A multi-user power consumption prediction method considering upstream and downstream linkage
By constructing a multi-dimensional feature system and linkage matrix, and combining it with an ensemble learning model, the problem of failing to effectively utilize upstream and downstream linkage information in existing technologies has been solved, thereby improving the accuracy and reliability of multi-user electricity consumption forecasting for highly correlated industrial clusters such as chemical manufacturing.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- STATE GRID SHANGHAI MUNICIPAL ELECTRIC POWER CO
- Filing Date
- 2026-03-16
- Publication Date
- 2026-07-10
AI Technical Summary
Existing electricity consumption forecasting methods fail to effectively utilize upstream and downstream linkage information, resulting in the inability to achieve accurate and reliable multi-user electricity consumption forecasting in highly interconnected industrial clusters. This is especially true in industries such as chemical manufacturing, where traditional methods lack modeling of supply chain transmission items, leading to significantly increased forecasting bias.
A multi-dimensional feature system is constructed, and upstream and downstream relationships are identified through homogeneous clustering and time-delay cross-correlation analysis. A linkage matrix is constructed, and an integrated learning model is built based on the linkage matrix. The lag term and weighted synchronization term of the upstream enterprise load are introduced as features to collaboratively predict the downstream enterprise load and to verify and correct the consistency of the industrial chain.
It significantly improves the accuracy and robustness of multi-user electricity consumption forecasting, and can accurately respond to the impact of upstream maintenance or shutdown events on downstream loads, thereby improving the accuracy and reliability of forecasting.
Smart Images
Figure CN122371074A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power load analysis and forecasting technology, and in particular to a multi-user power consumption forecasting method that takes into account upstream and downstream linkages. Background Technology
[0002] In modern manufacturing and energy management, electricity forecasting is a core component. Accurate forecasting is not only related to daily production planning, equipment maintenance, and energy consumption control, but also affects electricity costs, power supply reliability, and connection with the electricity market.
[0003] Manufacturing clusters, exemplified by the chemical industry, are characterized by long industrial chains, strong upstream and downstream linkages, and high production continuity. However, in actual operation, the maintenance, shutdowns, or load adjustments of individual enterprises often transmit synchronously or delayed to their downstream partners through product supply and energy flow paths, leading to adjustments in the electricity load of downstream enterprises. This interconnected coupling relationship within the industrial chain is a significant internal factor contributing to fluctuations in the electricity load of industrial clusters. Existing electricity forecasting methods, whether employing trend extrapolation, classical time series models, or traditional machine learning models, mostly treat enterprises as independent entities for forecasting. These methods ignore or fail to effectively utilize this upstream-downstream linkage information, making it difficult to handle clusters of highly homogeneous and correlated enterprises. Particularly when upstream enterprises undergo planned maintenance or sudden shutdowns, traditional methods, lacking modeling of the industrial chain transmission factors, result in significantly increased forecasting biases for downstream enterprise loads. Therefore, existing methods struggle to meet the stringent requirements for accurate electricity sales forecasting under the context of electricity market reform, especially in highly interconnected industrial clusters such as chemical manufacturing, where they cannot handle coupled loads within the industrial chain or achieve accurate and reliable multi-user electricity forecasting. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a multi-user electricity consumption forecasting method that considers upstream and downstream linkages. This method can explicitly model the upstream and downstream linkages of the industrial chain and incorporate them as core input features into multi-user collaborative forecasting, thereby improving the accuracy of multi-user electricity consumption forecasting.
[0005] The objective of this invention can be achieved through the following technical solution: a multi-user electricity consumption forecasting method considering upstream and downstream linkages, comprising the following steps: S1. Construct a multi-dimensional feature system that includes electricity consumption time-series characteristics, production and maintenance characteristics, external environmental characteristics, and industry tags; S2. For multidimensional feature systems, homogeneous clustering and time-delay cross-correlation analysis are used to identify upstream and downstream relationships and construct a linkage matrix; S3. Based on the linkage matrix, construct an integrated learning model, introduce the lag term and weighted synchronization term of the load of upstream enterprises as features, and perform collaborative prediction of the load of downstream enterprises. S4. Perform consistency verification and correction on the prediction results obtained in step S3 based on the industrial chain link, and output the single enterprise load prediction value and the total industry load prediction value.
