Multi-energy load prediction method based on feature screening and multi-model fusion

Abstract

The invention discloses a multi-energy load prediction method based on feature screening and multi-model fusion, and belongs to the field of electric power and comprehensive energy load prediction. The invention provides a three-stage hybrid learning prediction framework. In the first stage, dynamic feature screening is achieved through a recursive feature elimination cross validation method based on expert knowledge constraints, key meteorological elements and multi-energy load time sequence features are reserved, and redundant feature interference is reduced. In the second stage, a multi-task long-short-term memory network is constructed, and coupling relation modeling and collaborative prediction of cold, heat and electricity multi-energy loads are achieved through sharing time sequence characteristic representation and a task exclusive output structure. And in the third stage, a random forest is adopted to carry out nonlinear correction on the residual error of the sub-model, so that the precision and robustness of prediction in sudden disturbance and local non-stationary scenes are improved, the prediction error is effectively reduced, and the stability of multi-energy load prediction is improved. And reliable support is provided for optimized operation, scheduling decision and renewable energy consumption of the park integrated energy system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power and comprehensive energy load prediction, and in particular to a multi-energy load prediction method based on feature screening and multi-model fusion. BACKGROUND

[0002] With the wide construction of park comprehensive energy systems and the continuous growth of urban energy demand, multi-energy loads such as cold, heat and electricity show significant seasonal fluctuations and complex time-coupling relationships. Influenced by extreme weather such as high temperature in summer and severe cold in winter, the cooling and heating demand in the park will rapidly rise in a short time, and multiple energy loads may rise simultaneously, thereby increasing the pressure of energy supply, dispatching and distribution network operation. At the same time, due to the differences in different industrial types, building purposes and operation laws, the load characteristics of each park show obvious diversity and instability, which puts forward higher requirements for multi-energy joint prediction.

[0003] At present, multi-energy load prediction generally adopts independent modeling or prediction based on a single machine learning model, often ignoring the internal correlation between multiple energies due to equipment coupling, working condition association and meteorological driving. The traditional method usually relies on manual experience to select input features, lacks a systematic feature screening and evaluation mechanism, and is easy to introduce redundant features and noise features, resulting in low model accuracy. In addition, under the background of non-linear, multi-scale and non-stationary characteristics of multi-energy load, a single model is difficult to balance the accuracy, stability and generalization ability of multi-energy prediction; at the same time, the lack of effective residual correction mechanism makes the prediction prone to systematic deviation under extreme weather or sudden load fluctuation conditions.

[0004] Therefore, in order to adapt to the needs of fine operation and energy efficiency optimization of park comprehensive energy systems, a new prediction method is needed that can automatically screen key influencing factors, model the coupling relationship of multi-energy load, and has error correction capability. Such a method should be able to integrate meteorological, time, load and other multi-source data, take into account the common and difference characteristics between multiple energies, and improve the robustness of prediction under complex working conditions, in order to support the scientific formulation of park energy dispatching, equipment operation optimization and load management strategies. SUMMARY

[0005] To address the challenges of significant coupling, complex and redundant features among cold, heat, and electricity loads in multi-energy systems within industrial parks, the difficulty of traditional single-task prediction models in simultaneously characterizing the temporal features of multiple energy sources, and the insufficient prediction accuracy of existing methods under extreme weather or load surge scenarios, this invention proposes a multi-energy load prediction method for industrial parks based on feature optimization and multi-task deep learning. This method integrates dynamic feature selection, multi-task temporal modeling, and random forest residual correction techniques, enabling the model to maintain high prediction accuracy and stability even under complex multi-energy synergy and drastic environmental changes. This meets the practical needs of optimized scheduling, demand response strategy formulation, and load security assessment for integrated energy systems in industrial parks.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution: A multi-energy load forecasting method based on feature selection and multi-model fusion includes the following steps: Step S1: Collect historical cooling load, heating load, and electrical load sequences of the target area, and align them with meteorological and temporal characteristics to construct an initial feature set that includes load lag characteristics, meteorological characteristics, and periodic temporal characteristics; Step S2: Dynamic feature selection is performed using the Recursive Feature Elimination and Cross-Validation (RFECV) algorithm. While retaining the core feature set determined based on domain knowledge, redundant or low-contribution features are eliminated to obtain the optimal feature subset. Step S3: Input the optimal feature subset into the multi-task long short-term memory network (MTL-LSTM) model, use the shared feature representation layer to learn the implicit coupling relationship between cold, hot and electrical loads, and generate preliminary prediction values ​​for the three types of loads through the task-specific output layer respectively; Step S4: Calculate the residuals between the preliminary predicted values ​​and the actual values, and use the optimal feature subset and the preliminary predicted values ​​as inputs to the random forest regression model to learn the mapping relationship between the preliminary predicted values ​​and the residuals, thereby outputting the residual correction values. Finally, add the preliminary predicted values ​​and the corresponding residual correction values ​​to obtain the corrected prediction results for cooling load, heating load, and electrical load.

