Power distribution area net load prediction method based on global modeling and fusion optimization

Abstract

The invention discloses a power distribution area net load prediction method based on global modeling and fusion optimization, and the method comprises the steps: firstly constructing a global net load prediction model, then carrying out the construction and fine tuning of a local model, and finally carrying out the model integration and fusion prediction. According to the power distribution area net load prediction method based on global modeling and fusion optimization, global modeling, feature extraction and multi-model integration are fused, a global period modeling structure is constructed, area-level local fine tuning is performed on the basis, and a multi-model fusion mechanism is combined, so that the power distribution area net load prediction efficiency is improved. The problems that according to an existing method, modeling is conducted in multiple areas, and the area migration effect is unstable are solved. According to the method, the modeling sharing performance and generalization capability of the model are improved, the depiction capability of periodicity and disturbance characteristics in the transformer area load is enhanced, and high-precision, high-efficiency and high-adaptability net load prediction is realized.

Description

Technical Field

[0001] The present invention relates to the field of power load forecasting, and in particular to a method for forecasting net load in a distribution station area based on global modeling and fusion optimization. Background Art

[0002] With the widespread integration of renewable energy sources like distributed photovoltaics and the rapid development of electric vehicle charging stations, regional loads are exhibiting strong non-stationarity, complex periodicity, and frequent disturbances. Net load forecasting must simultaneously account for user electricity consumption behavior, fluctuations in distributed generation, and the dynamic changes caused by charging and discharging of charging stations. Traditional methods struggle to meet practical needs in terms of forecast accuracy and model adaptability.

[0003] In the existing technologies, there are load forecasting methods based on deep learning, such as those based on convolutional neural networks (CNN) or long short-term memory networks (LSTM). There are also load forecasting methods based on transfer learning, such as cross-regional forecasting methods based on pre-training and local fine-tuning mechanisms. However, these forecasting methods have the following disadvantages:

[0004] (1) Existing deep learning-based load forecasting methods require separate modeling when processing loads in multiple substations, making it difficult to achieve shared modeling capabilities. They also suffer from computational complexity and poor scalability.

[0005] (2) When there are significant structural differences between different stations, the model transfer effect is unstable and the prediction accuracy is limited based on the transfer learning method.

[0006] In recent years, load forecasting methods based on time series modeling and transfer learning have garnered significant attention. This architecture, combining global modeling with local fine-tuning, has become an effective approach for addressing multi-regional collaborative forecasting. Furthermore, periodic modeling methods such as TBATS (Temporal Bayesian Autoregressive with eXogenous inputs) are well-suited for handling complex periodic patterns in substation loads due to their multi-seasonal modeling capabilities. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method for net load forecasting in distribution substations based on global modeling and fusion optimization. By constructing a global periodic modeling structure and performing local fine-tuning at the substation level based on this structure, combined with a multi-model fusion mechanism, this method solves the problems of existing methods such as separate modeling in multiple substations and unstable regional migration effects. This method not only improves the model's modeling sharing and generalization capabilities, but also enhances the ability to characterize the periodic and perturbation characteristics of substation loads, achieving high-precision, high-efficiency, and highly adaptable net load forecasting.

[0008] The technical solution to achieve the above purpose is: a method for predicting net load in distribution area based on global modeling and fusion optimization, comprising the following steps:

[0009] S1, building a global net load forecasting model: Based on TBATS, a global net load forecasting model is built and pre-trained on historical data from multiple substations to extract common features of periodicity and trend;

[0010] S2, local model construction and fine-tuning: Clustering is used to divide the substation area into subsets, and a local model fine-tuning mechanism is introduced. The local model is fine-tuned on each subset to enhance personalized modeling capabilities. To more accurately characterize the impact of distributed generation and charging pile behavior on load, relevant disturbance features are explicitly introduced in the feature construction stage.

[0011] S3, model integration and fusion prediction: Use regression models such as random forest to learn and fuse the prediction results of global and local models, realize adaptive weighting between models, and build an integrated prediction framework with both high precision and strong generalization capabilities.

