Comprehensive energy system multi-load prediction method, device, equipment, medium and product
By optimizing the MVMD penalty factor and dynamic model selection mechanism through Q-learning, multivariate loads are adaptively decomposed into high-frequency, low-frequency, and residual subsequences, which solves the problem of low accuracy in multivariate load prediction and achieves high-precision and strong generalization capability in comprehensive energy system load prediction.
Patent Information
- Application Number
- CN202511718823.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-21
- Publication Date
- 2026-03-06
AI Technical Summary
Existing load forecasting methods are unable to effectively characterize the coupling characteristics between multiple loads. Traditional decomposition methods are prone to mode aliasing, and the decomposition parameters rely on human experience, resulting in low accuracy of multivariate load forecasting and weak model generalization ability.
The Q-learning algorithm is used to optimize the penalty factor in multivariate variational mode decomposition (MVMD). Combined with sample entropy (SE) and dynamic model selection mechanism (DMS), the multivariate loading is adaptively decomposed into high-frequency, low-frequency and residual subsequences, and the optimal base model is dynamically selected for prediction.
It significantly improves the accuracy of multivariate load forecasting, reduces RMSE by 2% to 15%, and increases R2 by 1% to 10%. It is applicable to short-term load forecasting of integrated energy systems of different climate zones and scales, and has strong generalization ability and engineering applicability.
Smart Images
Figure CN121615097A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of load forecasting for integrated energy systems, and in particular to a method, apparatus, equipment, medium, and product for multi-load forecasting of integrated energy systems. Background Technology
[0002] Building a new power system with new energy sources as the mainstay has become a core path, but the problem of absorbing a high proportion of renewable energy urgently needs to be solved. Integrated energy systems, by breaking down barriers between electricity, heat, cooling, and gas, achieve multi-energy synergy and efficient conversion, serving as a key vehicle for improving energy utilization efficiency and promoting the absorption of new energy. However, the strong randomness and intermittency of renewable energy sources such as wind and solar power, as well as the complex spatiotemporal coupling relationships among diverse loads, result in highly non-stationary and drastically fluctuating load sequences, posing significant challenges to real-time system scheduling and economic operation. Against this backdrop, high-precision short-term forecasting of diverse loads is a crucial foundation for the optimized operation of integrated energy systems. Its accuracy directly affects system stability, operating costs, and the level of clean energy absorption. Due to the limited research in this area, many issues still urgently need to be addressed.
[0003] Existing load forecasting methods are mostly designed for single energy types, making it difficult to effectively characterize the coupling characteristics between multiple loads. Meanwhile, traditional decomposition methods (such as EMD and CEEMDAN) are prone to mode aliasing, and their decomposition parameters rely on human experience, affecting the quality of feature extraction. Furthermore, a single forecasting model cannot simultaneously adapt to the modeling needs of different frequency components (such as trends, periods, and noise), resulting in limited overall forecast accuracy.
[0004] Therefore, there is an urgent need for a high-precision short-term forecasting method that can adaptively decompose multiple loads, accurately reconstruct frequency band characteristics, and dynamically match the optimal forecasting model to support the safe, economical, and efficient operation of integrated energy systems. Summary of the Invention
[0005] The purpose of this application is to provide a method, device, equipment, medium and product for multi-load forecasting of integrated energy systems, which can solve the shortcomings of existing technologies such as low accuracy of multi-load forecasting, weak model generalization ability and dependence of decomposition parameters on experience.
[0006] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a method for predicting multiple loads in an integrated energy system, including: Acquire historical time series data of multivariate loads and perform preprocessing; The preprocessed historical time series data of multivariate loads were divided into training and test sets; Adaptive optimization of the penalty factor in multivariate variational mode decomposition (MVMD) based on the Q-learning algorithm; The optimized penalty factor was used to perform MVMD decomposition on the preprocessed multivariate load history time series data to obtain several intrinsic mode components (IMF) and one residual component. Calculate the sample entropy SE of each IMF component and residual component, set a threshold based on the SE value, and reconstruct all components into high-frequency subsequences, low-frequency subsequences, and residual subsequences; A multivariate load forecasting model MVMD-SF-DMS is constructed; the MVMD-SF-DMS multivariate load forecasting model includes a preset base model pool and a dynamic model selection mechanism DMS. The MVMD-SF-DMS multivariate load prediction model is trained based on the training set. For the high-frequency subsequence, low-frequency subsequence, and residual subsequence, the optimal base model is dynamically selected using the DMS mechanism for prediction, and the prediction results for each subsequence are obtained. The DMS mechanism is based on the sliding window historical performance evaluation and the Softmax probability update strategy, and is the prediction model with the highest self-scheduled adaptability for each type of subsequence. The prediction results of each subsequence are weighted and summed according to the original reconstruction weights to obtain the final short-term prediction values of multiple loads including electricity, heat, cooling, and gas.