[0007] Furthermore, the construction of the initial feature set specifically includes: First, time-series data of cooling, heating, and electrical loads are collected from the historical operational data of the target park's integrated energy system, and external driving data strictly aligned with this time resolution is obtained. This external driving data includes meteorological and temporal characteristics. Meteorological characteristics include at least temperature, humidity, wind speed, and solar radiation intensity. To fully characterize the cumulative and delayed effects of meteorological conditions on energy load, features of multiple discontinuous lag times are simultaneously introduced for key meteorological factors, in addition to their current values. The temporal characteristics are derived from date and time information, including hour, day type, and weekday. These data collectively constitute the set of external driving factors for load changes.

[0008] Secondly, to characterize the time dependence and periodicity of the load itself, multi-scale lag features are constructed for the cold, heat, and electricity load sequences. Specifically, this includes: extracting continuous short-term lag values ​​from t-1 to t-24 to characterize the short-term inertia and intraday variation pattern of the load; and simultaneously extracting the weekly lag value at t-168 to capture the long-term periodic characteristics of the load affected by the weekly operating pattern.

[0009] Furthermore, the time features are preprocessed to enhance their periodic expressive power. Discrete hour and weekday variables are mapped into continuous periodic encoding vectors using a sine-cosine function to eliminate the step discontinuity of calendar time, allowing it to smoothly represent the day-night and week cycle in the feature space.

[0010] Before constructing the initial feature set, the original historical operational data needs to be preprocessed. Anomaly detection methods based on quantile thresholds are used to identify and remove outliers that significantly deviate from the normal range; missing values ​​and outliers are repaired using a time-neighborhood-based mean interpolation method. Specifically, for missing or outlier values ​​at time t... The interpolation result is calculated by referencing historical samples from the same time over several days:

[0011] in, This represents the historical load value that is at the same intraday cycle position as the current time t, and N is the number of historical samples used to construct the time neighborhood. By utilizing historical information consistent with intraday cycles, the temporal structure of the load sequence can be preserved to the greatest extent while repairing abnormal or missing data, making the repaired sequence closer to the actual operating state in terms of fluctuation characteristics and trend pattern.

[0012] Finally, the preprocessed load sequence, its multi-scale lag characteristics, meteorological characteristics, and encoded periodic time characteristics are combined to construct a high-dimensional initial feature set.

[0013] Furthermore, step S2 employs the Recursive Feature Elimination and Cross-Validation (RFECV) algorithm for dynamic feature selection, specifically including: First, the initial feature set is divided into two parts: a core feature set pre-determined based on domain knowledge, and the remaining candidate features excluding the core features. The core feature set is fixed and retained throughout the algorithm and does not participate in subsequent recursive elimination processes.

[0014] Furthermore, the core feature set determined based on domain knowledge specifically includes: Key driving factors with clear physical meaning or stable relevance in the energy system include at least the key lag characteristics of the load itself, key lag terms characterizing the impact of temperature, and periodic time characteristics encoded by sine-cosine. Their role is to ensure that critical variables essential to business operations are not mistakenly removed due to statistical fluctuations.

[0015] Furthermore, the recursive feature elimination and cross-validation process for the candidate feature set specifically includes: Specifically, to avoid the leakage of future information due to random data shuffling in time series forecasting, time series cross-validation is used in the feature evaluation stage. Let the sample set arranged in chronological order be represented as:

[0016] in, The input feature vector has dimension D; For load forecasting.

