Multi-dimensional attribute power consumer time-sharing electric quantity clustering analysis and prediction method and system

Abstract

The invention discloses a multi-dimensional attribute power consumer time-sharing electric quantity clustering analysis and prediction method and system. The method comprises the following steps: extracting multi-dimensional prediction input features including time sequence features, market features and business features from collected data; calling a typical power consumption mode clustering template, calculating the similarity between a to-be-predicted user and each clustering template based on the multi-dimensional prediction input features, and determining a target cluster and a corresponding typical power curve; performing prediction through the multi-branch collaborative prediction model to obtain an initial 24-hour time-sharing electric quantity weight of the user to be predicted; dynamically adjusting the initial weight in combination with the price signal of the electric power spot market and the price sensitivity level of the user to be predicted, and synchronously executing industry business rule verification and correction; and outputting the optimized time-sharing electric quantity weight and the hierarchical interpretation report. According to the method, the model prediction and the typical power consumption mode are fused through the clustering enhancement strategy, the prediction precision and stability are effectively improved, and technical support is provided for power selling companies to optimize spot market declaration strategies.

Description

Technical Field

[0001] This invention belongs to the field of power management technology, specifically relating to a method and system for clustering and predicting the time-of-use electricity consumption of multi-dimensional attribute power users. Background Technology

[0002] With the continuous deepening of my country's power market reform, the electricity spot market, as a core platform for optimal resource allocation, has been gradually established and improved. This has broken the traditional centralized power purchase and sales model, forming a market-oriented pattern where generation side bids for grid connection and users make independent choices. In the electricity spot market environment, user electricity consumption behavior is influenced by multiple factors, resulting in significant complexity and diversity in time-of-use electricity curves. From an external environmental perspective, changes in meteorological conditions such as temperature and humidity directly affect the electricity demand for loads such as air conditioning and heating. From a market perspective, real-time price fluctuations in the electricity spot market guide users to adjust their electricity consumption periods, especially commercial users and those with adjustable industrial loads, whose electricity consumption behavior is highly sensitive to price signals. From the perspective of user attributes, different industries, such as continuously producing chemical industries, intermittently operating commercial retail, and flexible residential users, as well as different geographical locations and different distributed energy capacity configurations (such as the installation or absence of photovoltaics and energy storage), all lead to significant differences in electricity consumption patterns. Furthermore, business factors such as adjustments to user production plans, equipment maintenance, and holiday arrangements also have a significant impact on time-of-use electricity curves, further increasing the difficulty of prediction.

[0003] However, existing power load forecasting methods are insufficient to meet the high-precision reporting requirements of the electricity spot market, mainly due to the following two core shortcomings:

[0004] On the one hand, the feature dimensions are too limited, failing to comprehensively cover the key factors influencing electricity consumption behavior. Traditional load forecasting methods mostly rely on historical electricity consumption data and basic meteorological information as core inputs, focusing only on the temporal correlation and meteorological sensitivity of electricity consumption behavior, while neglecting the essential impact of multi-dimensional attributes such as user industry characteristics, geographical location, and distributed energy capacity on electricity consumption patterns. At the same time, existing methods generally do not incorporate key information such as electricity spot market price signals and user business plans, making it difficult for forecast results to adapt to adjustments in user electricity consumption behavior under a market-oriented environment, and thus unable to support the market-based decision-making of electricity sales companies.

[0005] On the other hand, there is insufficient handling of user heterogeneity, lacking effective classification and pattern recognition mechanisms. Electricity consumption patterns vary significantly across different industries, cities, and installed capacity types: industrial users in continuous production have stable consumption curves with high nighttime loads; commercial users exhibit a single-peak characteristic with daytime peaks and nighttime troughs, significantly affected by holidays; residential users show a double-peak pattern with morning and evening peaks. Most existing methods use a uniform prediction model to batch predict for all users, failing to classify the electricity consumption characteristics of different user groups, resulting in models that cannot accurately adapt to the electricity consumption patterns of various user groups. While some methods attempt to use simple clustering algorithms for user classification, they suffer from fixed clustering parameters and fail to consider temporal continuity and business constraints, making it difficult to form typical electricity consumption patterns with practical guidance. Ultimately, this leads to low prediction accuracy, failing to meet the accuracy requirements submitted by electricity sales companies.