[0007] Optionally, the Q-learning algorithm models the optimization of the penalty factor as a reinforcement learning task, wherein the state space is a set of discrete candidate values for the penalty factor, the action space includes three operations: "increase", "decrease", and "remain unchanged", the immediate reward is defined by the inverse of the mean square error between the MVMD reconstructed signal and the original signal, and the Q table is iteratively updated through the Bellman equation, and finally the state corresponding to the largest Q value is selected as the optimal penalty factor.
[0008] Optionally, the expression for the optimal penalty factor is as follows: ; in, Indicates the optimal penalty factor. This represents the minimum value of the penalty factor. This represents the state corresponding to the maximum value in table Q. This indicates the step size of the penalty factor.
[0009] Optionally, the calculation of the sample entropy SE uses embedding dimension and similarity tolerance. IMF components with SE values below a set threshold are classified as low-frequency subsequences, and those above a high-frequency threshold are classified as high-frequency subsequences. The residual components are treated separately as mid-frequency or trend subsequences.
[0010] Optionally, the base model pool includes four models: XGBoost, LightGBM, SVR, and BiLSTM; among which, XGBoost or LightGBM is preferentially used for high-frequency and residual subsequence prediction, and SVR or BiLSTM is preferentially used for low-frequency subsequence prediction.
[0011] Optionally, the DMS mechanism samples all base models with equal probability during the initialization phase, calculates the weighted composite score of each model within the sliding window through k-fold cross-validation in subsequent phases, dynamically updates the model selection probability based on the Softmax function, and finally determines the model used for the current prediction through Monte Carlo sampling.
[0012] Secondly, this application provides a multi-load forecasting device for an integrated energy system, the multi-load forecasting device for an integrated energy system comprising: The preprocessing module is used to acquire and preprocess historical time series data of multivariate loads. The data partitioning module is used to divide the preprocessed historical time series data of multivariate loads into training and test sets. The optimization module is used to adaptively optimize the penalty factor in multivariate variational mode decomposition (MVMD) based on the Q-learning algorithm. The decomposition module is used to perform MVMD decomposition on the preprocessed multivariate load historical time series data using the optimized penalty factor, to obtain several intrinsic mode components (IMF) and a residual component. The reconstruction module is used to calculate the sample entropy SE of each IMF component and residual component, and to reconstruct all components into high-frequency subsequences, low-frequency subsequences and residual subsequences based on the SE value. A multivariate load forecasting model construction module is used to construct the MVMD-SF-DMS multivariate load forecasting model; the MVMD-SF-DMS multivariate load forecasting model includes a preset base model pool and a dynamic model selection mechanism (DMS); The training module is used to train the MVMD-SF-DMS multivariate load prediction model based on the training set. The prediction module is used to dynamically select the optimal base model for prediction of the high-frequency subsequence, low-frequency subsequence and residual subsequence using the DMS mechanism, respectively, to obtain the prediction results of each subsequence; wherein the DMS mechanism is based on the sliding window historical performance evaluation and Softmax probability update strategy, and is the prediction model with the highest self-scheduling adaptability for each type of subsequence. The weighted summation module is used to sum the prediction results of each subsequence according to the original reconstruction weights to obtain the final short-term prediction values of multiple loads such as electricity, heat, cooling and gas.
[0013] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the integrated energy system multi-load forecasting method described in any one of the above.
[0014] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the integrated energy system multi-load forecasting method described above.
[0015] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the integrated energy system multi-load forecasting method described above.
[0016] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a method, apparatus, equipment, medium, and product for predicting multiple loads in an integrated energy system, which has the following beneficial effects: Improved decomposition accuracy: By adaptively optimizing the MVMD penalty factor through Q-learning, the blindness of traditional grid search or empirical setting is avoided, and the accuracy and physical interpretability of mode decomposition are significantly improved. Enhanced feature separability: Sample entropy is introduced to perform complex quantification of IMF components, achieving effective separation of high frequency (transient / noise), low frequency (trend), and residual (period), laying the foundation for subsequent differential modeling; Achieving dynamic model adaptation: The DMS mechanism can automatically schedule the optimal prediction model based on the characteristics of the subsequence (e.g., LightGBM is good at high frequencies, and BiLSTM is good at low frequencies), giving full play to the advantages of various models; Improved overall forecast accuracy: Significantly outperforms single models and traditional decomposition methods (such as CEEMDAN) across multiple load types (electricity, heat, cooling, gas), reducing RMSE by 2%–15%, and R... 2 Increase by 1% to 10%; It possesses strong generalization ability and engineering applicability: the methodology is clear, the parameters are adaptive, and no manual intervention is required. It is applicable to short-term load forecasting scenarios for integrated energy systems of different climate zones and scales. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 A flowchart illustrating a multi-load forecasting method for an integrated energy system provided in an embodiment of this application; Figure 2 This is a box plot of seasonal load characteristics from an embodiment of this application; Figure 3 This is a box plot of monthly load characteristics from an embodiment of this application; Figure 4 This is a diagram showing the reconstruction results of the IMF components of the electrical load in this application; Figure 5 This is a graph showing the calculation results of the multivariate load IMF sample entropy in an embodiment of this application; Figure 6 This is a graph showing the results of multivariate load forecasting in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation
[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0020] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0021] In one exemplary embodiment, such as Figure 1 As shown, a multi-load forecasting method for an integrated energy system is provided. This method is executed by computer equipment, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, it includes the following steps: Step 101: Obtain historical time series data of multivariate loads and perform preprocessing.