[0006] Furthermore, the electricity consumption time sequence characteristics in S1 include monthly electricity consumption, daily electricity consumption, and electricity consumption at a specific time point, and the production and maintenance characteristics include maintenance plans, downtime status, and production continuity indicators.
[0007] Furthermore, S2 specifically employs a deep autoencoder (DAE) for dimensionality reduction preprocessing, followed by homogeneity clustering analysis based on the K-means algorithm; S2 uses the Pearson correlation coefficient to calculate the consistency of electricity changes among enterprises for time-delay cross-correlation analysis.
[0008] Furthermore, S2 specifically includes the following steps: S21. A deep autoencoder is used to extract and reduce the dimensions of high-dimensional electricity consumption features, and the corresponding low-dimensional feature vector is output. S22. For low-dimensional feature vectors, the K-means algorithm is used to identify homogeneity and identify groups of companies in the industry with similar electricity consumption behavior patterns. S23. Using the Pearson correlation coefficient, first calculate the synchronous correlation coefficient between enterprises, and then calculate the lagged correlation coefficient to determine the upstream and downstream linkage transmission relationship between enterprises. S24. Based on the synchronous correlation coefficient, the lag correlation coefficient and industry labels, a weighted fusion is performed to construct the upstream and downstream linkage matrix.
[0009] Furthermore, the formula for calculating the synchronization correlation coefficient in S23 is as follows: in, The synchronous correlation coefficient represents the correlation coefficient between enterprises. and enterprises Synchronous correlation between them; The formula for calculating the lagged correlation coefficient is: in, For the time step of lag, if exist If the value is reached at a certain time, then the enterprise is judged. As upstream, enterprises Downstream, and Its transmission lag period; when synchronicity and lag When both are significant, it indicates that there is a clear upstream and downstream linkage and transmission relationship between enterprises.
[0010] Furthermore, the upstream and downstream linkage matrix in S24 is specifically as follows: in, This refers to the synchronous correlation coefficient; The maximum lag correlation coefficient; Labels indicating upstream and downstream directions confirmed through research; These are the corresponding weighting coefficients.
[0011] Furthermore, S3 specifically employs the Extreme Gradient Boosting (XGBoost) algorithm to construct an ensemble learning model, wherein the input feature vector of the ensemble learning model contains a composite vector of supply chain transmission terms; The prediction function of the ensemble learning model consists of an additive model of multiple base learners. The ensemble learning model is optimized using a loss function with a regularization term, which is used to penalize the complexity of the model.
[0012] Furthermore, the composite vector includes the enterprise's own characteristics, the real-time weighted characteristics and the lagged weighted characteristics of upstream enterprises, external factor characteristics, and maintenance status characteristics.
[0013] Furthermore, when the maintenance status characteristics indicate that the enterprise is in a maintenance state, the prediction results output by the ensemble learning model will be corrected by hard constraints according to the maintenance plan.
[0014] Furthermore, S4 specifically addresses the prediction results of each enterprise. When the prediction results of an upstream enterprise are in a state of shutdown or low load, it verifies whether the prediction results of its downstream enterprise exceed the preset supply chain constraint range. If the verification finds an inconsistency, it triggers an automatic correction mechanism to adjust the prediction results of the downstream enterprise to the set range allowed by the industrial chain logic, and dynamically optimizes the parameters of the integrated learning model and the linkage matrix structure based on the prediction accuracy index.