[0017] In a total of K folds, the training set for the r-th fold and verification set The definition is as follows:

[0018] For any candidate feature subset Its temporal cross-validation loss The mean squared error (MSE) was calculated over all K-fold validation sets to obtain the following:

[0019] This loss serves as a criterion for evaluating the quality of feature subsets.

[0020] Secondly, the importance of features is calculated using a random forest regression model as the base learner, specifically the importance of feature j in the t-th decision tree. The following calculation is based on a weighted sum of the impurity reductions at all splitting nodes caused by this feature:

[0021] in, Let represent the set of all internal nodes in the t-th random forest tree that are split by feature j, where for any node... , Let m be the sample set before the split. and These represent the sample subsets of the left and right child nodes generated after the node is split; (symbols) This represents the sample size of the corresponding sample set and the node weights. This is to ensure that the contribution of different nodes to the overall feature is proportional to their sample size. This represents the sample variance of the target load and is used to measure node impurity.

[0022] The combined importance of feature j across all T trees The average importance of it in a single tree:

[0023] Based on the above importance assessment, the recursive feature elimination process progressively removes the features with the lowest importance. In the r-th iteration, the feature with the lowest importance is determined by the following formula:

[0024] in, This represents the current set of candidate features in the r-th iteration. This is the feature importance score obtained from the random forest in the previous stage. Then, this feature is removed from the set, resulting in the candidate feature set for the next iteration.

[0025] As the number of iterations gradually increases, the number of candidate features gradually decreases. The corresponding cross-validation loss is calculated in each iteration. This is used to track performance changes in feature subsets. To achieve optimal generalization performance in feature selection, the optimal candidate feature subset is... Defined as the combination of features that minimizes the cross-validation loss across all iteration rounds:

[0026] Finally, the core feature set P determined based on domain knowledge is combined with the optimal candidate feature set obtained through recursive feature elimination. Together they form, namely:

[0027] in, This represents the feature set that is ultimately input into the model, and its dimension is determined by the union of the two types of features in the formula.

[0028] Furthermore, in step S3, the optimal feature subset is input into a multi-task long short-term memory network (MTL-LSTM) model for joint prediction, specifically including: First, the optimal feature subset is organized in chronological order into a sliding window sequence of length T, which is then used as the model input:

[0029] in, This represents the filtered feature vector at time tt. is the dimension of the optimal feature subset.

[0030] Furthermore, the MTL-LSTM model consists of a shared LSTM encoding layer and three task-specific output layers, specifically including: Shared LSTM coding layers are used to extract from the input sequence The common cross-task temporal features and coupling relationships among cold, hot, and electrical loads are extracted synchronously. This layer can contain multiple stacked LSTM units, and the computation of each unit within time step t follows the following gating mechanism:

[0031] in, , , These represent the forget gate, input gate, and output gate, respectively. and These represent the cell state and its candidate values; It is in a hidden state; Use the Sigmoid activation function; This indicates element-wise multiplication; (·)and The high-level temporal representation of the final output of the trainable parameter-sharing layer for the corresponding gate. It is simultaneously fed into three task-specific output layers.

[0032] Furthermore, the three task-specific output layers are independent fully connected networks, specifically including: Each task-specific layer shares a representation. Using the input as input, preliminary predictions for the corresponding energy type are generated independently through linear transformation and activation functions:

[0033] Where k=1,2,3 correspond to the cold, heat, and electricity load forecasting tasks, respectively; and These are the trainable weights and biases for the k-th task-specific layer. This structure enables the model to learn unique dynamic patterns for various workloads based on shared information.

[0034] Furthermore, the MTL-LSTM model employs an uncertainty-weighted mechanism for multi-task joint training, and its loss function construction and optimization objectives specifically include: The prediction error for each task is defined as the mean squared error (MSE):

[0035] in, This is the set of time indices for the training set.