[0022] Step 102: Divide the preprocessed multivariate load historical time series data into training set and test set.
[0023] Historical data on various loads, including electricity, heat, cooling, and gas, were acquired, encompassing meteorological data, date-specific sequence data, and multivariate load data. This data was then analyzed and normalized to form the model's input data. The collected historical load data and time data were used to construct the model dataset. Date and time data, along with a portion of the historical load data, served as input to the training set, while other load data were used as output comparison data. The ratio of the training set to the test set was 7:3.
[0024] Step 103: Adaptively optimize the penalty factor in multivariate variational mode decomposition (MVMD) based on the Q-learning algorithm.
[0025] The MVMD method, optimized using Q-learning (QL), decomposes multivariate load time series into intrinsic mode components (IMFs) of multiple frequencies, thereby extracting richer signal features and accurately capturing subtle changes, rapid fluctuations, and transient behaviors in the sequence. This will help the prediction model better learn and capture the frequency characteristics of the load, improving prediction accuracy.
[0026] For containing Input data for each data channel (i.e., historical data on electricity, cooling, heating, and gas loads). MVMD needs to be decomposed into multiple IMF components. To solve the constrained optimization problem: in, It is the number of IMF components (i.e., the number of modes). for Harmonic shift, It is a modal component. Let be the center frequencies of each mode. To solve this constrained optimization problem, the Lagrange operator is introduced. and penalty factor Transforming the above equation into an unconstrained optimization form: The above equation is solved iteratively using the alternating direction multiplier algorithm, and the following iterative update formula is given: in, Indicates the first In the next iteration, the channel Inner Each component at frequency The estimated value at that location, Indicates channel Internal signal at frequency The estimated value at that location, Indicates channel Inner Each component In frequency The estimated value at that location, Indicates channel Inner Lagrange multipliers at frequency The estimated value at that location, This represents a frequency variable, used to describe the characteristics of a signal in the frequency domain. Indicates the first The center frequency corresponding to each component Indicates the first In the nth iteration, the 1st Center frequency estimates of each component Indicates channel Inner Each component at frequency The estimated value at that location.
[0027] The iterative constraints are: In the formula: As the stopping condition for iteration, The iteration stops when the above iterative condition is met.
[0028] QL (Quick Decision) is a classic algorithm in reinforcement learning that emphasizes obtaining the optimal decision through continuous trial and error, correction, and refinement during interaction with the environment, thereby achieving the optimal result. This strategy constructs a relationship between the target state and the action state. The value table is updated during interactions between the agent and the environment using a reward and punishment mechanism. The table selects the action that yields the maximum benefit. To address the issue of the penalty factor α depending on experience and grid search during MVMD decomposition, this application models the optimization of the α parameter as a reinforcement learning task under discrete states and employs the Q-learning algorithm to learn the optimal α online. The optimization process is as follows: 1) Construct the state space S: exist Discretize the interval with a fixed step size to obtain a set of candidate parameters and define the state space. : in, For the first One candidate penalty factor, For the first A state quantity, This represents the minimum value of the penalty factor. This represents the maximum value of the penalty factor. This represents the discretization step size of the penalty factor. N This indicates the number of candidate penalty factors.
[0029] 2) Constructing the action space Action sets contain three basic operations: In the formula, , , Corresponding to Increase, decrease, or remain unchanged.
[0030] Then based on Mechanism, intelligent agent The probability of choosing the optimal action is used to... The probability of choosing a random action is expressed as: In the formula: A random number uniformly distributed between [0,1]. The greed coefficient is set to 0.8 in this application.
[0031] Based on the selected action For the current Update the value; use the updated value. For the original signal Perform multivariate variational mode decomposition (MVMD) to compute the reconstructed signal. The instant reward is calculated as follows: in, This is the signal length.