[0015] Compared with the prior art, the present invention has the following advantages: This invention first constructs a multi-dimensional feature system based on historical electricity consumption data, production and maintenance plans, external environmental factors, and industry tags. Second, it identifies upstream and downstream relationships among enterprises through homogeneity clustering analysis and time-delay cross-correlation analysis, and constructs a quantified upstream-downstream linkage matrix. Then, using this linkage matrix as the core, it builds an ensemble learning model for multi-user collaborative prediction, introducing lag terms and weighted synchronization terms of upstream enterprise load as features to achieve collaborative prediction of downstream load. Finally, it performs consistency verification and correction on the prediction results based on the industrial chain link. This fully considers the linkage effect of upstream and downstream in the industrial chain, explicitly models the upstream and downstream linkage relationships, and utilizes the homogeneity among enterprises, upstream-downstream transmission relationships, and external influencing factors for multi-user collaborative prediction, thereby effectively solving the problem of insufficient prediction accuracy caused by neglecting the internal transmission relationships of the industrial chain.
[0016] This invention employs a deep autoencoder and the K-means algorithm for homogeneity identification, accurately identifying groups of enterprises within an industry with similar electricity consumption patterns, laying a structured foundation for subsequent identification of linkage relationships. Furthermore, using the Pearson correlation coefficient, it first calculates the synchronous correlation coefficient between enterprises, then the lagged correlation coefficient, determining the upstream and downstream linkage transmission relationships. Based on the synchronous correlation coefficient, lagged correlation coefficient, and industry labels, a weighted fusion is performed to construct an upstream and downstream linkage matrix, which accurately and comprehensively characterizes the load transmission relationship and coupling strength between enterprises. Using this linkage matrix as the core input feature provides structured and mechanism-driven load transmission constraint information for subsequent ensemble learning model prediction. By explicitly quantifying the upstream and downstream linkage relationships of the industrial chain into a linkage matrix and input features, this invention solves the problem that traditional algorithms cannot capture strong synchronous or lagged transmission relationships between enterprises, realizing a shift in the prediction model from a "single-enterprise independent perspective" to a "industrial chain collaborative perspective."
[0017] This invention integrates the linkage matrix into the input features of the ensemble learning model. The input feature vector of the ensemble learning model is designed as a composite vector containing the transmission term of the industrial chain. This composite vector includes the enterprise's own features, the real-time weighted features and lag weighted features of upstream enterprises, external factor features, and maintenance status features. This not only utilizes the homogeneity of electricity consumption among enterprises to improve the model's generalization ability, but also, by incorporating the lag term of upstream load, enables the model to effectively cope with the impact of planned or sudden events such as upstream maintenance and shutdown on downstream load, thereby improving the robustness of multi-user collaborative prediction.
[0018] When constructing the ensemble learning model, this invention considers the situation where the enterprise is under maintenance. When the enterprise is under maintenance, the load prediction value output by the ensemble learning model will be hard-constrained and corrected according to the maintenance plan. This allows for the linkage modeling of the production maintenance plan with the upstream and downstream transmission relationship, enabling accurate prediction of the significant load reduction effect downstream caused by upstream shutdowns. It is particularly suitable for industries with high industrial chain aggregation and high production continuity requirements, such as chemical, automobile manufacturing, and pharmaceutical manufacturing, and helps improve prediction accuracy.
[0019] This invention performs consistency verification on the upstream and downstream links of the forecast results. When an upstream enterprise is predicted to be in a state of shutdown or low load, the invention verifies whether the predicted load of its downstream enterprises exceeds a reasonable supply chain constraint. If inconsistency is found, an automatic correction mechanism is triggered to adjust the forecast value to a reasonable range allowed by the supply chain logic. Furthermore, the model parameters and linkage matrix structure are dynamically optimized based on the forecast accuracy index. This ensures that the predicted load of downstream enterprises does not exceed the maximum electricity load constrained by their upstream supply capacity, thus ensuring that the downstream forecast value does not exceed the upstream supply capacity and fully guaranteeing the reliability of the final forecast result. Attached Figure Description
[0020] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a flowchart illustrating the XGBoost algorithm in the embodiment. Figure 3 This is a flowchart illustrating the process of multi-user collaborative prediction. Detailed Implementation
[0021] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.
