) to ensure consistency and alignment accuracy of multi-source data in the time dimension. Finally, the raw measurement data aggregated to the data platform undergoes preliminary screening to remove null values ​​and garbled characters caused by communication interruptions or equipment failures. The data is then structured and stored according to a unified data format (e.g., a two-column structure of timestamp-load data) to form historical load time series data available for subsequent analysis.

[0074] Further, step S10 includes: using statistical methods to identify outliers based on historical load time series data before the current prediction time, marking data with deviations exceeding a preset threshold as outliers and removing them to obtain the removed data; using linear interpolation to fill missing values ​​in the removed data with valid data before the current prediction time to obtain the filled data; resampling the removed data to convert non-uniform sampling data into sampling data with fixed time intervals to obtain resampled data before the current prediction time; and normalizing the resampled data before the current prediction time using the minimum-maximum normalization method to obtain preprocessed time series data.

[0075] Specifically, firstly, for historical load time series data prior to the current prediction time, the mean and standard deviation of the load data within that time period are calculated. For each data point, the standard deviation factor of its deviation from the mean is calculated (i.e., the value of that point minus the mean, divided by the standard deviation). When the absolute value of this factor exceeds a preset threshold (e.g., 3), the data point is marked as an outlier and directly removed. This is done to eliminate the interference of sudden data changes caused by equipment failures or unexpected disturbances on model training, while ensuring that outlier identification relies solely on historical statistical characteristics prior to the current prediction time without introducing future information. Secondly, for the gaps created after removing outliers, a linear relationship is established between the effective load data point before the gap time and the first effective data point after that time (ensuring both are earlier than the current prediction time) for interpolation calculation. The gap is filled by allocating the difference between the two points according to the time distance ratio, resulting in a continuous and complete data sequence. This is done to restore data continuity while maintaining temporal causality and avoid time series breaks caused by missing data. Then, the non-uniformly sampled data after filling is resampled, with a fixed time interval (e.g., 15 minutes). For each target sampling time, the arithmetic mean of all valid data points within a preset time window before and after that time is calculated as the representative load value for that time. This transforms the originally randomly spaced sampling points into a strictly equally spaced time series. This is done to unify the data time granularity and ensure the consistency of the dimension of subsequent model inputs. For example, data that was originally collected irregularly every 10 or 20 minutes is unified into one point every 15 minutes. Finally, for the resampled data before the current prediction time, the minimum and maximum load data within that time period are identified. For each data point, the difference between its value and the minimum value is calculated and divided by the difference between the maximum and minimum values. The result is mapped to a closed interval between 0 and 1 to obtain the preprocessed time series data. This is done to eliminate the influence of differences in load baselines in different regions or seasons, allowing the model to focus on the relative patterns of load changes rather than absolute values, accelerating model convergence and improving numerical stability.

[0076] Step S20: Based on the preprocessed time series data, a strict causal sliding window mechanism is used to construct multi-scale features, generating structured features, a first time series matrix, and a second time series matrix, respectively.

[0077] It should be noted that the strict causal sliding window mechanism refers to constructing features using only historical data prior to the current prediction time, without introducing any information from future times, ensuring that the feature generation process conforms to the rules of temporal causality. Structured features refer to high-dimensional combined features designed for tree models. These features are formed by fusing continuous near-point loads and historical loads at the same time, adapting to the input format and fitting logic of tree models. The first time-series matrix refers to the time-series input designed for gated recurrent single-ridge regression models. This matrix is ​​composed of nearest-neighbor loads and rolling means, focusing on expressing short-term fluctuation features. The second time-series matrix refers to the time-series input designed for bidirectional recurrent neural network models. This matrix is ​​composed of nearest-neighbor loads, rolling means, and historical features at the same time, focusing on expressing long-term periodic features.

[0078] Further, step S20 includes: determining the preset nearest neighbor window length, the preset number of historical days, and the preset sliding window length; based on the preprocessed time series data, extracting continuous nearby historical load data using a strict causal sliding window with the preset nearest neighbor window length, with the current prediction time as the endpoint, to obtain the nearest neighbor historical load features; based on the preprocessed time series data, extracting historical load data at the same time within the preset number of days using a strict causal sliding window, with the current prediction time as the benchmark, to obtain historical contemporaneous periodic features; calculating the rolling mean using a strict causal sliding window with the preset sliding window length based on the preprocessed time series data, to obtain the sliding mean features; fusing the nearest neighbor historical load features and the historical contemporaneous periodic features to obtain structured features adapted to the input form of the tree model; fusing the nearest neighbor historical load features and the sliding mean features to obtain a first time series matrix adapted to the input form of the gated recurrent single-ridge regression model; fusing the nearest neighbor historical load features, the sliding mean features, and the historical contemporaneous periodic features to obtain a second time series matrix adapted to the input form of the bidirectional recurrent neural network model.

[0079] Specifically, firstly, three types of window parameters are determined, and basic features are extracted based on strict causal constraints: the nearest neighbor window length is set to 95 points (corresponding to approximately 24 hours), the historical contemporaneous window length is 10 days, and the moving average window length is 5 points (corresponding to 75 minutes), ensuring that the endpoint of all windows is the current prediction time and does not contain future information; then, starting from the current prediction time, 95 points are traced back to extract continuous historical load data as the nearest neighbor historical load feature, expressed by the following formula:

[0080] ,

[0081] in Indicates the predicted time. Indicates the first The characteristics of the nearest historical load at any given time. Indicates the first Historical load data at any given time Indicates the first Historical load data at any given time Indicates the first Historical load data at any given time.