[0012] In the above-mentioned method for predicting net load of distribution substation area based on global modeling and fusion optimization, step S1 includes the following process:

[0013] S11, data preparation and processing: collect and organize the historical operation data of multiple distribution substations to form an original data set, and construct a unified input feature matrix and its corresponding net load value;

[0014] S12, model structure design: establish a unified prediction model structure to achieve collaborative modeling for multiple substations or buses;

[0015] S13, model training and optimization: After completing the model structure design, the global TBATS model is trained so that it can fully learn the common periodic structure and trend characteristics in the net load sequences of multiple substations.

[0016] The specific process of step S11 of the above-mentioned method for predicting net load of distribution substation area based on global modeling and fusion optimization is as follows:

[0017] First, the historical operation data of multiple distribution substations are collected and sorted to form the original data set D raw , at any time t, the original data is represented as a vector:

[0018] X t ={L t ,G t ,C t ,W t ,T t} (1)

[0019] In formula (1), X t Represents the original data vector at time t, L t Indicates the total power load, G t For distributed power generation, C t is the net power of the charging pile, W t represents the meteorological characteristic vector, T t Indicates a time tag;

[0020] Based on the above raw data, the net load value of the substation at each moment is calculated and defined as follows:

[0021]

[0022] In formula (2), Indicates the time series value of the net load at that moment, L t Indicates the total power load, G t For distributed power generation, C t is the net power of the charging pile;

[0023] In order to enable the model to accurately learn the temporal patterns and disturbance features in the data, the original dataset D raw Perform unified preprocessing to generate a standardized data set D clean All numerical variables are standardized, and time features are converted to one-hot encoding to unify feature formats and ensure comparability of model training.

[0024] In order to enhance the model's ability to model historical dependency information, based on the net load sequence Construct the hysteresis feature. Specifically, a sliding window method is used to introduce the net load value of the previous K moments as one of the inputs of the current moment to represent the time dependency.

[0025] Finally, at each time t, the input feature matrix is constructed Its structure is as follows:

[0026]

[0027] In formula (3), Represents the net load hysteresis value at K historical moments, W t is the current meteorological characteristic vector, T t (enc) is the encoded time feature; the input matrix As the prediction input of the model at time t, it is used to predict the net load value at future time

[0028] After the above preprocessing process, a standardized and uniformly structured input sample pair is obtained. The dataset D required for model training clean , providing a high-quality input basis for the subsequent modeling stage.

[0029] In the above-mentioned method for predicting net load of distribution substation area based on global modeling and fusion optimization, the specific process of step S12 is as follows:

[0030] The unified input feature matrix constructed in step S11 and its corresponding net load value On this basis, a unified prediction model structure is established to achieve collaborative modeling for multiple substations or buses;

[0031] To uniformly model the net load variation patterns of multiple substations, a global forecasting model structure is proposed. This global forecasting model is jointly trained on historical data from all substations, with parameters shared between substations. It aims to extract common periodic structures and trend patterns across substations. The global forecasting model uses the TBATS model as its modeling framework, which is suitable for processing power load sequences with multi-periodic and non-stationary characteristics.

[0032] The global prediction model takes the input feature matrix constructed in step S11 as input and fits the target net load value by combining multiple time series components Its structure is as follows:

[0033]

[0034] In formula (4), B λ (·) is the Box-Cox transformation, Z t is the trend term, S t is the seasonal periodic term E t is the residual term; the global prediction model adaptively identifies multiple periodic structures and jointly models them to adapt to the characteristics of complex periodic disturbances;

[0035] This global prediction model is jointly trained on data from multiple substations and has good generalization capabilities. The parameters of the global prediction model can be migrated to each substation, and on this basis, local models can be further fine-tuned to achieve personalized prediction optimization.