[0022] Example 1 To address the shortcomings of traditional electricity consumption forecasting, which only considers single users and neglects the interconnected effects of upstream and downstream industries, especially in situations with strong internal correlations and multiple coupled factors, this invention aims to develop a multi-user collaborative forecasting method that simultaneously leverages inter-enterprise homogeneity, upstream and downstream transmission relationships, and external influencing factors to significantly improve forecast accuracy. To achieve this objective, this invention provides a multi-user electricity consumption forecasting method that considers upstream and downstream linkages, such as... Figure 1 As shown, it includes the following steps: S1. Construct a multi-dimensional feature system that includes electricity consumption time-series characteristics, production and maintenance characteristics, external environment characteristics, and industry labels. Among them, electricity consumption time-series characteristics include monthly electricity consumption, daily electricity consumption, and electricity consumption at a specific time. Production and maintenance characteristics include maintenance plans, downtime status, and production continuity indicators. S2. For multidimensional feature systems, homogeneous clustering and time-delay cross-correlation analysis are used to identify upstream and downstream relationships and construct a linkage matrix; This embodiment specifically employs a deep autoencoder (DAE) for dimensionality reduction preprocessing, followed by homogeneous clustering analysis based on the K-means algorithm to identify groups of enterprises within the industry with similar electricity consumption patterns; and uses the Pearson correlation coefficient to calculate the consistency of electricity consumption changes among enterprises for time-delay cross-correlation analysis, obtaining synchronous correlation coefficients and lag correlation coefficients between enterprises to determine the upstream and downstream linkage transmission relationships between enterprises. Then, based on the synchronous correlation coefficient, the lag correlation coefficient, and industry labels, a weighted fusion is performed to construct the upstream and downstream linkage matrix; S3. Based on the linkage matrix, construct an integrated learning model, introduce the lag term and weighted synchronization term of the load of upstream enterprises as features, and perform collaborative prediction of the load of downstream enterprises. In this embodiment, the Extreme Gradient Boosting (XGBoost) algorithm is used to construct an ensemble learning model. The input feature vector of the ensemble learning model contains a composite vector of supply chain transmission terms (including enterprise features, real-time weighted features and lag weighted features of upstream enterprises, external factor features, and maintenance status features). The prediction function of the ensemble learning model consists of an additive model of multiple base learners. The ensemble learning model is optimized using a loss function with a regularization term, which is used to penalize the complexity of the model. Furthermore, when the maintenance status characteristics indicate that the enterprise is in a maintenance state, the prediction results output by the ensemble learning model will be corrected by hard constraints based on the maintenance plan. S4. Perform consistency verification and correction on the prediction results obtained in step S3 based on the industrial chain link, and output the single enterprise load prediction value and the total industry load prediction value. Specifically, for each enterprise's forecast results, when the upstream enterprise's forecast results are in a state of shutdown or low load, it checks whether the forecast results of its downstream enterprise exceed the preset supply chain constraint range. If the check finds an inconsistency, it triggers an automatic correction mechanism to adjust the downstream enterprise's forecast results to the set range allowed by the industrial chain logic, and dynamically optimizes the parameters of the integrated learning model and the linkage matrix structure based on the forecast accuracy index.
[0023] Through the above process, guided by the principles of supply chain structured analysis and multi-user collaborative learning, and employing homogeneous clustering, correlation identification, upstream and downstream linkage matrix construction, and ensemble learning models as the main technical means, this research on novel multi-user electricity consumption forecasting within the supply chain will help accurately depict the upstream and downstream relationships among enterprises within the industry, enhance the structured capabilities and interpretability of enterprise electricity consumption forecasting, reflect the transmission mechanism of load changes in the supply chain, and improve the accuracy and stability of industry-level electricity consumption trend judgment. Furthermore, by introducing enterprise correlation characteristics and a collaborative forecasting framework, a new modeling perspective is provided for industry electricity consumption situation awareness and key user monitoring, which can significantly improve the intelligence level of load forecasting.