[0082] Using the current time point (e.g., 14:00) as a baseline, we trace back 10 days to extract preprocessed time-series data from the same time each day as historical synchronous periodic features. Let the number of daily sampling points be P=96, where P is the preset period length, i.e., the number of time points in a day. Then, the specific formula for the ten-day historical synchronous periodic features is expressed as follows:

[0083]

[0084] in Indicates the first Historical cyclical characteristics of a given moment Indicates the first Historical load data at any given time, i.e., the current forecast time. trace back one cycle Load records, Indicates the first Historical load data at any given time Indicates the first Historical load data at any given time.

[0085] The arithmetic mean of 5 points forward from the current prediction time is used as the moving average feature. The specific formula is as follows:

[0086]

[0087] in Indicates the first The time corresponding to the first The moving average characteristics of the nearest neighbor time points Indicates from arrive The historical load data within the window is accumulated, covering two time points before and after the target point. This represents historical load data shifted j time points before and after the target time point t−i. Secondly, based on the aforementioned fundamental features, differentiated fusion is performed to adapt to different models: neighboring historical load features and historical periodic features are vector-concatenated (directly linked end-to-end) to generate a 105-dimensional structured feature. The characteristics of nearby historical loads and moving averages are stacked along the time series dimension to generate a two-dimensional first time series matrix. The near-neighbor historical load features, moving average features, and historical contemporaneous periodic features are combined in a time series (near-neighbor sequences and moving averages are stacked, and historical contemporaneous features are broadcast, expanded, and aligned to the same time series dimension) to generate a second time series matrix. Finally, the structured features are input into the tree model for baseline load regression, the first time series matrix is ​​input into the gated recurrent single-ridge regression model for short-term fluctuation fitting, and the second time series matrix is ​​input into the bidirectional recurrent neural network model for long-term cycle memory. This approach aims to provide differentiated feature perspectives for models with different inductive biases while satisfying temporal causality, avoiding integration redundancy caused by feature homogenization. This allows each model to form positive interactive complements in the three dimensions of "baseline + fluctuation + cycle." For example, the tree model uses high-dimensional static features to capture overall load levels, the gated recurrent unit uses short sequences to capture local changes over the past 24 hours, and the bidirectional recurrent neural network uses sequences containing information from multiple days at the same time to capture daily cycle patterns.

[0088] Step S30: Determine the optimal hyperparameter combination of the tree model, the gated recurrent single-ridge regression meta-model, and the bidirectional recurrent neural network model based on the two-stage hyperparameter joint optimization strategy, and construct the optimal tree model, the optimal gated recurrent single-ridge regression meta-model, and the optimal bidirectional recurrent neural network model based on the optimal hyperparameter combination.

[0089] It should be noted that step S30 includes: performing a large-grid sequential coarse search on the tree model, the gated recurrent single-ridge regression meta-model, and the bidirectional recurrent neural network model, fixing the hyperparameters of the other two models to the default configuration, and searching for the optimal hyperparameter neighborhood for each individual model in the large grid to obtain the optimal hyperparameter neighborhood for each of the tree model, the gated recurrent single-ridge regression meta-model, and the bidirectional recurrent neural network model after the large-grid coarse search; constructing a small grid for each model based on the optimal hyperparameter neighborhood of each model to form a small grid space for joint hyperparameter search of the three models; performing a joint fine search on the three models in the small grid space, traversing all hyperparameter combinations and calculating the prediction error of each combination; and selecting the hyperparameter combination with the smallest prediction error as the optimal hyperparameter combination for each of the tree model, the gated recurrent single-ridge regression meta-model, and the bidirectional recurrent neural network model.

[0090] Specifically, a large-grid sequential coarse search is first performed on the tree model, the gated recurrent single-ridge regression meta-model, and the bidirectional recurrent neural network model. The hyperparameters of the other two models are fixed to the default configuration (e.g., learning rate 0.1, number of hidden units 64). Training and validation are performed only on a single target model in its large-grid candidate set (e.g., the number of leaves in the tree model {31, 63, 127, 255}, and the depth {6, 8, 10, 12}). The prediction error of the validation set under each parameter combination is recorded. The parameter point with the smallest error is selected as the center point and the neighborhood range is defined (e.g., the center point ± 1 grid step). The optimal hyperparameter neighborhood for each model is obtained. The total number of searches in this stage is the sum of the number of large-grid candidates for each model. That is, if each of the three models has 4 sets of candidates, the total number of searches is 4 + 4 + 4 = 12 times, which is much less than the 4 × 4 × 4 = 64 times of brute force joint search. This is done in order to quickly locate the potential optimal region of each model under the premise of decoupling the complex interaction between models, and to avoid ineffective exploration in the high-dimensional parameter space in the early stage. Secondly, based on the optimal hyperparameter neighborhoods obtained after coarse search in the large grid for each model (such as the leaf number neighborhood {63,95,127} for the tree model, the hidden layer unit number neighborhood {48,64,80} for the gated recurrent unit, and the hidden layer unit number neighborhood {56,64,72} for the bidirectional recurrent neural network), a small grid candidate set is constructed for each model, forming a small grid space for joint hyperparameter search of the three models (the space size is the product of the number of small grid candidates for each model, i.e., 3×3×3=27 combinations). In this space, a joint fine search is performed on the three models: traversing all hyperparameter combinations, configuring the parameters of the three models for each combination and performing joint training, calculating the multi-step prediction error (such as root mean square error) under this combination, and selecting the hyperparameter combination with the smallest prediction error as the final optimal hyperparameter combination for each model. The total computational cost of this two-stage strategy is 12+27=39 training iterations, which is about 39% less computation than the 64 iterations of brute-force joint search. At the same time, the global optimization capability of the coupling relationship between models is preserved through the joint traversal in the neighborhood in the second stage.