[0032] Subsequently, the Bellman equation was used to update... surface: In the formula: The learning factor is used to control the learning speed; in this application, it is set to 0.1. The discount factor is set to 0.9 in this application.
[0033] Repeat the above calculation and update process until the iteration count condition is met. Based on the number of training rounds, reset the exploration rate after each round. When the number of rounds reaches a preset value or... After the table converges, select The state corresponding to the maximum value in the value table And calculate the optimal penalty factor. : The above process can be completed online on the training sample set and decisions can be made through the QL strategy. This application sets the maximum number of iterations to 500, sets a reasonable reward and punishment mechanism, and based on the prediction error, QL can adaptively optimize the MVMD penalty factor. This reduces the reliance on parameter tuning for empirical applications and grid search, improves the accuracy and training efficiency of MVMD decomposition, and provides an optimal decomposition basis for subsequent processing.
[0034] Step 104: Perform MVMD decomposition on the preprocessed multivariate load historical time series data using the optimized penalty factor to obtain several intrinsic mode components (IMFs) and one residual component.
[0035] Step 105: Calculate the sample entropy SE of each IMF component and residual component, set a threshold based on the SE value, and reconstruct all components into high-frequency subsequences, low-frequency subsequences, and residual subsequences.
[0036] SE is an improvement on approximate entropy, offering better consistency in measuring time series complexity. By comparing the multivariate load IMF components and residual components obtained from MVMD decomposition using SE values, time series in three different frequency bands—high-frequency, low-frequency, and residual—are reconstructed to analyze data characteristics and reduce the complexity of prediction models. The SE calculation process is as follows: First, given the time series , convert the sequence Composed of multiple in sequence If the vector is dimensional, then: In the formula: .
[0037] Then calculate and Maximum difference between corresponding elements The expression is: for Calculate the difference The number of values is calculated, and its proportion in the total number of differences is denoted as . and for all Take the average value : In the formula: for The number.
[0038] Increase the dimension to Repeat the above steps to obtain and The average value is then used to calculate the theoretical value of the sample entropy of the random sequence: Finally, the SE threshold of the IMF component is calculated. : in From 1 to The index represents the first... The sample entropy statistics of a sample or component. This represents the number of samples or components involved in the calculation.
[0039] Studies have shown that the calculated SE value depends on the embedding dimension. Similarity tolerance This application selects and Calculate the entropy value of the data components.
[0040] Step 106: Construct the MVMD-SF-DMS multivariate load forecasting model; the MVMD-SF-DMS multivariate load forecasting model includes a preset base model pool and a dynamic model selection mechanism (DMS).
[0041] To address the problem of insufficient prediction accuracy caused by the complex coupling characteristics and strong non-stationarity of multi-load sequences in integrated energy systems, this application proposes a multi-load prediction framework, MVMD-SF-DMS, which integrates reinforcement learning optimization decomposition and dynamic model selection. A multi-dimensional fluctuation index evaluation system is designed, and a feature-driven dynamic model selection mechanism is used to autonomously schedule the optimal prediction model for different frequency band sequences, constructing a full-process adaptive prediction system for MVMD-SF-DMS. For the reconstructed high-frequency, low-frequency, and residual sub-sequences and auxiliary features, this application constructs the following optimal prediction models respectively. The feature-driven dynamic model selection mechanism can selectively choose the most suitable prediction model based on the fluctuation characteristics of the current sequence, thereby significantly improving the overall prediction accuracy.
[0042] The DMS mechanism targets the components obtained from multivariate variational mode decomposition and achieves precise matching between component characteristics and model advantages through dynamic model adaptation.
[0043] Define model pool Through weighted composite score Quantify model performance.
[0044] In the formula, Represents the number of samples; It is the fold index of cross-validation; For model indexing; Refers to the first Compromise The true value of each sample; Indicates the first The model for the first Compromise Predicted values for each sample; It is the first The mean of all true values.
[0045] The DMS mechanism comprises four core components: In the initialization phase (the first 56 days), each model participates in prediction with equal probability. Historical performance was then evaluated using a sliding window. Dynamically update the selection probability based on the Softmax function: Finally, the model MVMD-SF-DMS was determined through Monte Carlo sampling, a mechanism that achieves a precise match between component characteristics and model strengths.
[0046] Step 107: Train the MVMD-SF-DMS multivariate load prediction model based on the training set.
[0047] Step 108: For the high-frequency subsequence, low-frequency subsequence and residual subsequence, the optimal base model is dynamically selected using the DMS mechanism for prediction to obtain the prediction results for each subsequence; wherein the DMS mechanism is based on the sliding window historical performance evaluation and Softmax probability update strategy, and is the prediction model with the highest self-scheduled adaptability for each type of subsequence.