[0024] Example 2 This embodiment applies the solution of Embodiment 1, collects historical electricity consumption, production and maintenance records and external environment of typical industrial chain enterprises in an industrial park, constructs a complete multi-dimensional feature system, and performs normalization preprocessing. Subsequently, the dimensionality of enterprise electricity consumption characteristics was reduced based on deep autoencoder (DAE), and homogeneous enterprise clusters were identified by combining K-means; at the same time, the transmission cycle and intensity between enterprises were determined by hysteresis correlation analysis, and an upstream and downstream linkage matrix was constructed accordingly.
[0025] Specifically, the process includes the following: 1. Analysis of Enterprise Homogeneity Enterprise homogeneity analysis is used to identify groups of enterprises within an industry that have similar electricity consumption behaviors. This embodiment employs a deep autoencoder for dimensionality reduction preprocessing, followed by clustering based on the K-means algorithm. This effectively captures the overall electricity consumption patterns of enterprises, reduces redundant information and dimensionality interference in high-dimensional data, and improves clustering results.
[0026] (1) Depth autoencoder dimensionality reduction process A deep autoencoder (DAE) is an unsupervised learning network used to compress high-dimensional power consumption features and extract key latent patterns. It learns features by minimizing the error between the input data and its reconstructed data. The specific process is as follows: Step 1: Input Layer Data Import and Encoding. This involves importing and encoding the original high-dimensional electricity consumption feature vector. The input is fed into the encoding network and processed through a nonlinear mapping. Transform into low-dimensional code vectors ( ): (1) in, and These are the weight matrix and bias vector of the encoder, respectively. This is the activation function.
[0027] Step 2: Decoding and Reconstruction Training. This involves training the decoding network. Low-dimensional code Map back to high-dimensional space to obtain the reconstructed vector. : (2) During training, the reconstruction error is continuously minimized through backpropagation (BP). : (3) Step 3: Feature Extraction and Dimensionality Reduction. After reaching the preset number of iterations, the trained encoding network can transform a high-dimensional index set similar to the training set into a low-dimensional code form that well expresses the original data features, thus completing efficient feature extraction and dimensionality reduction.
[0028] (2) K-means clustering Based on the low-dimensional feature vector output by DAE, the K-means algorithm is used for homogeneity identification. The algorithm steps are as follows: Step 1: Random selection One initial cluster center; Step 2: Calculate the Euclidean distance between each enterprise's feature vector and the cluster center, and reassign enterprise categories according to the minimum distance principle; Step 3: Update the cluster centers according to the newly assigned categories; Step 4: Repeat steps 2–3 until the cluster structure converges, meaning the cluster centers no longer change significantly.
[0029] Homogeneity analysis can identify groups of companies in the industry with similar electricity consumption patterns, laying a structured foundation for subsequent identification of interrelationships.
[0030] 2. Enterprise Correlation Analysis Based on Pearson Correlation Coefficient Enterprise correlation is used to characterize the synchronous change relationships among multiple users and is a key basis for identifying potential upstream and downstream logic. This embodiment uses the Pearson correlation coefficient to calculate the consistency of electricity consumption changes among enterprises in order to quantify the synchronous relationship.
[0031] The steps of the enterprise correlation analysis method based on Pearson correlation coefficient are as follows.
[0032] Step 1: Calculate the relevant synchronicity The formula is as follows: (4) in, , indicating enterprises and enterprises Synchronous correlation between them.