[0036] In the above-mentioned method for predicting net load of distribution substation area based on global modeling and fusion optimization, the specific process of step S13 is as follows:

[0037] After completing the design of the global prediction model structure, the global prediction model is trained to enable it to fully learn the common periodic structure and trend characteristics of the net load series of multiple substations;

[0038] Model training is based on the standardized dataset constructed in step S11 An end-to-end approach is used to optimize the model parameters. To improve the prediction accuracy and generalization ability of the model, a joint loss function consisting of an error term and a regularization term is introduced during the training process, which is defined as follows:

[0039]

[0040] In formula (5), is the actual net load value, is the model prediction value, θ is the model parameter, and λ is the regularization coefficient; the first term is the mean absolute error, which is used to measure the model prediction accuracy; the second term is L1 regularization, which is used to control the model complexity and improve the generalization ability of the model on new areas;

[0041] During the training process, the global prediction model is jointly trained on data from all substations, and parameters are shared between substations, thus forming a global prediction model with universal modeling capabilities; after training is completed, the global prediction model can be used as the initialization basis for local modeling, and further parameter fine-tuning can be performed on a single substation to achieve personalized optimization.

[0042] The specific process of step S2 of the above-mentioned method for predicting net load of distribution substation area based on global modeling and fusion optimization is as follows:

[0043] S21, Substation Clustering and Stratification: After the global prediction model is trained, a local model fine-tuning mechanism is introduced to further improve the prediction accuracy and personalized adaptability of the model for each substation or busbar. The core idea of the local model is to fine-tune the historical data of each substation based on the structure and parameters of the global prediction model to adapt to its unique power consumption characteristics and disturbance patterns.

[0044] S22, local model initialization and fine-tuning: On the basis of completing the global prediction model training, in order to achieve personalized fitting of the substation load characteristics, the transfer learning strategy is used to fine-tune the local model.

[0045] In the above-mentioned method for predicting net load of distribution substations based on global modeling and fusion optimization, in step S21, to enhance the pertinence of local model structure selection and initialization, cluster analysis is first performed on all substations. The clustering process extracts statistical features based on the historical net load sequence of each substation to construct a substation feature space, and K-means clustering is performed at multiple clustering levels.

[0046] Net load sequence for each substation Extract statistical feature vector F i , including mean, trend slope and periodicity indicators, to characterize the typical behavior pattern of the net load in the substation area;

[0047] The characteristic set of all regions {F i} is used as input, and the K-means algorithm is used to construct clustering results at different clustering levels C = {2, 3, ..., 10}. The clustering structure at each level c is expressed as:

[0048]

[0049] In formula (6), Represents the kth subset of the cth level cluster; each station i in each cluster level c corresponds to a unique cluster label Ultimately, each station has a set of hierarchical clustering labels at different levels, which are used for subsequent local model initialization strategies, integrated weight calculations, or structural optimization operations.

[0050] The above-mentioned method for predicting net load of distribution substation area based on global modeling and fusion optimization has the following specific process in step S22:

[0051] First, for each station area i, the parameter θ of the global prediction model is global As the initial parameters of its local model

[0052] Subsequently, the historical data set of the area The local model is fine-tuned and trained. The training objectives remain consistent with the global stage. The joint loss function consisting of mean absolute error and L1 regularization term is still used. However, the training process is performed independently in each station area and the parameters are no longer shared.

[0053] The training objective is expressed as:

[0054]

[0055] In formula (7), represents the optimal local model parameters obtained after fine-tuning of the ith station, is the actual net load value of the station i, is the model prediction value of station area i, θ i is the model parameter of station i, and λ is the regularization coefficient.

[0056] The above-mentioned method for predicting the net load of distribution substations based on global modeling and fusion optimization, the fine-tuning strategy in step S22 takes into account the modeling capabilities of the global prediction model and the individual characteristics of local data, significantly improves the prediction accuracy of the model in different substations, and provides a diversified and high-quality candidate model basis for subsequent model integration.