[0048] Step 109: The prediction results of each subsequence are weighted and summed according to the original reconstruction weights to obtain the final short-term prediction values of multiple loads of electricity, heat, cooling and gas.
[0049] By inputting the date type, time data, and historical electricity, heat, cooling, and gas load data of the day to be predicted into the trained prediction model, the predicted values of electricity, heat, and cooling loads for that day can be obtained.
[0050] Load characteristic analysis, a crucial step in comprehensive energy load forecasting, is the core basis for distinguishing different datasets. It not only visually presents the characteristics of various loads in a comprehensive energy system but also provides a reliable data foundation for forecasting models. This example uses box plots to analyze the comprehensive energy load dataset from multiple time dimensions, based on data from the entire year of 2016. Specifically, it covers the four seasons (spring, summer, autumn, and winter) and the twelve months of the year, and studies four types of loads: electricity, heat, cooling, and gas.
[0051] Box plots include elements such as median, maximum, minimum, first quartile, third quartile, mean, and outliers. The median reflects the middle level of the corresponding time scale, the mean represents the average level of the corresponding time interval, and the box section covers the main load range of 25% to 75%.
[0052] Seasonal characteristics are used for load analysis with a quarter as the time scale. The seasons are divided into: spring (March to May), summer (June to August), autumn (September to November), and winter (December to February of the following year). Seasonal characteristics represent the energy consumption of the Tempe campus throughout the year. Figure 2 The following are seasonal distribution charts for electricity, heat, cooling, and gas. Figure 2 Part a of the diagram is a box plot showing the seasonal characteristics of cooling load, part b is a box plot showing the seasonal characteristics of heating load, part c is a box plot showing the seasonal characteristics of electrical load, and part d is a box plot showing the seasonal characteristics of gas load.
[0053] Cooling load increases significantly in summer, peaking at 400,000 Tons / h, while the narrowest box is in winter, indicating that cooling demand is concentrated during high-temperature periods. Heating load shows the opposite trend, with the highest median and peak value in winter, approaching 370 mmBTU / h, while the shortest box and least volatile load are in summer. Electricity load is driven by demand from air conditioning and other sources in summer, with a median exceeding 850,000 kW and a wide fluctuation range; demand is relatively stable in winter, with a median of approximately 500,000 kW. Gas load fluctuates moderately, with a median of approximately 1250 Tons / h in summer and approximately 600 Tons / h in winter, and is more concentrated in the box, indicating stable heating demand. Overall, cooling and electricity loads peak in summer and decline significantly in winter, while heating load peaks in winter and bottoms out in summer. Gas load is relatively stable throughout the year, with higher concentration in winter. Outliers are few in each season and have little impact on the overall distribution. These seasonal differences reveal the pattern of energy demand driven by climate change, providing a key reference for multi-energy load forecasting and scheduling optimization.
[0054] The monthly load characteristics are statistically analyzed based on the monthly time scale, covering all 12 months of the year, reflecting the energy usage of the Tempe campus in each month of the year. Figure 3The monthly box plots show the electricity, heat, cooling, and gas loads. Figure 3 Part a in the diagram represents the monthly characteristic box plot of cooling load, part b represents the monthly characteristic box plot of heating load, part c represents the monthly characteristic box plot of electrical load, and part d represents the monthly characteristic box plot of gas load.
[0055] Cooling load gradually increases from spring, with significant increases in both the cooling and overhead power supply from June to August, peaking at nearly 380,000 Tons / h in July. It then rapidly declines after autumn, with the lowest cooling load in December and a median below 100,000 Tons / h, indicating that cooling demand is highly concentrated in the high-temperature months. Heating load shows the opposite trend to cooling: the highest cooling and overhead power supply occurs in January and February, with a peak of approximately 360 mmBTU / h in January; the lowest heating load occurs from May to July, with a median around 110 mmBTU / h, indicating negligible heating demand in summer. Electricity load follows a similar trend to cooling load, with a median exceeding 850,000 kW in summer and the widest fluctuation range; the median in December is approximately 550,000 kW, and the load is more concentrated in the cooling and overhead power supply. Gas load fluctuates moderately monthly, with a median of approximately 1350 Tons / h in August and approximately 600 Tons / h in January and February, with slightly wider load ranges in the mid-to-high temperature months compared to the low temperature months. Overall, cooling and electricity loads peak in summer and decline significantly in winter, while heating load is highest in winter and lowest in summer. Gas load shows a steady upward trend from early winter to mid-summer. There are few outliers throughout the year, and their impact on the overall distribution is limited. These monthly characteristics reveal the response patterns of multi-energy loads to climate change, providing an important basis for short-term load forecasting and scheduling optimization.