[0091] Then, the optimal hyperparameter combinations of each model are configured into the corresponding base models for model training, resulting in the optimized tree model, the optimized gated recurrent single-ridge regression model, and the optimized bidirectional recurrent neural network model. Performance tests are then performed on the optimized tree model, the optimized gated recurrent single-ridge regression model, and the optimized bidirectional recurrent neural network model to obtain the test results. When the test results show that the optimized tree model, the optimized gated recurrent single-ridge regression model, and the optimized bidirectional recurrent neural network model all meet the preset requirements, the current optimized tree model, the optimized gated recurrent single-ridge regression model, and the optimized bidirectional recurrent neural network model are determined as the optimal tree model, the optimal gated recurrent single-ridge regression model, and the optimal bidirectional recurrent neural network model, respectively.

[0092] Specifically, the optimal hyperparameter combinations corresponding to each model (e.g., 95 leaves and 8 depth for the tree model, 64 hidden units for the gated recurrent unit, and a dropout rate of 0.2) are configured into the corresponding basic model architecture. The models are trained using the training set data until convergence, resulting in the optimized tree model, the optimized gated recurrent single-ridge regression meta-model, and the optimized bidirectional recurrent neural network model. Subsequently, these optimized models are tested on an independent test set, and various evaluation indicators are calculated. When the test results show that the prediction accuracy of each model meets the preset threshold requirements (e.g., the mean absolute percentage error (MAPE) is less than 1%), the current optimized model is officially determined as the optimal tree model, the optimal gated recurrent single-ridge regression meta-model, and the optimal bidirectional recurrent neural network model for subsequent ensemble prediction. This is done to ensure that each sub-model maintains a collaborative performance baseline after independent optimization, and to avoid ensemble failure caused by overfitting or underfitting of a single model.

[0093] Step S40: Input the structured features, the first time series matrix, and the second time series matrix into the optimal tree model, the optimal gated recurrent single-ridge regression model, and the optimal bidirectional recurrent neural network model for prediction, and obtain the corresponding first prediction result, second prediction result, and third prediction result.

[0094] It should be noted that step S40 includes: inputting structured features into the optimal tree model and performing baseline load regression prediction to obtain a first initial prediction result; inputting the first time series matrix into the optimal gated recurrent single-ridge regression model and performing short-term load fluctuation fitting prediction to obtain a second initial prediction result; inputting the second time series matrix into the optimal bidirectional recurrent neural network model and performing long-term load cycle memory prediction to obtain a third initial prediction result; and performing format standardization processing on the first, second, and third initial prediction results to obtain the first, second, and third prediction results.

[0095] Specifically, structured features (high-dimensional vectors containing nearest-neighbor historical load features and historical contemporaneous periodic features) are input into the optimal tree model (such as LightGBM after two-stage optimization). This model constructs an additive prediction function through multiple iterations, that is, it accumulates the index values ​​of the leaf nodes of each decision tree based on the current sample features to obtain the output value. Specifically, it traverses each decision tree sequentially, determines the leaf node where the sample falls based on the feature splitting condition (such as whether the historical load data of the 48th nearest neighbor is greater than 0.65) and obtains the weight of that node, sums the weight values ​​output by all trees, and maps them through the link function to obtain the first initial prediction function. The initial prediction is done to leverage the tree model's efficient fitting capability for high-dimensional tabular features. It captures daily cycle benchmarks based on historical points and recent trends based on nearest neighbor points, thus stabilizing the output of baseline estimates for the main load level. Secondly, the first time-series matrix (two-dimensional time-series data containing nearby historical load features and moving average features) is input into the optimally gated cyclic single-ridge regression meta-model step by step. For each time step, the fusion ratio between historical hidden states and the current input is calculated based on the update gate, and the degree of forgetting of historical information is determined based on the reset gate. Temporary memories for the current moment are generated through candidate states. Finally, the hidden state is updated and passed to the next time step. After traversing all time steps, the hidden state at the last time step is taken and mapped through a fully connected layer to obtain the second initial prediction result. This is done to utilize the selective memory capability of the gating mechanism for time-series dependencies, smooth noise interference through the rolling mean feature, and capture the local fluctuation pattern of the past 24 hours through the nearest neighbor sequence, thereby accurately fitting short-term dynamic changes. Then, the second time-series matrix (containing three-dimensional time-series data including the features of the nearest neighbor historical load, the moving average feature, and the historical periodic features) is input into the forward and backward layers of the optimal bidirectional recurrent neural network model, respectively. The forward layer calculates the hidden state layer by layer in chronological order from past to present, while the backward layer calculates the hidden state layer by layer in chronological order from present to past. At each time step, the forward hidden state and the backward hidden state are concatenated and fused to form a comprehensive representation. The fused hidden state at the last moment is taken and mapped through the output layer to obtain the third initial prediction result. This is done in order to simultaneously utilize the sensitivity of the forward layer to recent trends and the sensitivity of the backward layer to historical cycle patterns. By introducing the broadcast of historical features at the same moment, the cycle memory is extended to a very long history (such as the same moment 10 days ago), thereby effectively capturing long-term regular features such as daily cycles and weekly cycles.Finally, the first, second, and third initial prediction results are standardized, including unifying numerical precision (e.g., retaining four decimal places), unifying timestamp alignment (ensuring the three results strictly correspond to the same prediction time), and unifying dimensional restoration (if normalization was performed during training, the predicted values ​​are mapped back to the original load magnitude through inverse transformation). This is done to ensure that the outputs of the three heterogeneous models (tree model, gated recurrent unit, and bidirectional recurrent neural network) are comparable and fusionable, providing standardized input for subsequent enhanced stacked fusion, and avoiding deviations in fusion weight calculation or information loss due to differences in numerical precision, timestamp misalignment, or inconsistent dimensions.

[0096] Step S50: Enhanced stacking and fusion of the first prediction result, the second prediction result, and the third prediction result are performed to obtain the power load prediction result.