[0006] In summary, against the backdrop of deepening power market reform, the shortcomings of existing power load forecasting methods in terms of feature coverage and handling of user heterogeneity are becoming increasingly prominent. There is an urgent need for a time-of-use power forecasting technology that can integrate multi-dimensional attribute features, accurately identify typical power consumption patterns, and adapt to market-oriented business needs, so as to provide power sales companies with high-precision and robust time-of-use power forecasting results and help them enhance their core competitiveness in the power spot market. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method and system for multi-dimensional attribute time-of-use electricity consumption clustering analysis and prediction, which provides electricity sales companies with high-precision and robust time-of-use electricity consumption prediction results, helping them enhance their core competitiveness in the electricity spot market.

[0008] In a first aspect, this invention proposes a multi-dimensional attribute time-of-use electricity consumption clustering analysis and prediction method for electricity users, including: Collect basic data, electricity spot market data and business data of users to be predicted, and extract multi-dimensional prediction input features containing time series features, market features and business features; Call the pre-trained typical electricity consumption pattern cluster templates, calculate the similarity between the user to be predicted and each cluster template based on the multi-dimensional prediction input features, and determine the target cluster and the corresponding typical electricity consumption curve; The multidimensional prediction input features and target clustering information are input into the multi-branch collaborative prediction model to make predictions and obtain the initial 24-hour time-of-use electricity weights of the users to be predicted. By combining electricity spot market price signals with the price sensitivity levels of users to be predicted, the initial 24-hour time-of-use electricity weights are dynamically adjusted, and industry business rules are simultaneously verified and corrected. The optimized time-of-use electricity weights are output at multiple time scales, including day-ahead, intraday, and real-time, and a hierarchical explanatory report containing feature contributions and business compliance is generated simultaneously.

[0009] This invention effectively improves the accuracy and robustness of time-of-use electricity prediction through multi-dimensional feature fusion, precise clustering and classification, multi-branch collaborative prediction, and business adaptation optimization. It fully adapts to the heterogeneous electricity consumption of different industries and types of users, providing reliable technical support for electricity sales companies to optimize their reporting strategies and enhance their market competitiveness. It has significant industrial applicability and market application value.

[0010] Preferably, the typical electricity consumption pattern clustering template includes: The training data collected from multiple sources over the past three years were divided into clustered training sets and validation sets according to a certain ratio. The training set is trained using triple clustering; the triple clustering includes a first-level spatiotemporal clustering, a second-level density clustering, and a third-level business clustering. The median sequence of each cluster is used as the template for the typical charge curve of that cluster; Among them, the application status of each template in the real-time monitoring template database is triggered to update immediately when any of the following conditions are met: the actual user sample ratio corresponding to any template exceeds 15% of the change in user samples when the template is generated; or in the prediction application of any template, the actual electricity consumption data for 15 consecutive days exceeds 12% of the template's MAPE.

[0011] Preferably, the typical electricity consumption pattern clustering template further includes: The first spatiotemporal clustering uses a spatiotemporal density clustering algorithm, taking the feature sequence of a single user for 7 consecutive days as one spatiotemporal sample. The key parameters of ST-DBSCAN are determined through the K-distance graph, and all spatiotemporal samples in the clustering training set are clustered to output a time-continuous electricity consumption pattern cluster. The second density clustering uses the DBSCAN algorithm. Taking the pattern clusters generated by the first clustering as units, for each sample in each cluster, DBSCAN secondary clustering is performed on each time-series pattern cluster according to the dual criteria of electricity consumption pattern similarity and price sensitivity coefficient similarity, and the subdivided pattern clusters are output. In addition to the third level of business clustering, for each of the subdivided pattern clusters output by the second level, we verify whether the samples within the cluster conform to the business constraints of the corresponding industry and remove cross-industry mis-clustered samples that do not conform to the constraints. Finally, based on the business constraint verification results, sub-clusters with a similarity of ≥0.9 are merged, and 10-15 typical electricity consumption pattern clusters are finally determined.

[0012] Preferably, the multi-branch collaborative prediction model includes: A Transformer encoder with time position encoding is used as the time series branch. The input consists of a 24-hour electricity weight sequence of the user to be predicted over the past 7 days and an hourly electricity spot price sequence over the past 7 days. The time series feature vector is calculated. The LightGBM model is used as the attribute branch. The user's basic attribute features, market attribute features and target clustering information are input to calculate the attribute feature vector. A 3-layer MLP network is used as the business constraint branch. The industry production standard features and business feature vectors are input to obtain the business feature vector and the business compliance pre-score. The three branch outputs are weighted and fused using an attention mechanism to generate a fused feature vector that is then connected to the fully connected layer. Each sub-layer outputs an hourly power weight, resulting in a 24-dimensional time-sharing power weight prediction result.