[0057] In the above-mentioned method for predicting net load of distribution substation area based on global modeling and fusion optimization, step S3 specifically includes the following process:

[0058] S31, summarizing the prediction results of multiple models: for each station area i, the prediction results of the global prediction model and the local model are summarized to form the input of the integrated model;

[0059] For any time t, construct the prediction result vector:

[0060]

[0061] In formula (8), represents the prediction result of the global prediction model on station i, Indicates the k The prediction results of the local models for the substation i, K represents the number of local models involved in the substation, including local models at different clustering levels;

[0062] S32, random forest-based weight learning: To achieve optimal fusion of prediction results from different models, a random forest regression model is used to learn the nonlinear relationship between the prediction values of each model and the true net load value, thereby generating adaptive fusion weights. The training sample format is:

[0063]

[0064] In formula (9), is the prediction vector formed by each model, is the corresponding true net load value; based on this data, the fusion function is trained

[0065]

[0066] In formula (10), represents the random forest model trained on station i, which can adaptively allocate fusion weights according to the performance of different models at different times; is the prediction vector formed by each model;

[0067] S33, weighted fusion output final prediction: input the prediction results of each model in step S31 into the fusion function model trained in step S32 to generate the final prediction result Expressed as:

[0068]

[0069] In formula (11), represents the random forest model trained on station i, which can adaptively allocate fusion weights according to the performance of different models at different times; is the prediction vector formed by each model.

[0070] The present invention's distribution network net load forecasting method, based on global modeling and fusion optimization, addresses existing issues such as separate modeling for multiple substations and unstable regional migration by constructing a global periodic modeling structure and performing local fine-tuning at the substation level. This method not only improves the model's modeling sharing and generalization capabilities, but also enhances its ability to capture the periodic and perturbation characteristics of substation loads, achieving highly accurate, efficient, and adaptable net load forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 This is a flow chart of the net load prediction method for distribution station areas based on global modeling and fusion optimization of the present invention. DETAILED DESCRIPTION

[0072] In order to enable those skilled in the art to better understand the technical solution of the present invention, the specific implementation methods thereof are described in detail below with reference to the accompanying drawings:

[0073] See also Figure 1 , the best embodiment of the present invention, a method for predicting net load of distribution station area based on global modeling and fusion optimization, includes the following steps.

[0074] S1, building a global net load forecasting model: Build a global net load forecasting model based on TBATS and pre-train it on historical data from multiple substations to extract common features of periodicity and trend. The specific process includes the following:

[0075] S11, Data Preparation and Processing

[0076] First, the historical operation data of multiple distribution substations are collected and sorted to form the original data set D raw At any time t, the original data can be represented as a vector:

[0077] X t ={L t ,G t ,C t ,W t ,T t} (1)

[0078] In formula (1), X t Represents the original data vector at time t, L t Indicates the total power load, G t For distributed power generation, C t is the net power of the charging pile (charging is positive, discharging is negative), W t represents the meteorological characteristic vector (such as temperature, humidity, light, etc.), T t Indicates time labels (such as hours, days of the week, holidays, etc.).

[0079] Based on the above raw data, the net load value of the substation at each moment is calculated and defined as follows:

[0080]

[0081] In formula (2), Indicates the time series value of the net load at that moment.

[0082] In order to enable the model to accurately learn the temporal patterns and disturbance features in the data, the original dataset D raw Perform unified preprocessing to generate a standardized data set D clean All numerical variables are standardized, and time features are converted to one-hot encoding to unify the feature format and ensure the comparability of model training.

[0083] In addition, in order to enhance the model's ability to model historical dependency information, the net load sequence The hysteresis feature is constructed. Specifically, a sliding window method is adopted to introduce the net load value of the previous K moments as one of the inputs of the current moment to represent the time dependency.

[0084] Finally, at each time t, the input feature matrix is constructed Its structure is as follows:

[0085]

[0086] In formula (3), Represents the net load hysteresis value at K historical moments, W t is the current meteorological characteristic vector, T t (enc) is the encoded time feature. The input matrix As the prediction input of the model at time t, it is used to predict the net load value at future time

[0087] After the above preprocessing process, a standardized and uniformly structured input sample pair is obtained. The dataset D required for model training clean , providing a high-quality input basis for the subsequent modeling stage.

[0088] S12, model structure design

[0089] In step S11, a unified input feature matrix is constructed and its corresponding net load value On this basis, this step establishes a unified prediction model structure to achieve collaborative modeling for multiple substations (or buses).