[0033] Step 2: To determine the direction and period of load transmission in the industrial chain, a lag correlation coefficient is introduced: (5) in, This is the lag time step. If exist If the value is reached at a certain time, then the enterprise can be determined. As upstream, enterprises Downstream, and Its transmission lag period. When synchronicity and lag When both are significant, it can be determined that there is a clear upstream and downstream linkage and transmission relationship between enterprises.
[0034] 3. Construction of upstream and downstream linkage matrix To more accurately and comprehensively characterize the load transmission relationship and coupling strength between enterprises, an upstream-downstream linkage matrix is constructed. This matrix can systematically describe the load coupling mechanism between enterprises in the industrial chain. (6) in, This refers to the synchronous correlation coefficient; The maximum lag correlation coefficient; Labels indicating upstream and downstream directions confirmed through research; These are the weighting coefficients.
[0035] Linkage Matrix It can serve as a core input feature, providing structured and mechanism-driven load transmission constraint information for subsequent prediction models.
[0036] Then, an ensemble learning model (XGBoost) is used, which takes the real-time and delayed weighted loads of upstream enterprises included in the linkage matrix as core input features, and inputs them together with enterprise features and external features to train a collaborative prediction model.
[0037] like Figure 2 As shown, an ensemble learning model, XGBoost, is used to construct a multi-user collaborative prediction framework. The core of this framework is to integrate the linkage matrix into the input features to achieve collaborative prediction for multiple users.
[0038] The enterprise's input feature vector is innovatively defined as a composite vector containing supply chain transmission terms: (7) in, For the characteristics of the enterprise itself; Real-time weighted features of upstream enterprises; The lagging weighted characteristics of upstream enterprises; Characteristics of external factors; This indicates a maintenance / repair status.
[0039] When a business is under maintenance, the load forecast will be adjusted for hard constraints based on the maintenance plan. (8) in, This is the impact factor of the maintenance on the load. This is a maintenance status variable.
[0040] Prediction function Depend on Tree base learner The additive model consists of: (9) The model uses a loss function with a regularization term. Optimize: (10) Its regular term The formula used to penalize the complexity of the model is as follows: (11) In summary, the implementation process of multi-user collaborative prediction is as follows: Figure 3 As shown.
[0041] After the model outputs the forecast results for each enterprise, a consistency check is performed on the upstream and downstream links. In this embodiment, when an upstream enterprise is predicted to be in a state of shutdown or low load, it is checked whether the predicted load of its downstream enterprise exceeds the reasonable supply chain constraint range. If the check finds that the predicted load of the downstream enterprise exceeds the reasonable supply chain constraint range, an automatic correction mechanism is triggered to adjust the predicted value of the downstream enterprise to a reasonable range allowed by the supply chain logic; and the model parameters and linkage matrix structure are dynamically optimized based on the prediction accuracy index.
[0042] In summary, by making the topology and transmission time delay of the industrial chain explicit and characteristic, the accuracy and robustness of electricity load forecasting in highly interconnected industrial clusters such as chemical and manufacturing industries can be significantly improved.
Claims
1. A multi-user electricity consumption forecasting method considering upstream and downstream linkages, characterized in that, Includes the following steps: S1. Construct a multi-dimensional feature system that includes electricity consumption time-series characteristics, production and maintenance characteristics, external environmental characteristics, and industry tags; S2. For multidimensional feature systems, homogeneous clustering and time-delay cross-correlation analysis are used to identify upstream and downstream relationships and construct a linkage matrix; S3. Based on the linkage matrix, construct an integrated learning model, introduce the lag term and weighted synchronization term of the load of upstream enterprises as features, and perform collaborative prediction of the load of downstream enterprises. S4. Perform consistency verification and correction on the prediction results obtained in step S3 based on the industrial chain link, and output the single enterprise load prediction value and the total industry load prediction value.
2. The multi-user electricity consumption forecasting method considering upstream and downstream linkage as described in claim 1, characterized in that, The electricity consumption time sequence characteristics in S1 include monthly electricity consumption, daily electricity consumption, and electricity consumption at a specific time point. The production and maintenance characteristics include maintenance plans, downtime status, and production continuity indicators.