[0056] In model performance evaluation and result analysis, the selection of error evaluation indicators and the rationality of data preprocessing directly affect the reliability of the conclusions. This application uses root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²). 2 ) and comprehensive evaluation indicators As evaluation criteria, RMSE reflects the overall error level between predicted and actual values, and is more sensitive to larger errors; MAPE presents the error magnitude intuitively as a percentage, facilitating error comparison of data of different magnitudes; R... 2 This measures the model's ability to explain data variation; the closer the value is to 1, the better the model fits the data.
[0057] To ensure the stability of model training and the scientific validity of evaluation index calculation, the raw data needs to be processed. Data normalization is the first step in preprocessing. Its core purpose is to eliminate the influence of differences in the units and magnitudes of different feature data, ensuring that all indicators are under a unified unit when calculating the comprehensive evaluation index, and avoiding the dominance of a certain feature in the model training or evaluation results due to an excessively large numerical range. The raw data is mapped to the [0,1] interval, and the calculation formula is as follows: in, The original data, and These are the minimum and maximum values in the feature data, respectively. This is the result after normalization. This operation not only accelerates the convergence speed of the model and improves training efficiency, but also ensures that each evaluation index has equal weight and reference value in the comprehensive calculation, thus ensuring the objectivity of the evaluation results.
[0058] To determine the optimal number of decomposition levels for MVMD, over-decomposition is defined as occurring when the Pearson correlation coefficient (PCC) between a sub-mode and the original data is less than 0.1. The calculation formula is as follows: In the formula, It is the PCC value; and It is the first of two sequences One sample; and The value is the average of two sequences. Based on calculations, this application determines the decomposition level to be 5, meaning a single load sequence is decomposed into 5 modal components and a single residual sequence. MVMD effectively reduces the non-stationarity of complex sequences. After decomposition, each characteristic mode function can be considered a unique feature extracted from the original load sequence. Through accurate decomposition and localization, an appropriate number of IMFs are obtained, providing effective additional feature information for load forecasting, enriching the input feature space of the forecasting model, and enhancing the model's ability to capture load changes. To effectively achieve grouping and reconstruction of IMF components, the SE values of all IMF components generated by the decomposition are calculated. Simultaneously, an IMF frequency threshold is set as the basis for component division. The calculation results are as follows: Figure 5 As shown. Taking electrical load as an example, from Figure 5As can be seen, the SE values of IMF1, IMF3, and IMF4 components are below the frequency threshold, indicating that these three components contain more trend components and the data is relatively stable. Therefore, these three IMF components are superimposed as low-frequency components for prediction. The SE values of IMF2 and IMF5 are above the mid-to-high frequency threshold, indicating that these components contain more nonlinear components, and they are classified as high-frequency components. The residual sequence contains more periodic feature information from the original sequence, so it is treated as a mid-frequency component. Taking electrical load as an example, the decomposed and reconstructed modal components are as follows: Figure 4 As shown.
[0059] Because different labels correspond to different dynamic characteristics, the same model cannot adapt to the information contained in different components. Therefore, this application selects the optimal base model corresponding to different components and constructs XGBoost, SVR, BiLSTM, and LightGBM models respectively to predict the high-frequency, low-frequency, and residual components of the electricity, cooling, heating, and gas loads of the integrated energy system. The prediction results of different models are presented below. , , and comprehensive indicators As shown in Table 1.
[0060] Table 1: Comparison of Base Model Performance
[0061] (a)
[0062] (b)
[0063] (c)
[0064] (d) The results in Table 1 show that the LightGBM and XGBoost models exhibit significant advantages in predicting high-frequency and residual components, with their accuracy significantly outperforming other baseline models. This is because LightGBM and XGBoost, as Gradient Boosting Decision Tree (GBDT) frameworks, can efficiently learn complex and nonlinear local patterns and interactions. High-frequency components typically contain rapidly fluctuating, detailed transient information and noise from the time series, while residual components contain nonlinear and non-stationary fluctuations in the original sequence that are difficult to explain by deterministic models. The tree model's splitting mechanism, feature importance evaluation, and ability to handle high-dimensional sparse features make it particularly adept at capturing the complex, fine-grained patterns and irregular structures contained in these components. In contrast, the SVR and BiLSTM models perform better in predicting low-frequency components. This is mainly due to the fact that the SVR model, when optimizing the regularized risk function, can effectively model smooth, global functional relationships through the kernel function; while the BiLSTM model, with its gating mechanism and bidirectional information flow, can fully learn and remember the long-term trends, slowly evolving periodic patterns, and historical dependencies inherent in the time series, which are the core components of low-frequency components. Finally, this application adopts a model selection and fusion strategy based on component characteristics: for each independent component obtained after decomposition, the single model with the best prediction performance on that component is selected, and the prediction results of these selected models are weighted and summed according to the reconstruction weights of their corresponding components, thereby integrating them to obtain the final comprehensive prediction result with higher accuracy.