[0090] To uniformly model the net load variations across multiple substations, a global forecasting model structure is proposed. This model is jointly trained on historical data from all substations, with parameters shared across substations. It aims to extract common periodic structures and trend patterns across substations. The TBATS model was selected as the modeling framework, which is suitable for handling power load series with multi-periodic and non-stationary characteristics.

[0091] The global prediction model takes the input feature matrix constructed in step S11 as input and fits the target net load value by combining multiple time series components Its structure is as follows:

[0092]

[0093] In formula (4), B λ (·) is the Box-Cox transformation, Z t is the trend term, S t is the seasonal periodic term E t The model can adaptively identify multiple periodic structures (such as daily and weekly cycles) and jointly model them to adapt to complex periodic disturbance characteristics.

[0094] This global prediction model is jointly trained on data from multiple substations and has good generalization capabilities. Its parameters can be transferred to each substation, and further fine-tuned to build local models to achieve personalized prediction optimization.

[0095] S13, model training and optimization

[0096] After completing the design of the global prediction model structure, the global prediction model needs to be trained so that it can fully learn the common periodic structure and trend characteristics in the net load sequences of multiple substations.

[0097] Model training is based on the standardized dataset constructed in step S11 The model parameters are optimized in an end-to-end manner. To improve the prediction accuracy and generalization ability of the model, a joint loss function consisting of an error term and a regularization term is introduced during the training process, which is defined as follows:

[0098]

[0099] In formula (5), is the actual net load value, is the model prediction value, θ is the model parameter, and λ is the regularization coefficient. The first term is the mean absolute error (MAE), which is used to measure the model's prediction accuracy; the second term is L1 regularization, which is used to control model complexity and improve the model's generalization ability on new platforms.

[0100] During training, the model is trained jointly on data from all substations, and parameters are shared across substations, forming a global prediction model with universal modeling capabilities. After training, this model can be used as the initialization basis for local modeling, and further parameter fine-tuning can be performed on individual substations to achieve personalized optimization.

[0101] S2, local model construction and fine-tuning: The substation area is divided into subsets through clustering methods, and a local model fine-tuning mechanism is introduced. The local model is fine-tuned on each subset to enhance personalized modeling capabilities. To more accurately characterize the impact of distributed generation and charging pile behavior on the load, relevant disturbance features are explicitly introduced in the feature construction stage. The specific process is as follows:

[0102] S21, Clustering and Stratification of Substation Areas

[0103] After the global prediction model is trained, a local model fine-tuning mechanism is introduced to further improve the model's prediction accuracy and personalized adaptability at each substation (or busbar). The core idea of the local model is to fine-tune the historical data of each substation based on the structure and parameters of the global prediction model to adapt to its unique power consumption characteristics and disturbance patterns.

[0104] To enhance the specificity of local model structure selection and initialization, this step first performs a cluster analysis on all substations. The clustering process extracts statistical features based on each substation's historical net load series to construct a substation feature space, and then performs K-means clustering at multiple clustering levels.

[0105] Net load sequence for each substation Extract statistical feature vector F i , including common mean, trend slope, periodic indicators, etc., to characterize the typical behavior pattern of the net load in the substation.

[0106] The characteristic set of all regions {F i} is input, and the K-means algorithm is used to construct clustering results at different clustering levels C = {2, 3, ..., 10}. The clustering structure at each level c is expressed as:

[0107]

[0108] In formula (6), Represents the kth sub-area subset in the cth level cluster. Each sub-area i in each cluster level c corresponds to a unique cluster label Ultimately, each station has a set of hierarchical clustering labels at different levels, which can be used for subsequent local model initialization strategies, integrated weight calculations, or structural optimization operations.

[0109] S22, local model initialization and fine-tuning

[0110] On the basis of completing the global prediction model training, in order to achieve personalized fitting of the load characteristics of the substation, this step adopts the transfer learning strategy to fine-tune the local model. Specifically, for each substation i, the parameters θ of the global prediction model are global As the initial parameters of its local model

[0111] Subsequently, the historical data set of the area The local model is fine-tuned and trained. The training objective remains the same as the global stage, and the joint loss function consisting of mean absolute error (MAE) and L1 regularization is still used. However, the training process is performed independently in each station, and the parameters are no longer shared.