[0036] The model's total loss function is the learnable weighted sum of the losses from each task:

[0037] in, This is a learnable log-variance parameter used to characterize the inherent noise level and prediction uncertainty of task k. This parameter is optimized synchronously with the model weights during training, and its loss weights can be adaptively adjusted according to the noise intensity and error scale of different tasks, enhancing the stability and robustness of the prediction results.

[0038] By optimizing the model parameters by minimizing the total loss function, the finally trained MTL-LSTM model can simultaneously output a sequence of preliminary predicted values ​​for cooling, heating, and electrical loads as input for the subsequent residual correction stage.

[0039] Furthermore, the residual correction based on the random forest model in step S4 specifically includes: First, the residuals between the initial predictions generated by the Multi-Task Long Short-Term Memory (MTL-LSTM) model and the actual load values ​​are calculated. For each prediction time t and each energy type k (k=1,2,3 corresponding to cooling, heating, and electrical loads respectively), the residuals are calculated. Defined as:

[0040] in, This is the actual load value. These are the initial predictions output by the MTL-LSTM model. The residual sequence reflects the systematic biases and nonlinear fluctuations that the deep learning model failed to fully capture in local time periods or under specific operating conditions.

[0041] Furthermore, the input feature vector used to train the random forest regression model is constructed in the following specific ways: To learn the mapping relationship between residuals and system states, the optimal feature subset obtained in step S2 is... Compared with the preliminary predicted values ​​of all energy types obtained in step S3 The features are concatenated to form an enhanced feature vector. :

[0042] This vector simultaneously contains the original external driving factors, historical load characteristics, periodic information, and high-level temporal representations extracted by the deep learning model, providing comprehensive conditional information for residual modeling.

[0043] Furthermore, the establishment and training process of the random forest residual correction model specifically includes: For each energy type k, construct an independent random forest residual regression model. Let... Let represent the prediction function of the m-th regression tree in the forest for task k. If the forest contains M trees, then the corresponding residual estimate is defined as:

[0044] Random forests construct training subsets through Bootstrap resampling and divide each tree node based on the criterion of minimizing the target load variance, thereby learning the nonlinear structure of the residuals.

[0045] Finally, the residual correction value is superimposed on the initial prediction value of the MTL-LSTM to obtain the final prediction result:

[0046] Through the aforementioned residual correction mechanism, random forests can effectively compensate for local irregular patterns that deep learning models fail to fully capture under complex operating conditions, thereby significantly improving the overall robustness and accuracy of predictions while maintaining the time-series modeling capabilities of LSTM. This hybrid MTL-LSTM-RF model achieves higher generalization ability in multi-energy load prediction tasks, enabling high-precision prediction of cooling, heating, and electrical loads.

[0047] The present invention has the following beneficial effects: This invention addresses the problem that in multi-energy systems within industrial parks, cold, heat, and electricity loads are significantly coupled due to the combined influence of meteorological conditions and operational modes. Traditional forecasting methods struggle to simultaneously consider the characteristics of multiple energy sources, lack feature selection mechanisms, and exhibit low prediction accuracy under extreme weather or load surge scenarios. To address this, a multi-energy load forecasting method based on feature selection and multi-model fusion is proposed. This method constructs a comprehensive multi-source feature system encompassing load, meteorological, and temporal data, integrating recursive feature elimination and domain knowledge constraints to automate the selection of key driving factors for multiple energy sources. It learns the temporal coupling patterns among cold, heat, and electricity loads through a multi-task long short-term memory network, utilizing shared feature representations and task-specific output structures to balance the commonalities and differences among different energy sources. Furthermore, a random forest residual correction mechanism is introduced to nonlinearly correct the systematic biases of the deep learning model. This method can accurately capture load change trends and significantly improve prediction stability and robustness even under conditions of strong time-varying loads, high coupling complexity, and significant external disturbances. This provides reliable data support and technical means for integrated energy dispatching, multi-energy coordinated optimization operation, and load peak early warning in industrial parks. Attached Figure Description

[0048] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below: Figure 1 This is the overall flowchart of the present invention; Figure 2 This is a schematic diagram of the multi-task long short-term memory network (MTL-LSTM) model structure in this invention; Figure 3 This is a comparison chart of typical daily forecast results and error statistics in an embodiment of the present invention; Figure 4 This is a scatter plot of the predicted residual distribution according to an embodiment of the present invention. Detailed Implementation