[0097] It should be noted that the first prediction result refers to the load forecast value output by the optimal tree model. This result mainly reflects the baseline level change of power load and is a fundamental component of integrated forecasting. The second prediction result refers to the load forecast value output by the optimal gated cyclic single-ridge regression model. This result mainly captures the short-term fluctuation characteristics of power load and improves the forecast's adaptability to sudden changes. The third prediction result refers to the load forecast value output by the optimal bidirectional recurrent neural network model. This result mainly characterizes the long-term cyclical pattern of power load and enhances the forecast's ability to grasp trend changes.

[0098] Specifically, firstly, the first, second, and third prediction results are concatenated to construct a basic prediction value vector. The arithmetic mean of all elements in this vector is calculated as the prediction mean feature, and the standard deviation of all elements in this vector is calculated as the prediction standard deviation feature. Secondly, the basic prediction value vector, the prediction mean feature, and the prediction standard deviation feature are concatenated to construct a five-dimensional feature vector. The validation set is divided into a predetermined number (e.g., 3) mutually exclusive data subsets according to time order. The five-dimensional feature vector is input into the ridge regression model with regularization terms corresponding to each data subset for local training. The local fusion weights corresponding to each subset are obtained by minimizing the mean square error and adding a weight decay penalty term. Finally, the arithmetic mean of the local fusion weights corresponding to each data subset is calculated to obtain the global fusion weight coefficient. Based on the global fusion weight coefficient, the five-dimensional feature vector is weighted and fused to obtain the power load prediction result. This is done to enable the Ridge Regressor Model to explicitly perceive differences in model confidence at the sample level (automatically reducing the prediction weight of a sample or adjusting the output bias when the predictions of the three models differ significantly), prevent the Ridge Regressor Model from overfitting to the validation set through regularization constraints, and enhance robustness to data distribution drift by employing a segmented learning strategy. This effectively avoids poorly performing individual sub-models dragging down the overall integration effect during the fusion stage, achieving true complementary advantages and dynamic weight allocation.

[0099] This embodiment preprocesses historical load time-series data by cleaning and normalizing it. Then, it constructs multi-scale features using a strict causal sliding window mechanism. A two-stage hyperparameter joint optimization strategy determines the optimal parameters for the tree model, the gated cyclic single-ridge regression meta-model, and the bidirectional recurrent neural network, and constructs the corresponding models. The structured features, the first time-series matrix, and the second time-series matrix are input separately for orthogonal division of labor prediction. Finally, the results are output through enhanced stacking fusion. This approach avoids data leakage at its source, does not rely on exogenous variables, reduces the computational complexity of parameter tuning, achieves model differentiation and complementarity, and improves the accuracy and robustness of power load forecasting.

[0100] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 The multi-model integrated prediction method for power load further includes steps S201 to S206 in step S50:

[0101] Step S201: Calculate the arithmetic mean of the three prediction results based on the first prediction result, the second prediction result, and the third prediction result to obtain the prediction mean feature.

[0102] Specifically, firstly, the specific values ​​of the first, second, and third prediction results are obtained. The sum of these three values ​​is then divided by the number of values ​​(3), and the arithmetic mean is calculated as the prediction mean feature. For example, when the first prediction result is 100 MW, the second is 105 MW, and the third is 95 MW, (100+105+95) / 3 = 100 MW is calculated, and this 100 MW is used as the prediction mean feature. Secondly, this prediction mean feature is used as a component of the subsequent meta-feature vector, concatenated with the basic prediction value vector and the prediction standard deviation feature to construct a complete five-dimensional meta-feature representation. Finally, this prediction mean feature is used as a benchmark reference for the prediction central trend, used to measure the degree of deviation of each sub-model's prediction value from the overall center, and to provide a centralization adjustment benchmark for subsequent fusion weight calculation. This is done to extract the central tendency of the prediction values ​​of the three heterogeneous models through a simple arithmetic mean, serving as a benchmark reference for the prediction results.

[0103] Step S202: Calculate the standard deviation of the three prediction results based on the first prediction result, the second prediction result, and the third prediction result to obtain the prediction standard deviation feature that characterizes the model's degree of divergence.

[0104] Specifically, firstly, the specific values ​​of the first, second, and third prediction results are obtained, and their arithmetic mean is calculated as the baseline mean. Then, the difference between each prediction result and this baseline mean is calculated and squared (e.g., the square of the difference between the first, second, and third prediction results and the mean). The sum of the three squared differences is divided by the number of values ​​to obtain the variance. The square root of the variance is then taken to obtain the prediction standard deviation feature. Secondly, this calculated prediction standard deviation feature is used as a component of the five-dimensional feature vector and concatenated with the base prediction value vector and the prediction mean feature. Finally, this prediction standard deviation feature is used as a quantitative indicator of model divergence and input into the ridge regression model. When the standard deviation exceeds a preset divergence threshold (e.g., 10 MW), it indicates severe divergence among the three models and low confidence. The ridge regression model is a linear regression model with regularization constraints, which can effectively suppress overfitting and ensure the stability and generalization ability of the fused weights. The ridge regression model adaptively adjusts the fusion weights accordingly to reduce the prediction confidence of the sample or weaken the influence of sub-models with larger biases. This is done to explicitly quantify the degree of divergence of the predicted values ​​of the three heterogeneous models on a specific sample using the standard deviation statistic, transforming the consistency information between models into computable features. This enables the fusion layer to perceive whether the models have reached a consensus, thereby adopting a conservative prediction strategy when there is significant divergence, effectively avoiding the dragging down of the overall ensemble performance by a single outlier model, and improving the robustness and reliability of the prediction results.

[0105] Step S203: The first prediction result, the second prediction result, the third prediction result, the predicted mean feature, and the predicted standard deviation feature are concatenated according to a preset dimension to obtain a five-dimensional feature vector.