[0013] Preferably, the step of dynamically adjusting the initial 24-hour time-of-use electricity weights by combining the electricity spot market price signal with the price sensitivity level of the user to be predicted, and simultaneously performing industry business rule verification and correction, includes: The K-means clustering algorithm is used to cluster the day-ahead price curve for the forecast date into time periods; the clustering results include price peak periods, price average periods, and price trough periods; price volatility is calculated separately for each type of time period. The initial 24-hour time-of-use electricity weights are dynamically adjusted and normalized based on user sensitivity levels and price volatility. The normalized weights are then iterated hourly according to industry business rules, and any violations are corrected according to the principle of minimum adjustment until there are no violations, at which point the final 24-hour time-of-use electricity weights are output.

[0014] Preferably, the step of dynamically adjusting the initial 24-hour time-of-use electricity weights based on user sensitivity levels and price volatility includes: For highly sensitive users, adjustments are made based on price peaks, averages, and troughs. The initial 24-hour time-of-use electricity weights are dynamically adjusted using a multiplication factor, taking into account price volatility and sensitivity coefficients. For users with moderate sensitivity, the weighting is adjusted only for price peak and trough periods, while the initial weighting remains for the flat period. In addition, adjustments will be made for low-sensitivity users, with the weight of non-core loads being reduced only when the volatility exceeds 0.2 during peak price periods.

[0015] Secondly, the present invention also provides a multi-dimensional attribute time-of-use electricity consumption clustering analysis and prediction system for electricity users, the prediction system comprising: The data acquisition unit collects basic data, electricity spot market data, and business data from users to be predicted. The data preprocessing unit extracts multidimensional predictive input features, including time-series features, market features, and business features, from the collected data. The first processing unit calls the pre-trained typical electricity consumption pattern clustering templates, calculates the similarity between the user to be predicted and each clustering template based on the multi-dimensional prediction input features, and determines the target cluster and the corresponding typical electricity consumption curve. The second processing unit inputs the multi-dimensional prediction input features and target clustering information into the multi-branch collaborative prediction model to make predictions and obtain the initial 24-hour time-of-use electricity weights of the users to be predicted. The third processing unit dynamically adjusts the initial 24-hour time-of-use electricity weights by combining the electricity spot market price signals with the price sensitivity levels of the users to be predicted, and simultaneously executes industry business rule verification and correction. The output unit outputs optimized time-of-use electricity weights at multiple time scales, including day-ahead, intraday, and real-time, and simultaneously generates a hierarchical explanation report containing feature contributions and business compliance.

[0016] Preferably, the first processing unit further includes: a typical electricity consumption pattern clustering template, the typical electricity consumption pattern clustering template including: Dataset module: Collects multi-source training data from the past 3 years and divides it into clustering training set and validation set according to proportion; The clustering module performs triple clustering training on the training set; the triple clustering includes a first-level spatiotemporal clustering, a second-level density clustering, and a third-level business clustering. The output module uses the median sequence of each cluster as a template for the typical power curve of that cluster.

[0017] Preferably, the second processing unit further includes: a multi-branch collaborative prediction model, the multi-branch collaborative prediction model comprising: Input layer: Aggregates all features required by the model and integrates multi-source data according to branch requirements; Parallel branching layer: includes time-series branch, attribute branch, and business constraint branch; among them, a Transformer encoder with time position encoding is used as the time-series branch, inputting the 24-hour electricity weight sequence and the hourly electricity spot price sequence of the user to be predicted for the past 7 days, to calculate the time-series feature vector; a LightGBM model is used as the attribute branch, inputting the user's basic attribute features, market attribute features, and target clustering information, to calculate the attribute feature vector; a 3-layer MLP network is used as the business constraint branch, inputting industry production standard features and business feature vector, to obtain the business feature vector and business compliance pre-score; Fusion layer: The three-branch outputs are weighted and fused through an attention mechanism to generate a fused feature vector, which is then fed into the fully connected layer for output. Output layer: Each sub-layer outputs the power weight for one hour, ultimately yielding a 24-dimensional time-division power weight prediction result.