[0112] The training objective can be expressed as:

[0113]

[0114] In formula (7), represents the optimal local model parameters obtained after fine-tuning of the i-th station; is the actual net load value of the station i, is the model prediction value of station area i, θ i is the model parameter of station i, and λ is the regularization coefficient.

[0115] This fine-tuning strategy takes into account both the modeling capabilities of the global prediction model and the individual characteristics of local data. It can significantly improve the prediction accuracy of the model in different substations and provide a diverse and high-quality candidate model foundation for subsequent model integration.

[0116] S3, model integration and fusion prediction: Use regression models such as random forest to learn and fuse the prediction results of global and local models, implement adaptive weighting between models, and build an integrated prediction framework with both high accuracy and strong generalization capabilities. The specific process includes the following:

[0117] S31, summarizing the prediction results of multiple models: for each station area i, the prediction results of the global prediction model and the local model are summarized to form the input of the integrated model;

[0118] For any time t, construct the prediction result vector:

[0119]

[0120] In formula (8), represents the prediction result of the global prediction model on station i, represents the prediction result of the kth local model for the station area i, K represents the number of local models involved in the station area, including local models at different clustering levels;

[0121] S32, random forest-based weight learning: To achieve optimal fusion of prediction results from different models, a random forest regression model is used to learn the nonlinear relationship between the prediction values of each model and the true net load value, thereby generating adaptive fusion weights. The training sample format is:

[0122]

[0123] In formula (9), is the prediction vector formed by each model, is the corresponding true net load value; based on this data, the fusion function is trained

[0124]

[0125] In formula (10), represents the random forest model trained on station i, which can adaptively allocate fusion weights according to the performance of different models at different times; is the prediction vector formed by each model;

[0126] S33, weighted fusion output final prediction: input the prediction results of each model in step S31 into the fusion function model trained in step S32 to generate the final prediction result Expressed as:

[0127]

[0128] In formula (11), represents the random forest model trained on station i, which can adaptively allocate fusion weights according to the performance of different models at different times; is the prediction vector formed by each model.

[0129] In summary, the present invention's distribution network net load forecasting method, based on global modeling and fusion optimization, addresses the problems of existing methods, such as separate modeling for multiple substations and unstable regional migration, by constructing a global periodic modeling structure and performing local fine-tuning at the substation level based on this structure. This method not only improves the model's modeling sharing and generalization capabilities, but also enhances the ability to characterize the periodic and perturbation characteristics of substation loads, achieving highly accurate, efficient, and adaptable net load forecasting.

[0130] Those skilled in the art should recognize that the above embodiments are merely intended to illustrate the present invention and are not intended to limit the present invention. As long as they are within the spirit of the present invention, any changes or modifications to the above embodiments will fall within the scope of the claims of the present invention.

Claims

1. A method for predicting net load in distribution area based on global modeling and fusion optimization, characterized in that: The following steps are involved: S1, building a global net load forecasting model: Based on TBATS, a global net load forecasting model is built and pre-trained on historical data from multiple substations to extract common features of periodicity and trend; S2, local model construction and fine-tuning: Clustering is used to divide the substation area into subsets, and a local model fine-tuning mechanism is introduced. The local model is fine-tuned on each subset to enhance personalized modeling capabilities. To more accurately characterize the impact of distributed generation and charging pile behavior on load, relevant disturbance features are explicitly introduced in the feature construction stage. S3, model integration and fusion prediction: Use regression models such as random forest to learn and fuse the prediction results of global and local models, realize adaptive weighting between models, and build an integrated prediction framework with both high precision and strong generalization capabilities.

2. A method for predicting net load in a distribution area based on global modeling and fusion optimization according to claim 1, characterized in that: Step S1 includes the following process: S11, data preparation and processing: collect and organize the historical operation data of multiple distribution substations to form an original data set, and construct a unified input feature matrix and its corresponding net load value; S12, model structure design: establish a unified prediction model structure to achieve collaborative modeling for multiple substations or buses; S13, model training and optimization: After completing the model structure design, the global TBATS model is trained so that it can fully learn the common periodic structure and trend characteristics in the net load sequences of multiple substations.