[0106] Specifically, firstly, the first, second, and third prediction results are used as basic prediction components. The calculated predicted mean and predicted standard deviation features are used as statistical enhancement components. These are then horizontally concatenated according to a preset dimensional order (e.g., first prediction result, second prediction result, third prediction result, predicted mean feature, predicted standard deviation feature) to construct a five-dimensional row vector. For example, when the first prediction result is 100 MW, the second prediction result is 105 MW, and the third prediction result is 95 MW, the calculated predicted mean feature is 100 MW, and the predicted standard deviation feature is approximately 4 MW. The resulting five-dimensional meta-feature vector is [100, 105, 95, 100, 4]. Secondly, the five-dimensional meta-feature vector undergoes numerical precision unification (e.g., uniformly retaining four decimal places) and dimensionality checks to ensure that the values ​​of each dimension are within a reasonable range and in a consistent format. Finally, the constructed five-dimensional meta-feature vector is used as the input representation of a single sample to batch construct the meta-feature matrices of the validation and test sets, which serve as the input data for the subsequent ridge regression meta-model. This is done to retain the original prediction information (first three dimensions) of the three base models (tree model, gated recurrent unit model, and bidirectional recurrent neural network model) while explicitly injecting two types of statistical information: prediction central tendency (fourth dimension) and model divergence (fifth dimension). This allows the meta-model to learn the dynamic weight allocation strategy of each sub-model based on the five-dimensional features. It uses the original prediction values ​​for regression fitting and the mean and standard deviation features for confidence correction, thereby improving the accuracy and robustness of the fused prediction.

[0107] Step S204: Input the five-dimensional feature vector into the ridge regression model and solve it using a robust piecewise fusion weight learning strategy to obtain the fusion weight coefficients.

[0108] It should be noted that step S204 includes: first, retrieving the five-dimensional meta-feature vector, and dividing the validation set corresponding to the five-dimensional meta-feature vector into sub-validation sets to obtain the first sub-validation set, the second sub-validation set, and the third sub-validation set. Specifically, retrieving the constructed five-dimensional meta-feature vector and its corresponding real payload data, and dividing the validation set into a preset number (3) mutually exclusive and equal-length sub-validation sets according to time order to obtain the first sub-validation set. Second Subset of Verification and the third sub-validation set This ensures the temporal continuity of samples within each subset to preserve temporal characteristics.

[0109] Next, the first, second, and third sub-validation sets are input into the ridge regression model and solved based on the minimum error criterion to obtain the corresponding first, second, and third fusion weight coefficients. Specifically, firstly, the validation set is divided into three mutually exclusive sub-validation sets (first, second, and third sub-validation sets) according to time sequence, ensuring that each sub-set contains a complete five-dimensional feature vector (composed of the first prediction result, second prediction result, third prediction result, prediction mean feature, and prediction standard deviation feature) and its corresponding actual load data, and that each sub-set is continuously distributed on the time axis to preserve time-series characteristics; secondly, the first, second, and third sub-validation sets are input into the ridge regression model respectively, and a constrained minimization problem is solved independently for each sub-validation set, i.e., under the constraint that the sum of the weights is 1 ( Minimize the sum of squared residuals between the actual values ​​and the weighted predicted values, mathematically expressed as:

[0110]

[0111] in This represents the actual historical load data for this subset. Let represent the predicted load data of the i-th base model (tree model, gated recurrent unit model, and bidirectional recurrent neural network model) on this subset. Based on the above formula, the constrained optimization problem is solved analytically or numerically to obtain the first fusion weight coefficient (corresponding to the first sub-validation set), the second fusion weight coefficient (corresponding to the second sub-validation set), and the third fusion weight coefficient (corresponding to the third sub-validation set) that minimize the prediction error. This is done to maintain the causality and temporal continuity of the data through time-sequential segmentation, avoiding future information leakage due to random shuffling. Simultaneously, the constraint that the weight sum is 1 on each subset ensures that the fusion result is within the convex hull of the base model's prediction, preventing over-extrapolation. Furthermore, the stability of the fusion strategy on different data distribution segments is tested through independent multi-segment solutions.

[0112] Then, the arithmetic mean of the first, second, and third fusion weight coefficients is calculated to obtain the average weight coefficient; finally, the average weight coefficient is determined as the fusion weight coefficient adapted to the five-dimensional feature vector. Specifically, the arithmetic mean of the first, second, and third fusion weight coefficients is performed, and the three weight vectors are added according to their corresponding dimensions and divided by the number of subsets (3) to calculate the average weight coefficient; secondly, the average weight coefficient is determined as the final fusion weight coefficient adapted to the five-dimensional feature vector, and a fixed mapping relationship from the five-dimensional feature to the prediction result is established; finally, in the prediction stage, the five-dimensional feature vector is weighted and summed based on the fusion weight coefficient to output the final power load prediction result. This is done to smooth out potential overfitting or local optima issues in a single subset of validation sets by using the arithmetic mean of multiple weights, thereby reducing the model's sensitivity to specific data distribution segments. This ensures that the fused weights remain stable under different power grid operating phases (such as peak-valley and flat load periods, and different seasonal operating conditions), enhancing robustness to data distribution drift, and improving the reproducibility and reliability of engineering deployment.

[0113] Step S205: Configure the fusion weight coefficients to perform weighted fusion calculation on the five-dimensional feature vector of the ridge regression model to obtain the fusion intermediate prediction result.

[0114] It should be noted that step S205 includes: parsing the fusion weight coefficients to obtain weight allocation values ​​that match the dimensions of the five-dimensional metafeature vector; configuring the weight allocation values ​​to the feature weight layer of the ridge regression model to update the weight parameters, thereby obtaining the updated ridge regression model; inputting the five-dimensional metafeature vector into the updated ridge regression model to perform weighted linear calculation of the feature dimensions, thereby obtaining the weighted calculation result; performing regularization constraint processing on the updated ridge regression model to obtain the constrained calculation result; and determining the constrained calculation result as the fusion intermediate prediction result.