[0065] To verify the superiority and generalization ability of the proposed model MVMD-SF-DMS, the CEEMDAN-SF-DMS (M2) model and the proposed MVMD-SF-DMS (M1) model were constructed to predict the electricity, cooling, and heating loads of the integrated energy system. Simultaneously, to verify the effect of DMS on improving model accuracy, the predictions of the base models MVMD-SF-XGBOOST (M3), MVMD-SF-SVR (M4), MVMD-SF-BiLSTM (M5), and MVMD-SF-LightGBM (M6) in the model pool were selected for ablation experiments for comparison. The results are shown in Table 2 and... Figure 6 As shown.
[0066] Table 2: Comparison of Multiple Load Forecasting Indicators
[0067] Combining the performance of different models in various load forecasting scenarios, from the perspectives of error and adaptability, MVMD-SF-DMS has a significant advantage over CEEMDAN-SF-DMS. CEEMDAN decomposition is prone to mode aliasing due to noise interference, leading to chaotic component feature extraction. In contrast, MVMD, based on variational mode decomposition theory, can more accurately separate different frequency components. In multivariate load data, it makes the physical meaning of each component clearer, allowing subsequent models to learn more efficiently. Furthermore, the results show that dynamic model selection exhibits significant advantages in processing multivariate load data: in total load forecasting, MVMD-SF-DMS... The score reached 0.9561, an improvement of approximately 1.32% compared to the outstanding MVMD-SF-LightGBM. The value is 0.0359, compared to 0.0367 for MVMD-SF-XGBOOST, representing a reduction of approximately 2.18%. Gradient boosting decision tree models like XGBoost and LightGBM, with their splitting mechanism, efficiently capture complex nonlinear relationships in load data and outperform SVR and BiLSTM in most load forecasting scenarios. However, ensemble forecasting models, by selecting the best single model for each load component and then weighting and summing them according to the reconstruction weights, integrate the advantages and compensate for the fitting shortcomings of single models in specific components. This significantly improves the accuracy of multi-energy load forecasting, laying a solid foundation for reliable energy system load forecasting and providing an effective solution.
[0068] The results show that the framework significantly improves the robustness and accuracy of multivariate load forecasting under fluctuating environments, providing reliable technical support for the efficient scheduling and stable operation of integrated energy systems.
[0069] In one exemplary embodiment, a multi-load forecasting device for an integrated energy system is provided, comprising: The preprocessing module is used to acquire and preprocess historical time series data of multivariate loads. The data partitioning module is used to divide the preprocessed historical time series data of multivariate loads into training and test sets. The optimization module is used to adaptively optimize the penalty factor in multivariate variational mode decomposition (MVMD) based on the Q-learning algorithm. The decomposition module is used to perform MVMD decomposition on the preprocessed multivariate load historical time series data using the optimized penalty factor, to obtain several intrinsic mode components (IMF) and a residual component. The reconstruction module is used to calculate the sample entropy SE of each IMF component and residual component, and to reconstruct all components into high-frequency subsequences, low-frequency subsequences and residual subsequences based on the SE value. A multivariate load forecasting model construction module is used to construct the MVMD-SF-DMS multivariate load forecasting model; the MVMD-SF-DMS multivariate load forecasting model includes a preset base model pool and a dynamic model selection mechanism (DMS); The training module is used to train the MVMD-SF-DMS multivariate load prediction model based on the training set. The prediction module is used to dynamically select the optimal base model for prediction of the high-frequency subsequence, low-frequency subsequence and residual subsequence using the DMS mechanism, respectively, to obtain the prediction results of each subsequence; wherein the DMS mechanism is based on the sliding window historical performance evaluation and Softmax probability update strategy, and is the prediction model with the highest self-scheduling adaptability for each type of subsequence. The weighted summation module is used to weight and sum the prediction results of each sub-sequence according to the original reconstruction weights to obtain the final short-term prediction values of multiple loads (electricity, heat, cooling, and gas). It includes: In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 7 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores multi-load forecasting data for integrated energy systems. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a multi-load forecasting method for integrated energy systems.
[0070] Those skilled in the art will understand that Figure 7 The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.
[0071] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.
[0072] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.
[0073] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.
[0074] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).
[0075] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.