[0018] Thirdly, the present invention also provides a multi-dimensional attribute electricity user time-of-use electricity clustering analysis and prediction model, wherein the prediction model is a typical electricity consumption pattern clustering template or a multi-branch collaborative prediction model in the multi-dimensional attribute electricity user time-of-use electricity clustering analysis and prediction method described in the first aspect.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention integrates multi-dimensional data such as user basic attributes, electricity spot market price signals, and business plans, extracting three core features: time series, market, and business. It particularly incorporates key attributes such as distributed energy capacity (PV / energy storage) and price sensitivity coefficients to comprehensively capture the influencing factors of user electricity consumption behavior in a market-oriented environment, addressing the technical pain point of single feature dimensions. Through a triple clustering mechanism, combining the advantages of ST-DBSCAN and DBSCAN algorithms, it ensures the temporal continuity of electricity consumption patterns while eliminating erroneous clustering samples through business rule constraints, ultimately forming 10-15 high-purity typical electricity consumption pattern templates. Simultaneously, the templates have a dynamic update mechanism, adapting to user electricity consumption pattern drift and effectively solving the problem of insufficient prediction accuracy caused by differences in electricity consumption patterns among different industries and user types.

[0020] This invention presents an innovative multi-branch collaborative prediction model that accurately captures temporal dependencies using a Transformer encoder, fits nonlinear attribute associations using a LightGBM model, embeds business constraints using a 3-layer MLP network, and then dynamically weights and fuses these features through an attention mechanism, achieving deep collaboration among the three types of features. Compared to a single model, this architecture takes into account temporal patterns, attribute associations, and business compliance, significantly improving prediction accuracy and robustness.

[0021] Furthermore, this invention achieves differentiated dynamic adjustment of time-of-use electricity weight based on user price sensitivity levels, and at the same time, through industry business rule verification and correction, it ensures that the prediction results are both adapted to market price orientation and in line with production and operation logic. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the multi-dimensional attribute electricity user time-of-use electricity clustering analysis and prediction method in an embodiment of the present invention.

[0023] Figure 2 This is a structural diagram of a multi-dimensional attribute electricity user time-of-use electricity clustering analysis and prediction system shown in an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0025] Example 1: As Figure 1 As shown, this invention proposes a multi-dimensional attribute time-of-use electricity consumption clustering analysis and prediction method for electricity users, including: Collect basic data, electricity spot market data and business data of users to be predicted, and extract multi-dimensional prediction input features containing time series features, market features and business features; Call the pre-trained typical electricity consumption pattern cluster templates, calculate the similarity between the user to be predicted and each cluster template based on the multi-dimensional prediction input features, and determine the target cluster and the corresponding typical electricity consumption curve; The multidimensional prediction input features and target clustering information are input into the multi-branch collaborative prediction model to make predictions and obtain the initial 24-hour time-of-use electricity weights of the users to be predicted. By combining electricity spot market price signals with the price sensitivity levels of users to be predicted, the initial 24-hour time-of-use electricity weights are dynamically adjusted, and industry business rules are simultaneously verified and corrected. The optimized time-of-use electricity weights are output at multiple time scales, including day-ahead, intraday, and real-time, and a hierarchical explanatory report containing feature contributions and business compliance is generated simultaneously.

[0026] Preferably, the typical electricity consumption pattern clustering template includes: Three years of multi-source training data were collected and proportionally divided into clustering training and validation sets. Preferably, the division ensured that both datasets contained user samples from different industries and installation types in equal proportions to avoid uneven sample distribution. The multi-source training data included historical user electricity consumption data, historical market data, and historical business data.

[0027] The training set is trained using triple clustering; the triple clustering includes a first-level spatiotemporal clustering, a second-level density clustering, and a third-level business clustering. The median sequence of each cluster is used as the template for the typical power curve of that cluster; that is, for the 12 finalized typical clusters, the 24-hour power weight sequence of all samples in each cluster is extracted, and the median is taken by hour to form the template for the typical power curve of each cluster.

[0028] Preferably, the typical electricity consumption pattern clustering template further includes: template verification.

[0029] Specifically, template validation includes: Accuracy verification: The MAPE of samples within each cluster and the corresponding template was calculated. The average MAPE of the 12 clusters was 6.8% (≤8%), with the industrial continuous production cluster having the lowest MAPE (4.2%) and the residential flexible electricity consumption cluster having the highest MAPE (7.9%). Quantitative verification: The silhouette coefficient of the clustering results is 0.72 (≥0.6), and the Davies-Bouldin index is 1.05 (≤1.2), which meets the clustering quality requirements; Expert verification: Five experts from the power industry and user industry were invited to review the business rationality of the 12 templates. The verification accuracy rate reached 93% (≥90%), confirming that the templates are highly matched with actual power consumption scenarios. Preferably, tags are added to each typical template, including industry attributes (industrial / commercial / residential), price sensitivity level (high / medium / low), time series characteristics (continuous and stable / peak and fluctuating), and installation compatibility type (photovoltaic / energy storage / none), to facilitate rapid matching later.