[0115] Specifically, firstly, the fusion weight coefficients are analyzed for dimensionality, mapping them to five weight allocation values ​​that are matched one-to-one with the dimensions of the five-dimensional feature vector. These weights correspond to the weighted coefficients of the first prediction result, the second prediction result, the third prediction result, the predicted mean feature, and the predicted standard deviation feature, respectively. For example, the fusion weight coefficients can be analyzed from the mean value of the weight coefficients. Five numerical values ​​are assigned. First, these five weights are configured into the feature weight layer of the ridge regressor model, serving as model parameters for fixation and updating. This replaces the randomly initialized or temporary weights from the previous stage, resulting in an updated ridge regressor model with deterministic weighting capabilities. Then, the five-dimensional feature vector is input into this updated model, and a weighted linear calculation of the feature dimensions is performed. This involves multiplying each feature dimension value by its corresponding weight and summing the results to obtain the weighted calculation result. Subsequently, based on the L2 regularization constraint characteristic embedded in the ridge regressor model during training (which penalizes large weights to prevent overfitting), the weighted calculation result is constrained to suppress numerical instability caused by multicollinearity among features (such as the high correlation between the three basic prediction results and the predicted mean feature), resulting in a constrained calculation result. Finally, this constrained calculation result is determined as the fused intermediate prediction result, serving as the final power load prediction value. This is done to achieve refined linear fusion of five-dimensional features through dimensional matching weight allocation, and to automatically handle the variance inflation problem caused by the correlation between features by using the regularization mechanism of ridge regression. This ensures that the fusion result will not fluctuate drastically due to the extreme prediction values ​​of a single base model or the linear dependence between features. At the same time, by fixing the weight parameters, the model has deterministic prediction behavior, which is convenient for stable deployment and result reproduction in the power dispatching system.

[0116] Step S206: Perform data inverse normalization on the fused intermediate prediction results to restore them to the original dimensions of the load data, and obtain the power load prediction results.

[0117] It should be noted that, firstly, the minimum and maximum values ​​of the original load data recorded in the preprocessing stage are retrieved (e.g., the minimum value of the historical load in the training set is 500 MW and the maximum value is 1500 MW), and the intermediate prediction result is obtained (e.g., the dimensionless value after normalization, 0.6). Based on the inverse normalization, the dimension is restored, that is, 0.6 multiplied by the range (1000 MW) and added to the minimum value (500 MW) to get 1100 MW. Secondly, the physical rationality of the inverse normalized value is checked to see if it is within the preset normal load range (e.g., 300 MW to 2000 MW). If it exceeds the reasonable range, an abnormal alarm is triggered or truncation is performed. Finally, the value that passes the verification is formatted into standard engineering units (e.g., retaining two decimal places and adding the unit "MW") and determined as the power load prediction result. This is done to map the predicted values ​​learned by the model in a dimensionless space (range of 0 to 1) back to the real physical dimensions, so that the prediction results have interpretable engineering significance and can be directly used by grid dispatchers for actual production decisions such as power generation planning and reserve capacity configuration, avoiding dispatching errors caused by numerical scale confusion.

[0118] This embodiment calculates the mean and standard deviation of the three types of prediction results, concatenates them with the original prediction values ​​to form a five-dimensional feature vector, and then obtains the fusion weight coefficients through a ridge regression model combined with robust piecewise weight learning. The final load prediction result is obtained after weighted fusion and inverse normalization. By introducing confidence features to improve the rationality of fusion, piecewise learning to enhance the robustness of weights, and restoring the dimensions to ensure the practicality of the results, the prediction accuracy and stability are effectively improved.

[0119] Based on the first embodiment of this application, this application also provides a multi-model integrated prediction device for power load, please refer to... Figure 3 The device includes:

[0120] The preprocessing module 10 is used to clean and normalize the collected historical load time series data to obtain preprocessed time series data.

[0121] The feature construction module 20 is used to construct multi-scale features based on preprocessed time series data using a strict causal sliding window mechanism, generating structured features, a first time series matrix, and a second time series matrix, respectively.

[0122] The optimization module 30 is used to determine the optimal hyperparameter combination of the tree model, the gated recurrent unit model, and the bidirectional recurrent neural network model based on a two-stage hyperparameter joint optimization strategy, and to construct the optimal tree model, the optimal gated recurrent unit model, and the optimal bidirectional recurrent neural network model based on the optimal hyperparameter combination.

[0123] The prediction module 40 is used to input the structured features, the first time series matrix, and the second time series matrix into the optimal tree model, the optimal gated recurrent unit model, and the optimal bidirectional recurrent neural network model for prediction, and to obtain the corresponding first prediction result, second prediction result, and third prediction result.

[0124] The results module 50 is used to perform enhanced stacking and fusion of the first prediction result, the second prediction result and the third prediction result to obtain the power load prediction result.

[0125] The power load multi-model integrated forecasting device provided in this application, employing the power load multi-model integrated forecasting method in the above embodiments, can solve the technical problem of how to avoid data leakage and improve load forecasting accuracy and robustness without relying on exogenous variables. Compared with the prior art, the beneficial effects of the power load multi-model integrated forecasting device provided in this application are the same as those of the power load multi-model integrated forecasting method provided in the above embodiments, and other technical features in the power load multi-model integrated forecasting device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0126] This application provides a power load multi-model integrated prediction device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the power load multi-model integrated prediction method in the above embodiment 1.

[0127] The following is for reference. Figure 4 This document illustrates a structural schematic diagram of a power load multi-model integrated forecasting device suitable for implementing embodiments of this application. The power load multi-model integrated forecasting device in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 4 The power load multi-model integrated prediction device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0128] like Figure 4As shown, the power load multi-model integrated forecasting device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 1002 or a program loaded from storage device 1003 into random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the power load multi-model integrated forecasting device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the power load multi-model integrated forecasting device to communicate wirelessly or wiredly with other devices to exchange data. Although various power load multi-model integrated forecasting devices are shown in the figures, it should be understood that implementation or possession of all of them is not required. More or fewer may be implemented alternatively.