3. The multi-user electricity consumption forecasting method considering upstream and downstream linkage as described in claim 1, characterized in that, Specifically, S2 employs a deep autoencoder for dimensionality reduction preprocessing, followed by homogeneity clustering analysis based on the K-means algorithm; S2 uses the Pearson correlation coefficient to calculate the consistency of electricity changes among enterprises for time-delay cross-correlation analysis.
4. The multi-user electricity consumption forecasting method considering upstream and downstream linkage as described in claim 3, characterized in that, S2 specifically includes the following steps: S21. A deep autoencoder is used to extract and reduce the dimensions of high-dimensional electricity consumption features, and the corresponding low-dimensional feature vector is output. S22. For low-dimensional feature vectors, the K-means algorithm is used to identify homogeneity and identify groups of companies in the industry with similar electricity consumption behavior patterns. S23. Using the Pearson correlation coefficient, first calculate the synchronous correlation coefficient between enterprises, and then calculate the lagged correlation coefficient to determine the upstream and downstream linkage transmission relationship between enterprises. S24. Based on the synchronous correlation coefficient, the lag correlation coefficient and industry labels, a weighted fusion is performed to construct the upstream and downstream linkage matrix.
5. A multi-user electricity consumption forecasting method considering upstream and downstream linkages as described in claim 4, characterized in that, The formula for calculating the synchronization correlation coefficient in S23 is as follows: in, The synchronous correlation coefficient represents the correlation coefficient between enterprises. and enterprises Synchronous correlation between them; The formula for calculating the lagged correlation coefficient is: in, For the time step of lag, if exist If the value is reached at a certain time, then the enterprise is judged. As upstream, enterprises Downstream, and Its transmission lag period; when synchronicity and lag When both are significant, it indicates that there is a clear upstream and downstream linkage and transmission relationship between enterprises.
6. The multi-user electricity consumption forecasting method considering upstream and downstream linkage as described in claim 5, characterized in that, The upstream and downstream linkage matrix in S24 is specifically as follows: in, This refers to the synchronous correlation coefficient; The maximum lag correlation coefficient; Labels indicating upstream and downstream directions confirmed through research; These are the corresponding weighting coefficients.
7. The multi-user electricity consumption forecasting method considering upstream and downstream linkage as described in claim 1, characterized in that, Specifically, S3 employs the Extreme Gradient Boosting (XGBoost) algorithm to construct an ensemble learning model. The input feature vector of the ensemble learning model contains a composite vector of supply chain transmission terms. The prediction function of the ensemble learning model consists of an additive model of multiple base learners. The ensemble learning model is optimized using a loss function with a regularization term, which is used to penalize the complexity of the model.
8. A multi-user electricity consumption forecasting method considering upstream and downstream linkages as described in claim 7, characterized in that, The composite vector includes the enterprise's own characteristics, the real-time weighted characteristics and the lagged weighted characteristics of upstream enterprises, external factor characteristics, and maintenance status characteristics.
9. A multi-user electricity consumption forecasting method considering upstream and downstream linkages as described in claim 8, characterized in that, When the maintenance status characteristics indicate that the enterprise is in a maintenance state, the prediction results output by the ensemble learning model will be corrected by hard constraints according to the maintenance plan.
10. A multi-user electricity consumption forecasting method considering upstream and downstream linkages as described in claim 1, characterized in that, Specifically, S4 involves checking whether the prediction results of upstream enterprises exceed the preset supply chain constraints when the prediction results of upstream enterprises are in a state of shutdown or low load. If the verification finds inconsistency, an automatic correction mechanism is triggered to adjust the prediction results of downstream enterprises to the set range allowed by the industrial chain logic, and dynamically optimize the parameters of the integrated learning model and the linkage matrix structure based on the prediction accuracy index.