3. The method for predicting net load of distribution network area based on global modeling and fusion optimization according to claim 2 is characterized in that: The specific process of step S11 is: First, the historical operation data of multiple distribution substations are collected and sorted to form the original data set D raw , at any time t, the original data is represented as a vector: X t ={L t ,G t ,C t ,W t ,T t } (1) In formula (1), X t Represents the original data vector at time t, L t Indicates the total power load, G t For distributed power generation, C t is the net power of the charging pile, W t represents the meteorological characteristic vector, T t Indicates a time tag; Based on the above raw data, the net load value of the substation at each moment is calculated and defined as follows: In formula (2), Indicates the time series value of the net load at that moment, L t Indicates the total power load, G t For distributed power generation, C t is the net power of the charging pile; In order to enable the model to accurately learn the temporal patterns and disturbance features in the data, the original dataset D raw Perform unified preprocessing to generate a standardized data set D clean ; All numerical variables are standardized, and time features are converted to one-hot encoding to unify feature formats and ensure comparability of model training; In order to enhance the model's ability to model historical dependency information, based on the net load sequence Construct the hysteresis feature. Specifically, a sliding window method is used to introduce the net load value of the previous K moments as one of the inputs of the current moment to represent the time dependency. Finally, at each time t, the input feature matrix is constructed Its structure is as follows: In formula (3), Represents the net load hysteresis value at K historical moments, W t is the current meteorological characteristic vector, T t (enc) is the encoded time feature; the input matrix As the prediction input of the model at time t, it is used to predict the net load value at future time After the above preprocessing process, a standardized and uniformly structured input sample pair is obtained. The dataset D required for model training clean , providing a high-quality input basis for the subsequent modeling stage.

4. The method for predicting net load of distribution network area based on global modeling and fusion optimization according to claim 3 is characterized in that: The specific process of step S12 is: The unified input feature matrix constructed in step S11 and its corresponding net load value On this basis, a unified prediction model structure is established to achieve collaborative modeling for multiple substations or buses; To uniformly model the net load variation patterns of multiple substations, a global forecasting model structure is proposed. This global forecasting model is jointly trained on historical data from all substations, with parameters shared between substations. It aims to extract common periodic structures and trend patterns across substations. The global forecasting model uses the TBATS model as its modeling framework, which is suitable for processing power load sequences with multi-periodic and non-stationary characteristics. The global prediction model takes the input feature matrix constructed in step S11 as input and fits the target net load value by combining multiple time series components Its structure is as follows: In formula (4), B λ (·) is the Box-Cox transformation, Z t is the trend term, S t is the seasonal periodic term E t is the residual term; the global prediction model adaptively identifies multiple periodic structures and jointly models them to adapt to the characteristics of complex periodic disturbances; This global prediction model is jointly trained on data from multiple substations and has good generalization capabilities. The parameters of the global prediction model can be migrated to each substation, and on this basis, local models can be further fine-tuned to achieve personalized prediction optimization.

5. The method for predicting net load of distribution network area based on global modeling and fusion optimization according to claim 4 is characterized in that: The specific process of step S13 is: After completing the design of the global prediction model structure, the global prediction model is trained to enable it to fully learn the common periodic structure and trend characteristics of the net load series of multiple substations; Model training is based on the standardized dataset constructed in step S11 An end-to-end approach is used to optimize the model parameters. To improve the prediction accuracy and generalization ability of the model, a joint loss function consisting of an error term and a regularization term is introduced during the training process, which is defined as follows: In formula (5), is the actual net load value, is the model prediction value, θ is the model parameter, and λ is the regularization coefficient; the first term is the mean absolute error, which is used to measure the model prediction accuracy; the second term is L1 regularization, which is used to control the model complexity and improve the generalization ability of the model on new areas; During the training process, the global prediction model is jointly trained on data from all substations, and parameters are shared between substations, thus forming a global prediction model with universal modeling capabilities; after training is completed, the global prediction model can be used as the initialization basis for local modeling, and further parameter fine-tuning can be performed on a single substation to achieve personalized optimization.