[0129] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0130] The power load multi-model integrated forecasting device provided in this application, employing the power load multi-model integrated forecasting method in the above embodiments, can solve the technical problem of how to avoid data leakage and improve load forecasting accuracy and robustness without relying on exogenous variables. Compared with the prior art, the beneficial effects of the power load multi-model integrated forecasting device provided in this application are the same as those of the power load multi-model integrated forecasting method provided in the above embodiments, and other technical features in this power load multi-model integrated forecasting device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0131] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0132] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0133] This application provides a computer-readable medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the power load multi-model integrated prediction method in the above embodiments.

[0134] The computer-readable medium provided in this application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor devices, or any combination thereof. More specific examples of computer-readable media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable medium may be any tangible medium containing or storing a program that can be executed by instructions, used by a device, or used in conjunction with it. The program code contained on the computer-readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0135] The aforementioned computer-readable medium may be included in the power load multi-model integrated forecasting device; or it may exist independently and not be assembled into the power load multi-model integrated forecasting device.

[0136] The aforementioned computer-readable medium carries one or more programs that, when executed by the power load multi-model integrated forecasting device, enable the power load multi-model integrated forecasting device to write computer program code for performing the operations of this application in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—and conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0137] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of this application. In this regard, all blocks in the flowcharts or block diagrams may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that all blocks in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using dedicated hardware-based implementations that perform the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0138] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0139] The readable medium provided in this application is a computer-readable medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described multi-model integrated forecasting method for power load. This solves the technical problem of how to avoid data leakage and improve load forecasting accuracy and robustness without relying on exogenous variables. Compared with the prior art, the beneficial effects of the computer-readable medium provided in this application are the same as those of the multi-model integrated forecasting method for power load provided in the above embodiments, and will not be elaborated upon here.

[0140] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the multi-model integrated prediction method for power load as described above.

[0141] The computer program product provided in this application can solve the technical problem of how to avoid data leakage and improve the accuracy and robustness of load forecasting without relying on exogenous variables. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the power load multi-model integrated forecasting method provided in the above embodiments, and will not be repeated here.

[0142] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A multi-model integrated forecasting method for power load, characterized in that, The method includes: The collected historical load time series data are cleaned and normalized to obtain preprocessed time series data; Based on the preprocessed time-series data, a strict causal sliding window mechanism is used to construct multi-scale features, generating structured features, a first time-series matrix, and a second time-series matrix, respectively. The optimal hyperparameter combination of the tree model, the gated recurrent single-ridge regression model, and the bidirectional recurrent neural network model is determined based on the two-stage hyperparameter joint optimization strategy, and the optimal tree model, the optimal gated recurrent single-ridge regression model, and the optimal bidirectional recurrent neural network model are constructed based on the optimal hyperparameter combination. The structured features, the first time series matrix, and the second time series matrix are respectively input into the optimal tree model, the optimal gated recurrent single-ridge regression model, and the optimal bidirectional recurrent neural network model for prediction, to obtain the corresponding first prediction result, second prediction result, and third prediction result; The first, second, and third prediction results are enhanced and stacked together to obtain the power load prediction result, including: The arithmetic mean of the three prediction results is calculated based on the first prediction result, the second prediction result, and the third prediction result to obtain the prediction mean feature. Based on the first prediction result, the second prediction result and the third prediction result, the standard deviation of the three prediction results is calculated to obtain the prediction standard deviation feature that characterizes the model's degree of divergence. The first prediction result, the second prediction result, the third prediction result, the predicted mean feature, and the predicted standard deviation feature are concatenated according to a preset dimension to obtain a five-dimensional feature vector; The five-dimensional feature vector is input into the ridge regression model, and a robust piecewise fusion weight learning strategy is used to solve the problem, resulting in fusion weight coefficients, specifically including: The five-dimensional meta-feature vector is retrieved, and the verification set corresponding to the five-dimensional meta-feature vector is divided into sub-verification sets to obtain the first sub-verification set, the second sub-verification set, and the third sub-verification set. The first sub-validation set, the second sub-validation set, and the third sub-validation set are input into the ridge regression model and solved based on the minimum error criterion to obtain the corresponding first fusion weight coefficient, second fusion weight coefficient, and third fusion weight coefficient. The arithmetic mean of the first fusion weight coefficient, the second fusion weight coefficient, and the third fusion weight coefficient is calculated to obtain the average weight coefficient. The average of the weight coefficients is determined as the fusion weight coefficients that adapt to the five-dimensional feature vector; The fusion weight coefficients are configured into the ridge regression model to perform a weighted fusion calculation on the five-dimensional feature vector to obtain the fusion intermediate prediction result; The fused intermediate prediction results are subjected to data inverse normalization to restore them to the original dimensions of the load data, thus obtaining the power load prediction results.

2. The method as described in claim 1, characterized in that, The step of constructing multi-scale features based on the preprocessed time-series data using a strict causal sliding window mechanism, and generating structured features, a first time-series matrix, and a second time-series matrix respectively, includes: Determine the preset neighbor window length, preset historical days, and preset sliding window length; Based on the preprocessed time series data, the historical load data of continuous nearby points are extracted using a strict causal sliding window with a preset nearest neighbor window length, with the current prediction time as the endpoint, to obtain the nearest neighbor historical load characteristics. Based on the preprocessed time series data, using the current prediction time as a benchmark, a strict causal sliding window is used to extract historical load data at the same time within a preset number of days to obtain historical synchronous periodic characteristics. Based on the preprocessed time series data, a strict causal sliding window with a preset sliding window length is used to calculate the rolling mean, and the sliding mean feature is obtained. By fusing the nearest-neighbor historical load features and the historical contemporaneous period features, a structured feature in the input form of the adaptive tree model is obtained; By integrating the nearest-neighbor historical load characteristics and the moving average characteristics, a first time series matrix adapted to the input form of the gated cyclic single-ridge regression metamodel is obtained; By integrating the nearest neighbor historical load features, the moving average features, and the historical periodic features, a second time-series matrix adapted to the input form of a bidirectional recurrent neural network model is obtained.