[0030] Preferably, the typical electricity consumption pattern clustering template further includes a template dynamic update mechanism, such as periodic updates or triggered updates.

[0031] Regular updates, with a preference for updates every 3 months: At the end of each quarter, the template library is updated using the incremental DBSCAN algorithm based on 3 months of newly added historical data. Similarity is calculated only for the newly added data, eliminating the need for full retraining. For example, during the Q2 update in 2024, the proportion of samples in the commercial new energy hotel cluster increased by 8% in the newly added data, which did not reach the trigger threshold. Only the typical curve of this cluster was slightly adjusted, and the MAPE of the updated template remained at 6.5%. Triggered Updates: Scenario 1: An industrial machinery processing cluster initially had a sample share of 20%. Due to the addition of 20 similar users, the sample share rose to 36% after 3 months (a change exceeding 15%), triggering an immediate update. Incremental DBSCAN was used to cluster the data of the new users, merging with the original template. After the update, the average MAPE of the samples within the cluster decreased from 7.3% to 6.1%. Scenario 2: A residential winter heating cluster experienced an extreme cold wave, with the actual electricity consumption data and the template's MAPE reaching 13.5% (exceeding 12%) for 15 consecutive days, triggering an immediate update. The median curve for this cluster was recalculated, incorporating electricity consumption characteristics under extreme weather conditions. After the update, the MAPE decreased to 9.2% for 7 consecutive days, returning to a reasonable range.

[0032] In summary, this embodiment monitors the application status of each template in the template database in real time. When any of the following conditions are met, an immediate update is triggered: the actual user sample corresponding to any template accounts for more than 15% of the change in user samples when the template was generated; or in the prediction application of any template, the actual electricity consumption data for 15 consecutive days exceeds the MAPE of the template by more than 12%.

[0033] Preferably, the typical electricity consumption pattern clustering template further includes: The first layer of spatiotemporal clustering is used to capture continuous electricity consumption patterns. This first layer employs a spatiotemporal density clustering algorithm, using a single user's feature sequence over seven consecutive days as one spatiotemporal sample. For example, user A's 24-hour weight sequence over seven consecutive days, daily price sensitivity coefficient, industry attributes, and other features are integrated into one spatiotemporal sample to ensure that the sample reflects the user's continuous electricity consumption patterns over time. Key parameters of ST-DBSCAN are determined using a K-distance graph, and all spatiotemporal samples in the clustering training set are clustered to output clusters of continuous electricity consumption patterns. Preferably, the step of determining the key parameters of ST-DBSCAN through the K-distance plot and clustering all spatiotemporal samples in the clustering training set specifically includes: calculating the pairwise Euclidean distance between all spatiotemporal samples in the clustering training set (a total of 1200×112=134400 samples, 112=788 days÷7 days / sample rounded down); sorting the distances of each sample in ascending order, and plotting a K-distance plot (the horizontal axis is the sample number, and the vertical axis is the distance value) using the top 20% of the distance values; selecting the distance value corresponding to the inflection point in the plot as the neighborhood radius eps=0.32, and setting the minimum number of samples min_samples=5 based on the sample density. ST-DBSCAN clustering is then performed on all spatiotemporal samples, outputting 20 temporally continuous electricity consumption pattern clusters. Examples of some cluster labels include: industrial continuous production, no photovoltaic, weekday stable cluster; commercial retail, weekend peak cluster; residential, winter energy-saving energy storage cluster.

[0034] The second-level density clustering uses the DBSCAN algorithm, taking the 20 pattern clusters output from the first level as independent units and performing secondary clustering one by one to avoid cross-cluster interference. For samples within each cluster, DBSCAN secondary clustering is performed on each time-series pattern cluster based on the dual criteria of electricity consumption pattern similarity and price sensitivity coefficient similarity, splitting niche patterns with large differences in price sensitivity and different electricity consumption details within the cluster, and outputting subdivided pattern clusters; for example, the original industrial continuous production - no photovoltaic - stable working day cluster is split into industrial continuous production - high price sensitive sub-cluster and industrial continuous production - low price sensitive sub-cluster, solving the problem of intra-cluster heterogeneity.