[0049] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The advantages and features of the present invention will become clearer from the following description. It should be noted that the drawings are in a very simplified form and use non-precise proportions, used only to facilitate and clearly illustrate the embodiments of the present invention. Please refer to the drawings to make the objectives, features, and advantages of the present invention more apparent and understandable. It should be understood that the structures, proportions, sizes, etc., depicted in the accompanying drawings are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed in the specification, and are not intended to limit the implementation conditions of the present invention. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to the size, without affecting the effects and objectives achieved by the present invention, should still fall within the scope of the technical content disclosed in the present invention.

[0050] This embodiment relates to a multi-energy load forecasting method based on feature selection and multi-model fusion. Although the present invention has been described in detail through the above preferred embodiments, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above content. Therefore, the scope of protection of the present invention should be defined by the appended claims.

[0051] This invention focuses on a typical commercial park-level integrated energy system in East China. The data used comes from hourly cooling, heating, and electrical load data recorded throughout 2021 by the park's Building Energy Management System (BEMS), as well as synchronous meteorological data collected by the park's meteorological monitoring station. To verify the performance of the proposed prediction method, the 8760 hours of valid sample data for the entire year were divided chronologically into a training set, a validation set, and a test set in a ratio of 8:1:1 to ensure the timeliness and fairness of the evaluation.

[0052] This embodiment strictly follows the workflow described in this invention: "data preprocessing and feature construction → expert knowledge-guided RFECV feature selection → MTL-LSTM collaborative prediction → RF residual correction". The detailed operation and parameter settings for each step are as follows: Step 1: Data Preprocessing and Feature Construction In this embodiment, historical data of the multi-energy microgrid is first preprocessed and features are constructed to ensure data quality and the validity of model input. During data cleaning, the interquartile range (IQR) method is used for outlier detection. In the face of invalid data, outliers, and missing data, the time-series mean imputation method is used to handle missing values, thereby preserving the time dependence of load data.

[0053] Subsequently, a high-dimensional initial feature set containing time attributes, historical loads, and meteorological factors was constructed, totaling 84 dimensions, as shown in Table 1.

[0054] The time characteristics mainly include hours, days of the week, and week number, which can capture the periodic patterns between daily routines and cycles; meteorological characteristics include temperature, humidity, radiation, and wind speed, which have significant correlations with load, especially temperature, which has a significant impact on cooling and electrical loads; load characteristics include historical load data for electricity, heat, and cooling, and combine time delay terms (lag characteristics) to describe the system's inertia and cross-energy coupling characteristics.

[0055] Step 2: Expert-guided recursive feature elimination and cross-validation To screen key driving factors and eliminate redundancy from high-dimensional features, this embodiment employs an RFECV method that integrates domain knowledge. First, based on the understanding of the PIES energy flow coupling mechanism, a set of core features (P) is predefined and forcibly retained. This set includes: encoded time-period features, key temperature lag terms (t-1, t-3, t-6), and autocorrelation terms (t-1, t-24 lag terms) for each energy load core. For the remaining features, a Random Forest (RF) is used as the base estimator. Specific parameter settings are determined using a grid search method, with the following ranges: the number of decision trees in the Random Forest is 100, 150, 200, or 400; the maximum depth is 8, 12, or 20; and the cross-validation folds K=5. In each iteration, the importance scores of all candidate features are calculated, the features with the lowest scores are removed, and the cross-validation performance of the new feature subset is evaluated (using mean squared error (MSE) as the metric). Finally, the feature subset that minimizes the cross-validation loss is selected. The final model input feature set is the union of the feature subset with the minimum cross-validation loss and the core feature set preserved based on domain knowledge.

[0056] Step 3: Multi-task Long Short-Term Memory Network (MTL-LSTM) Collaborative Prediction The filtered feature sequences are organized chronologically into sliding windows of 24 hours in length, and then input into a multi-task Long Short-Term Memory (MTL-LSTM) network for collaborative prediction. Figure 2 As shown, this model adopts a shared-proprietary architecture, specifically consisting of a shared LSTM encoding layer and three independent task-specific fully connected output layers.