3. The method as described in claim 1, characterized in that, The steps of determining the optimal hyperparameter combination of the tree model, the gated recurrent single-ridge regression meta-model, and the bidirectional recurrent neural network model based on a two-stage hyperparameter joint optimization strategy, and constructing the optimal tree model, the optimal gated recurrent single-ridge regression meta-model, and the optimal bidirectional recurrent neural network model based on the optimal hyperparameter combination, include: A large-grid sequential coarse search is performed on the tree model, the gated recurrent single-ridge regression meta-model, and the bidirectional recurrent neural network model. The hyperparameters of the other two models are fixed to the default configuration. The optimal hyperparameter neighborhood of each individual model is searched in the large grid to obtain the optimal hyperparameter neighborhood of each of the tree model, the gated recurrent single-ridge regression meta-model, and the bidirectional recurrent neural network model after the large-grid coarse search. Based on the optimal hyperparameter neighborhood of each model, a small grid is constructed for each model to form a small grid space for joint search of hyperparameters of the three models; A joint fine search is performed on the three models within the small grid space, traversing all hyperparameter combinations and calculating the prediction error of each combination; The hyperparameter combination with the smallest prediction error is selected as the optimal hyperparameter combination for the tree model, the gated recurrent single-ridge regression model, and the bidirectional recurrent neural network model, respectively. The optimal hyperparameter combinations of each model are configured into the corresponding base model for model training to obtain the optimized tree model, the optimized gated recurrent single-ridge regression meta-model, and the optimized bidirectional recurrent neural network model. The performance of the optimized tree model, the optimized gated recurrent single-ridge regression meta-model, and the optimized bidirectional recurrent neural network model were tested, and the test results were obtained. When the test results show that the optimized tree model, optimized gated cyclic single-ridge regression model, and optimized bidirectional recurrent neural network model all meet the preset requirements, the current optimized tree model, optimized gated cyclic single-ridge regression model, and optimized bidirectional recurrent neural network model are determined as the optimal tree model, optimal gated cyclic single-ridge regression model, and optimal bidirectional recurrent neural network model.

4. The method as described in claim 1, characterized in that, The step of inputting the structured features, the first time series matrix, and the second time series matrix into the optimal tree model, the optimal gated recurrent single-ridge regression meta-model, and the optimal bidirectional recurrent neural network model respectively for prediction, and obtaining the corresponding first prediction result, second prediction result, and third prediction result, includes: The structured features are input into the optimal tree model, and the baseline load regression prediction is performed to obtain the first initial prediction result; The first time series matrix is ​​input into the optimal gated cyclic single-ridge regression model to perform short-term load fluctuation fitting prediction and obtain the second initial prediction result. The second time series matrix is ​​input into the optimal bidirectional recurrent neural network model to perform long-term load cycle memory prediction and obtain the third initial prediction result. The first initial prediction result, the second initial prediction result, and the third initial prediction result are format-standardized to obtain the first prediction result, the second prediction result, and the third prediction result.

5. The method as described in claim 1, characterized in that, The step of configuring the fusion weight coefficients to perform weighted fusion calculations on the five-dimensional feature vectors of the ridge regression model to obtain the fusion intermediate prediction results includes: The fusion weight coefficients are analyzed to obtain weight allocation values ​​that match the dimensions of the five-dimensional feature vector; The weight allocation values ​​are configured into the feature weight layer of the ridge regression model to update the weight parameters, resulting in the updated ridge regression model. The five-dimensional feature vector is input into the updated ridge regression model to perform a weighted linear calculation of the feature dimensions, and the weighted calculation result is obtained. The updated ridge regression model is subjected to regularization constraints to obtain the constrained calculation results; The calculation results after the constraints are determined as the fusion intermediate prediction results.

6. A multi-model integrated prediction device for power load, characterized in that, The apparatus is applied to the power load multi-model integrated prediction method as described in any one of claims 1-5, and the apparatus comprises: The preprocessing module is used to clean and normalize the collected historical load time series data to obtain preprocessed time series data. The feature construction module is used to construct multi-scale features based on the preprocessed time-series data using a strict causal sliding window mechanism, generating structured features, a first time-series matrix, and a second time-series matrix, respectively. The optimization module is used to determine the optimal hyperparameter combination of the tree model, the gated recurrent single-ridge regression model, and the bidirectional recurrent neural network model based on a two-stage hyperparameter joint optimization strategy, and to construct the optimal tree model, the optimal gated recurrent single-ridge regression model, and the optimal bidirectional recurrent neural network model based on the optimal hyperparameter combination. The prediction module is used to input the structured features, the first time series matrix, and the second time series matrix into the optimal tree model, the optimal gated recurrent single-ridge regression model, and the optimal bidirectional recurrent neural network model for prediction, and obtain the corresponding first prediction result, second prediction result, and third prediction result. The results module is used to perform enhanced stacking and fusion of the first prediction result, the second prediction result, and the third prediction result to obtain the power load prediction result.

7. A multi-model integrated forecasting device for power load, characterized in that, The device includes: a memory, a processor, and a power load multi-model integrated prediction program stored in the memory and running on the processor, the power load multi-model integrated prediction program being configured to implement the steps of the power load multi-model integrated prediction method as described in any one of claims 1-5.

8. A storage medium, characterized in that, The storage medium stores a multi-model integrated prediction program for power load, which, when executed by a processor, implements the steps of the multi-model integrated prediction method for power load as described in any one of claims 1-5.

Images