[0035] Preferably, the similarity of power consumption patterns is used to calculate the cosine similarity of the 24-hour weighted sequence between the sample and the cluster center; Price sensitivity coefficient similarity is calculated as the absolute difference between the price sensitivity coefficient of the sample and the center of the cluster.

[0036] In addition to the third level of business clustering, for each of the subdivided pattern clusters output by the second level, the samples within the cluster are checked to see if they conform to the business constraints of the corresponding industry, and cross-industry mis-clustered samples that do not conform to the constraints are removed; preferably, if ≥10% of the samples in a certain cluster do not conform to the business constraints of the corresponding industry, it is determined to be a cross-industry mis-clustered cluster. Finally, based on the business constraint verification results, the cosine similarity of the cluster centers between effective sub-clusters is calculated, and sub-clusters with a similarity ≥ 0.9 are merged to ultimately determine 10-15 typical electricity consumption pattern clusters. For example, the commercial hotel - weekday peak cluster and the commercial office building - weekday peak cluster have a similarity of 0.92 and are merged into the commercial office - weekday peak cluster.

[0037] Preferably, the construction of an industry business constraint rule base includes: For industrial users with continuous production: the total electricity consumption weight from 22:00 to 6:00 the next day must be ≥30%, and the weight for a single hour must be ≥0.025; Business users: Total electricity consumption weighting between 8:00 and 22:00 during the day should be ≥70%, and the peak weighting on holidays should be ≥15% higher than on weekdays; Photovoltaic installation users: Their daytime electricity consumption weighting from 10:00 to 15:00 is ≥12% lower than that of users without photovoltaic systems in the same industry; Energy storage users: The weight of off-peak hours (23:00-5:00) is ≥8% higher than that of normal hours (charging characteristics).

[0038] Preferably, the multi-branch collaborative prediction model includes: A Transformer encoder with temporal position encoding is used as a temporal branch to capture the long-term and short-term temporal dependencies between electricity consumption weights and prices. The input consists of a 24-hour electricity consumption weight sequence and an hourly electricity spot price sequence for the user to be predicted over the past 7 days. The temporal feature vector is calculated. The Transformer encoder with temporal position encoding has 6 structural encoder layers, 8 attention heads, 256 hidden layer dimensions, and 336 word embedding dimensions. It also uses sine and cosine position encoding and incorporates hourly timestamps.

[0039] The LightGBM model is used as the attribute branch to fit the nonlinear relationship between user static attributes, market dynamic attributes and cluster information on electricity consumption weight. The user basic attribute features, market attribute features and target cluster information are input first, and the attribute feature vector is calculated. A 3-layer MLP network is used as the business constraint branch. The industry production standard features and business feature vectors are input to obtain the business feature vector and the business compliance pre-score. The three branch outputs are weighted and fused using an attention mechanism to generate a fused feature vector that is then connected to the fully connected layer. Each sub-layer outputs an hourly power weight, resulting in a 24-dimensional time-sharing power weight prediction result.

[0040] This preferred embodiment is the core optimization scheme of the aforementioned multi-branch collaborative prediction model. For the output feature vectors of the time-series branch, attribute branch, and business constraint branch, a two-dimensional adaptive attention mechanism is used to complete weighted fusion. The generated fused feature vector is then connected to a fully connected layer, outputting the power weights hourly, ultimately obtaining a 24-dimensional time-sharing power weight prediction result. The specific implementation is as follows: Step 1: Initialize the basic weights for the two-dimensional attention; This embodiment employs a dual-dimensional attention weight system, consisting of industry-specific basic weights and dynamically adjusted branch quality weights, which is the core of this preferred solution. First, basic branch weights are configured for different types of power users, tailored to their electricity consumption characteristics. These weights are empirically optimal values, set in conjunction with the electricity consumption patterns of the power industry. The specific basic weight allocation is as follows, and the sum of the basic weight coefficients is always equal to 1: For industrial continuous production users (low-sensitivity users): Time-series branch weight W t0 =0.3, attribute branch weight W a0 =0.3, Business constraint branch weight W b0 =0.4; For this type of user, power stability is prioritized, and business production specifications have the greatest impact on power consumption weight, so the business branch has the highest weight ratio.