[0057] The shared LSTM encoding layer is responsible for synchronously extracting and learning the common temporal dynamic patterns and coupled correlation features behind the three types of loads—cold, heat, and electricity—from the input multi-energy feature sequences. This layer consists of two stacked LSTM units, with each layer having 128 hidden units, and a dropout operation with a dropout rate of 0.2 is used between layers to prevent overfitting. This encoding layer achieves effective modeling of long-term dependencies through its internal forget gate, input gate, and output gate mechanisms.

[0058] Based on the temporal features learned by the shared LSTM layers, three task-specific output layers independently predict cold, hot, and electrical loads, respectively. The model employs the Adam optimizer with an initial learning rate of 0.001, a batch size of 64, and a maximum training epoch of 200. The following strategies are implemented to improve training efficiency and generalization ability: early stopping (training stops when the validation set loss fails to decrease for 20 consecutive epochs), gradient clipping (maximum gradient norm of 5 to prevent gradient explosion), and L2 weight decay (coefficient of 0.0001 to prevent overfitting). Furthermore, the model uses a multi-task loss function based on uncertainty, adaptively adjusting its weight in the total loss to achieve automatic balancing and collaborative optimization of multi-task learning, avoiding the subjectivity of manual parameter tuning.

[0059] Step 4: Random Forest (RF) Residual Correction MTL-LSTM can effectively capture the temporal dependencies and coupling characteristics of multi-energy loads, but systematic prediction biases may still occur in scenarios with sudden load changes, extreme weather disturbances, or enhanced local nonlinearities. To further improve the accuracy and robustness of the model, this embodiment introduces a random forest (RF) residual correction module to compensate for errors in the initial predictions of MTL-LSTM.

[0060] Specifically, the predicted residual sequences of cooling, heating, and electrical loads output by the MTL-LSTM are first calculated, representing the differences between the actual loads and the preliminary predicted values. Then, the final feature set obtained in the feature selection stage is used as input, combined with the preliminary MTL-LSTM predictions, to construct an RF residual regression model, which learns the nonlinear error structure inherent in the residuals. Finally, the residual compensation output by the RF correction model is superimposed with the preliminary prediction results of the MTL-LSTM to obtain the optimized final predicted values ​​for cooling, heating, and electrical loads.

[0061] To verify the effectiveness of the three-stage method of "RFECV feature optimization + MTL-LSTM collaborative prediction + RF residual correction", this embodiment selects four benchmark models, namely LSTM, MTL-LSTM, Pearson+MTL-LSTM, and RFECV+MTL-LSTM, as control groups, and conducts model performance tests under the same dataset partitioning, feature preprocessing, and training configuration. The evaluation metrics for the three types of workloads include: mean absolute percentage error (MAPE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²).

[0062] Table 2. Model Performance Comparison Results

[0063] As can be seen from the comparison results in Table 2, the RFECV+MTL-LSTM-RF joint prediction method proposed in this invention has achieved the best performance in the prediction tasks of cold, heat and electricity loads, and the improvement is significant. In terms of cooling load forecasting, the RMSE decreased from 1.859 GJ / h of the traditional LSTM to 1.159 GJ / h (a decrease of approximately 37.6%), the MAPE decreased to 4.65%, and the R² increased to 0.932, indicating a significant improvement in both fitting accuracy and stability. In terms of heat load forecasting, the RMSE decreased from 2.636 GJ / h of the MTL-LSTM to 2.081 GJ / h, the MAPE decreased to 2.79%, and the R² increased to 0.926, demonstrating that the model more accurately portrays changes in the heating system load. In terms of electricity load forecasting, the RMSE decreased from 2.809 kW to 1.641 kW (a decrease of over 41%) compared to the ordinary LSTM, the MAPE decreased to 3.01%, and the R² increased to 0.936, indicating significant improvements in both forecasting accuracy and peak-valley tracking capability.

[0064] Figure 3 This is a comparison chart of typical daily forecast results and error statistics according to an embodiment of the present invention. The left side shows a comparison of hourly forecast curves for three types of loads—cold, hot, and electric—on a typical day; the right side shows the corresponding relative error box plot. The results show that RFECV+MTL-LSTM-RF outperforms the control model in trend following ability and error convergence characteristics in all three load forecasting tasks.