6. The method for predicting net load of distribution network area based on global modeling and fusion optimization according to claim 1 is characterized in that: The specific process of step S2 is: S21, Substation Clustering and Stratification: After the global prediction model is trained, a local model fine-tuning mechanism is introduced to further improve the prediction accuracy and personalized adaptability of the model for each substation or busbar. The core idea of the local model is to fine-tune the historical data of each substation based on the structure and parameters of the global prediction model to adapt to its unique power consumption characteristics and disturbance patterns. S22, local model initialization and fine-tuning: On the basis of completing the global prediction model training, in order to achieve personalized fitting of the substation load characteristics, the transfer learning strategy is used to fine-tune the local model.

7. The method for predicting net load of distribution substation area based on global modeling and fusion optimization according to claim 6 is characterized in that: In step S21, to enhance the pertinence of local model structure selection and initialization, cluster analysis is first performed on all substations. The clustering process extracts statistical features based on the historical net load sequence of each substation to construct a substation feature space, and K-means clustering is performed at multiple clustering levels. Net load sequence for each substation Extract statistical feature vector F i , including mean, trend slope and periodicity indicators, to characterize the typical behavior pattern of the net load in the substation area; The characteristic set of all regions {F i } is used as input, and the K-means algorithm is used to construct clustering results at different clustering levels C = {2, 3, ..., 10}. The clustering structure at each level c is expressed as: In formula (6), Represents the kth subset of the cth level cluster; each station i in each cluster level c corresponds to a unique cluster label Ultimately, each station has a set of hierarchical clustering labels at different levels, which are used for subsequent local model initialization strategies, integrated weight calculations, or structural optimization operations.

8. The method for predicting net load of distribution network area based on global modeling and fusion optimization according to claim 6 is characterized in that: The specific process of step S22 is: First, for each station area i, the parameter θ of the global prediction model is global As the initial parameters of its local model Subsequently, the historical data set of the area The local model is fine-tuned and trained. The training objectives remain consistent with the global stage. The joint loss function consisting of mean absolute error and L1 regularization term is still used. However, the training process is performed independently in each station area and the parameters are no longer shared. The training objective is expressed as: In formula (7), represents the optimal local model parameters obtained after fine-tuning of the ith station, is the actual net load value of the station i, is the model prediction value of station area i, θ i is the model parameter of station i, and λ is the regularization coefficient.

9. The method for predicting net load of distribution network area based on global modeling and fusion optimization according to claim 8 is characterized in that: The fine-tuning strategy in step S22 takes into account both the modeling capability of the global prediction model and the individual characteristics of local data, significantly improving the prediction accuracy of the model in different areas and providing a diverse and high-quality candidate model foundation for subsequent model integration.

10. The method for predicting net load of distribution network area based on global modeling and fusion optimization according to claim 1, characterized in that: Step S3 specifically includes the following process: S31, summarizing the prediction results of multiple models: for each station area i, the prediction results of the global prediction model and the local model are summarized to form the input of the integrated model; For any time t, construct the prediction result vector: In formula (8), represents the prediction result of the global prediction model on station i, Indicates the k The prediction results of the local models for the substation i, K represents the number of local models involved in the substation, including local models at different clustering levels; S32, random forest-based weight learning: To achieve optimal fusion of prediction results from different models, a random forest regression model is used to learn the nonlinear relationship between the prediction values of each model and the true net load value, thereby generating adaptive fusion weights. The training sample format is: In formula (9), is the prediction vector formed by each model, is the corresponding true net load value; based on this data, the fusion function is trained In formula (10), represents the random forest model trained on station i, which can adaptively allocate fusion weights according to the performance of different models at different times; is the prediction vector formed by each model; S33, weighted fusion output final prediction: input the prediction results of each model in step S31 into the fusion function model trained in step S32 to generate the final prediction result Expressed as: In formula (11), represents the random forest model trained on station i, which can adaptively allocate fusion weights according to the performance of different models at different times; is the prediction vector formed by each model.

Images