[0041] For business users (highly sensitive users): Time-series branch weight W t0 =0.4, attribute branch weight W a0 =0.4, Business constraint branch weight W b0 =0.2; The electricity consumption behavior of this type of user is significantly affected by both time-series patterns and market price attributes, so the weight of time-series and attribute branches is higher.

[0042] For residential and distributed energy installation users: Time-series branch weight W t0 =0.5, attribute branch weight W a0 =0.3, Business constraint branch weight W b0 =0.2; This type of user's electricity consumption behavior has the strongest temporal periodicity, so the temporal branch weight accounts for the highest proportion.

[0043] Step 2: Dynamically adjust attention weights based on branch output quality; Based on the industry-specific weights, and considering the actual output quality of each branch, the weight coefficients are dynamically adjusted by ±0.05 to ±0.12. After adjustment, the weights of each branch still satisfy the constraint that "the sum of the weight coefficients = 1". The adjustment logic aligns with the output characteristics of each branch without subjective adjustments. The specific adjustment rules are as follows: Temporal branch weight correction: Based on the temporal fitting error correction of the Transformer encoder, if the temporal fitting error is ≤4.8% (good fitting effect), then W t =W t0 +0.08; If the time series fitting error is >6.0% (the fitting effect is generally poor), then W t =W t0 0.05; the fitting error is between 4.8% and 6.0%, and the weights remain unchanged from the base value.

[0044] Attribute branch weight adjustment: Based on the feature importance contribution adjustment of the LightGBM model, if the cumulative contribution of the top 5 core features is ≥0.7, then W a =W a0 +0.06; if the cumulative contribution is <0.5, then W a =W a0 0.07; the contribution rate is between 0.5 and 0.7, and the weight remains unchanged at the base value.

[0045] Business constraint branch weight adjustment: Based on the pre-score of business compliance, this is the core basis for adjustment of this branch; if the pre-score of compliance is ≥0.9 (high compliance), then W b =W b0 +0.12; If the preliminary score is between 0.8 and 0.9 (good compliance), then W b =W b0 +0.05; if the pre-score is <0.7 (poor compliance), then W b =W b0 0.08.

[0046] Step 3: Normalization and calibration of attention weight coefficients; After completing the weight assignment and correction for the two dimensions, the preliminary attention weight coefficients W for the three branches are obtained. t W a W b To ensure the effectiveness and binding force of the weighting coefficients, L1 normalization calibration is performed on them.

[0047] ; After normalization calibration, the final attention weight coefficient W is obtained. t′ Wa′ W b′ Satisfying W t′ +W a′ +W b′ ≡1, this coefficient is the final weighting coefficient of the feature vectors of each branch, with no overflow and no invalid values.

[0048] Step 4: Weighted fusion to generate fused feature vectors; Based on the calibrated attention weight coefficients, a weighted summation operation is performed on the dimension-aligned temporal feature vector Ft, attribute feature vector Fa, and business feature vector Fb to complete feature fusion. The fusion formula is: Fmerge = W t′ Ft+W a′ F a +W b′ F b .

[0049] It should be noted that the above formula ultimately generates a 512-dimensional fusion feature vector Fmerge. This fusion feature vector is not a simple feature splicing, but a comprehensive feature that deeply integrates the time-series dependence of user electricity consumption, the correlation of multi-dimensional attributes, and the constraints of industry business. It fully preserves the core feature information of the three branches, and at the same time, it realizes the differentiated empowerment of features through attention weights, allowing the core influencing features to play a leading role in the prediction process.

[0050] The 512-dimensional fused feature vector Fmerge generated above is fed into a pre-defined fully connected layer network to complete the final prediction of the 24-dimensional time-sharing power weights. This step is the mapping step from fused features to prediction results, and the specific implementation is as follows: Fully connected layer network structure: It adopts an adaptive structure of "single backbone + multiple sub-layers". The backbone layer is a 512-dimensional input layer with ReLU activation function, and the hidden layer has a dimension of 256. After the backbone layer, 24 independent fully connected sub-layers are set in parallel. The 24 sub-layers correspond one-to-one with the 24-hour period from 0:00 to 23:00 of the next day. Each sub-layer is an independent feature mapping unit and does not interfere with each other.

[0051] Hourly weight output: The input of each fully connected sub-layer is a 256-dimensional hidden layer feature, and the output is a 1-dimensional numerical value. After processing by the Sigmoid activation function, the numerical value is mapped to the [0,1] interval to obtain the hourly electricity weight value of the corresponding hour. The 24 fully connected sub-layers operate synchronously and output 24 independent hourly weight values ​​to form the initial 24-dimensional hourly electricity weight sequence [w0,w1,w2...w23].