[0065] Regarding hourly forecast curves, both traditional LSTM and Pearson+MTL-LSTM models exhibit varying degrees of lag in capturing rapid load increases and decreases, with significant biases in local fluctuations. While the RFECV+MTL-LSTM model improves overall smoothness through feature selection, it still retains some errors near peak and trough values. In contrast, the RFECV+MTL-LSTM-RF method achieves a high degree of fit to the actual load curve across peak, trough, and drastically changing ranges for both cooling, heating, and electricity loads, significantly improving upon the baseline model's lag bias and local oscillation issues.

[0066] The error box plot further validated the above conclusions from a statistical perspective. For cooling load forecasting, the mean relative error of RFECV+MTL-LSTM-RF was only 0.0329, significantly lower than that of RFECV+MTL-LSTM (0.0503) and Pearson+MTL-LSTM (0.0535); its error box was narrower, with lower dispersion and fewer outliers, indicating a significant improvement in forecast stability. In heat load forecasting, the mean error of RFECV+MTL-LSTM-RF was 0.0250, about 16% lower than the model without residual correction, and the upper limit of the error contracted significantly, indicating that this method can effectively suppress peak forecasting errors. Electricity load forecasting also performed excellently, with the mean error of RFECV+MTL-LSTM-RF at 0.0279, better than the comparison models, and its error distribution had the smallest box range and the least dispersion, demonstrating the best overall robustness.

[0067] Figure 4 This is a scatter plot of the predicted residual distribution from an embodiment of the present invention, visually comparing the error characteristics of different models in predicting cold, hot, and electric loads. Overall, the Pearson+MTL-LSTM model exhibits the most discrete residual scatter plot distribution, with a large number of high-error points significantly deviating from the zero axis. The RFECV+MTL-LSTM model narrows the residual range through feature filtering, but its upper and lower boundaries remain relatively wide, indicating the existence of uncaptured systematic biases. In contrast, the RFECV+MTL-LSTM-RF method proposed in this invention shows a residual point cloud highly concentrated near the zero axis, with the vast majority of scatter points falling within a very narrow band, significantly reducing the number of outliers and significantly suppressing the overall fluctuation of the residuals.

[0068] Specifically, for various load types, the RFECV+MTL-LSTM-RF model exhibited the most compact residual distribution band and the smallest dispersion, confirming that the random forest residual correction module effectively identified and compensated for the nonlinear systematic errors in the initial prediction of MTL-LSTM, thereby significantly improving the stability and accuracy of the final prediction. In contrast, other comparative models, especially during periods of drastic or large-fluctuation nonlinear load changes, still showed a clear trend of diffused or asymmetrical residual distribution.

[0069] The accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent. The embodiments described above merely illustrate several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of this patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. A multi-energy load forecasting method based on feature selection and multi-model fusion, characterized in that, Includes the following steps: Step S1: Collect historical operational data for the target area, including cooling load, heating load, electrical load sequences, and time-aligned meteorological and temporal characteristics; An initial feature set containing multi-dimensional input variables is constructed based on the historical operational data; Step S2: The Recursive Feature Elimination and Cross-Validation (RFECV) algorithm is used to retain the core feature set based on domain knowledge, and the remaining features are dynamically selected to obtain the optimal feature subset; Step S3: Input the optimal feature subset into the Multi-Task Long Short-Term Memory Network (MTL-LSTM) model. The MTL-LSTM model learns the coupling temporal relationship between cold, hot and electrical loads through a shared feature representation layer, and outputs preliminary predicted values ​​of cold load, hot load and electrical load through task-specific prediction layers respectively. Step S4: Calculate the residual sequence between the preliminary predicted value and the actual load value; train a random forest model with the optimal feature subset and the preliminary predicted value as input to learn the mapping relationship between the preliminary predicted value and the residual, and obtain the residual correction value based on the random forest model; The preliminary predicted value is added to the corresponding residual correction value to obtain the final predicted values ​​for cooling, heating, and electrical loads.