[0076] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0077] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A method for multi-load forecasting of an integrated energy system, characterized in that, The comprehensive energy system multi-load prediction method comprises: obtaining historical time series data of multi-element load and preprocessing; dividing the historical time series data of the preprocessed multi-element load into a training set and a test set; based on the Q-learning algorithm, the penalty factor in the multi-element variation mode decomposition (MVMD) is adaptively optimized; using the optimized penalty factor to perform MVMD decomposition on the preprocessed multi-element load historical time series data, to obtain a plurality of intrinsic mode components (IMFs) and a residual component; calculating the sample entropy (SE) of each IMF component and residual component, setting a threshold value according to the SE value, and reconstructing all components into high-frequency subsequences, low-frequency subsequences and residual subsequences; constructing an MVMD-SF-DMS multi-load prediction model; the MVMD-SF-DMS multi-load prediction model comprises a pre-set base model pool and a dynamic model selection mechanism (DMS); training the MVMD-SF-DMS multi-load prediction model based on the training set; for the high-frequency subsequences, low-frequency subsequences and residual subsequences, respectively, using the DMS mechanism to dynamically select the optimal base model for prediction to obtain the prediction results of each subsequence; wherein the DMS mechanism is based on a sliding window historical performance evaluation and a Softmax probability updating strategy, and independently schedules the prediction model with the highest adaptation degree for each type of subsequence; weighting and summing the prediction results of each subsequence according to the original reconstruction weight to obtain the final short-term prediction values of the electric, heat, cold and gas multi-element load. 2.The integrated energy system multi-load forecasting method according to claim 1, characterized in that, The Q-learning algorithm models the optimization of the penalty factor as a reinforcement learning task, wherein the state space is a set of discrete candidate values of the penalty factor, the action space includes three operations of "increase", "decrease" and "keep unchanged", the immediate reward is defined by the reciprocal of the mean square error between the MVMD reconstruction signal and the original signal, and the Q table is updated iteratively through the Bellman equation, and finally the state corresponding to the maximum Q value is selected as the optimal penalty factor. 3.The integrated energy system multi-load forecasting method of claim 1, wherein, The expression of the optimal penalty factor is as follows: ； wherein, represents the optimal penalty factor, represents the minimum value of the penalty factor, represents the state corresponding to the maximum value in the Q-table, represents the step size of the penalty factor. 4.The method of claim 1, wherein, The calculation of the sample entropy (SE) adopts embedding dimension and similarity tolerance, and the IMF components with SE values lower than the set threshold are classified into low-frequency subsequences, and the IMF components with SE values higher than the high-frequency threshold are classified into high-frequency subsequences, and the residual component is processed as a medium-frequency or trend subsequence. 5.The integrated energy system multi-load forecasting method of claim 1, wherein, The base model pool comprises four models of XGBoost, LightGBM, SVR and BiLSTM; wherein XGBoost or LightGBM is preferentially used for high-frequency and residual subsequence prediction, and SVR or BiLSTM is preferentially used for low-frequency subsequence prediction. 6.The integrated energy system multi-load forecasting method of claim 1, wherein, The DMS mechanism equally samples all base models in the initialization stage, calculates the weighted comprehensive score of each model in the sliding window through k-fold cross-validation in the subsequent stage, dynamically updates the model selection probability based on the Softmax function, and finally determines the model used for the current prediction through Monte Carlo sampling.
7. An integrated energy system multi-load forecasting device, characterized in that, The comprehensive energy system multi-load prediction device comprises: a preprocessing module for obtaining historical time series data of multi-element load and preprocessing; The data division module is configured to divide the preprocessed historical time series data of the multi-energy load into a training set and a test set. The optimization module is configured to perform adaptive optimization on a penalty factor in multi-variate modal decomposition (MVMD) based on a Q-learning algorithm. The decomposition module is configured to perform MVMD decomposition on the preprocessed historical time series data of the multi-energy load by using the optimized penalty factor, to obtain a plurality of intrinsic modal components (IMFs) and a residual component. The reconstruction module is configured to calculate sample entropy (SE) of each IMF component and the residual component, set a threshold according to the SE value, and reconstruct all the components into a high-frequency subsequence, a low-frequency subsequence, and a residual subsequence. The multi-energy load prediction model construction module is configured to construct an MVMD-SF-DMS multi-energy load prediction model. The training module is configured to train the MVMD-SF-DMS multi-energy load prediction model based on the training set. The prediction module is configured to predict the high-frequency subsequence, the low-frequency subsequence, and the residual subsequence by respectively using the DMS mechanism to dynamically select an optimal base model, to obtain prediction results of the subsequences. The weighted summation module is configured to perform weighted summation on the prediction results of the subsequences according to original reconstruction weights, to obtain final short-term prediction values of the electric, heating, cooling, and gas multi-energy loads.
8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the multi-energy load prediction method of the integrated energy system according to any one of claims 1-6.
9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the multi-energy load prediction method of the integrated energy system according to any one of claims 1-6.
10. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the multi-energy load prediction method of the integrated energy system according to any one of claims 1-6.