2. The multi-energy load forecasting method based on feature selection and multi-model fusion according to claim 1, characterized in that, The construction of the initial feature set in step S1 includes: (1) Extract multiple time lag features from the cooling load, heating load and electrical load sequences respectively. The time lag features include short-term lag values ​​from t-1 to t-24 and periodic lag values ​​from t-168. (2) Introduce meteorological features that are time-aligned with the load sequence, the meteorological features including at least temperature, humidity, wind speed and solar radiation; (3) Generate time features based on date and time information, the time features including hours and weekdays, and perform periodic sine-cosine encoding on the hours and weekdays; (4) Preprocessing the historical operation data, including: removing outliers by an anomaly detection method based on quantile thresholds and repairing missing data by time neighborhood interpolation.

3. The multi-energy load forecasting method based on feature selection and multi-model fusion according to claim 2, characterized in that, The recursive feature selection process in step S2 specifically includes: (1) The training set and the validation set are divided by time series cross-validation strategy to maintain the causal order of the time series and avoid future information leakage; (2) Using the random forest model as the base learner, evaluate the importance score of each feature in each iteration and remove a preset number of features with the lowest importance score; (3) During the recursive feature elimination process, the core feature set based on the domain knowledge is always retained. The core feature set includes key temperature lag features, key load lag features, and encoded periodic time features.

4. The multi-energy load forecasting method based on feature selection and multi-model fusion according to claim 3, characterized in that, The core feature set pre-determined based on domain knowledge includes at least the following three types of features: Multiple temperature hysteresis features characterizing the effect of temperature, including short-term and discontinuous hysteresis terms; key hysteresis features characterizing load autocorrelation, including short-period and daily-period hysteresis features for cooling, heating, and electrical loads respectively. And the encoding features that characterize time periodicity, namely periodic time features generated based on hour and weekday information and encoded by sine-cosine functions.

5. The multi-energy load forecasting method based on feature selection and multi-model fusion according to claim 1, characterized in that, The multi-task long short-term memory network (MTL-LSTM) in step S3 consists of a shared LSTM encoding layer and three independent task-specific output layers; the shared LSTM encoding layer is used to synchronously extract cross-task temporal features common to cold, hot and electrical loads from the optimal feature subset; The three task-specific output layers are fully connected layers, which take the high-level time series representation output by the shared coding layer as input and independently generate preliminary predicted values ​​for cold, heat, and electrical loads.

6. The multi-energy load forecasting method based on feature selection and multi-model fusion according to claim 5, characterized in that, The MTL-LSTM model employs an uncertainty weighting mechanism for multi-task joint training; its overall loss function is the weighted sum of the prediction errors of each task, where the weight of each task is automatically adjusted by the learnable parameters of the corresponding task, adaptively balancing the contribution of different prediction tasks in training.

7. The multi-energy load forecasting method based on feature selection and multi-model fusion according to claim 1, characterized in that, The random forest residual correction process in step S4 includes: training a random forest regression model by using the optimal feature subset and the preliminary prediction values ​​generated by the MTL-LSTM model as input vectors; the random forest model constructs multiple regression trees through Bootstrap resampling and divides them based on the criterion of minimizing the variance of node samples, so as to learn the mapping relationship between the preliminary prediction values ​​and the residuals and output the residual correction values.

8. The multi-energy load forecasting method based on feature selection and multi-model fusion according to claim 7, characterized in that, The random forest model includes three independent residual regression models established for cold, heat, and electricity loads, respectively. Each residual regression model uses the residual sequence of the corresponding energy type and its associated optimal feature subset and preliminary prediction value as training samples, and outputs only the residual correction value of that energy type.

9. The multi-energy load forecasting method based on feature selection and multi-model fusion according to claim 7, characterized in that, The recursive feature elimination and cross-validation adopts five-fold time series cross-validation. The input of the multi-task long short-term memory network model is a sliding time window sequence of length 24, and its shared feature representation layer is composed of two stacked LSTM units.

10. The multi-energy load forecasting method based on feature screening and multi-model fusion according to claim 8, characterized in that, The hyperparameters of the random forest model were determined through joint optimization using grid search and time series cross-validation.

Images