[0052] After the above fully connected layer operations, the final 24-dimensional time-of-use electricity weight prediction result for the user to be predicted is obtained. This result is a standardized weight sequence, which can be directly used in the subsequent dynamic weight adjustment process combining electricity spot market price signals and user sensitivity levels, or it can be directly used as the basic data for the day-ahead declaration of electricity sales companies. Compared with the fixed weight splicing and simple feature superposition of existing technologies, the attention mechanism weighted fusion in this embodiment can dynamically allocate weights according to the user's industry characteristics and the output quality of each branch. The fit and effectiveness of feature fusion are significantly improved, allowing the prediction model to adapt to the electricity consumption patterns of different types of users, and solving the problem of poor adaptability of single fusion methods.

[0053] Preferably, the step of dynamically adjusting the initial 24-hour time-of-use electricity weights by combining the electricity spot market price signal with the price sensitivity level of the user to be predicted, and simultaneously performing industry business rule verification and correction, includes: The K-means clustering algorithm is used to cluster the day-ahead price curve for the forecast date into time periods; the clustering results include price peak periods, price average periods, and price trough periods; price volatility is calculated separately for each type of time period. Preferably, the clustering parameters are: K=3 (fixed as peak segment, average segment, and valley segment), and the K-means++ algorithm is used to initialize the cluster centers, with 100 iterations. Input the price curve P_day before the date to be predicted, perform K-means clustering, and automatically divide it into 3 time periods. Example results: Price peak period T_peak=[10,11,18,19] (10:00-12:00, 18:00-20:00, a total of 4 hours); Price flat period T_flat=[8,9,12,13,14,15,16,17,20,21] (8:00-10:00, 12:00-18:00, 20:00-22:00, a total of 10 hours); Price trough range T_valley=[0,1,2,3,4,5,6,7,22,23] (22:00-8:00 the next day, a total of 10 hours); Output: Time period segmentation result T=[T_peak,T_flat,T_valley], indicating the price time period type corresponding to each hour.

[0054] Price volatility calculation for different time periods: Average price over a period: P_avg_peak=mean(P_day[T_peak]), P_avg_flat=mean(P_day[T_flat]), P_avg_valley=mean(P_day[T_valley]); Standard deviation of prices over time: P_std_peak=std(P_day[T_peak]), P_std_flat=std(P_day[T_flat]), P_std_valley=std(P_day[T_valley]); Volatility calculation: V_peak = P_std_peak / P_avg_peak, V_flat = P_std_flat / P_avg_flat, V_valley = P_std_valley / P_avg_valley; Constraint handling: Volatility is limited to 0.05≤V≤0.3. In the example, V_peak=0.22 (compliant), V_flat=0.11 (compliant), and V_valley=0.08 (compliant). If the volatility exceeds 0.3 in a certain period, it is taken as 0.3; if it is below 0.05, it is taken as 0.05.

[0055] Intraday price correction (intraday forecast scenario): For intraday rolling forecasts (updated every hour), adjust prices and volatility using an intraday price correction factor α: Corrected price: P_day_corr[i] = P_day[i] × α[i] (i = 0-23); Corrected volatility: V_peak_corr = V_peak × mean(α[T_peak]) (in the example, mean(α[T_peak]) = 1.08, then V_peak_corr = 0.22 × 1.08 = 0.2376), the same applies to the flat and trough segments; Output: For day-ahead forecasts, use V=[V_peak,V_flat,V_valley]; for intraday forecasts, use V_corr=[V_peak_corr,V_flat_corr,V_valley_corr].

[0056] The initial 24-hour time-of-use electricity weights are dynamically adjusted and normalized based on user sensitivity levels and price volatility. The normalized weights are then iterated hourly according to industry business rules, and any violations are corrected according to the principle of minimum adjustment until there are no violations, at which point the final 24-hour time-of-use electricity weights are output.

[0057] Preferably, the step of dynamically adjusting the initial 24-hour time-of-use electricity weights based on user sensitivity levels and price volatility includes: For highly sensitive users, adjustments are made based on price peaks, averages, and troughs. The initial 24-hour time-of-use electricity weights are dynamically adjusted using a multiplication factor, taking into account price volatility and sensitivity coefficients. For users with moderate sensitivity, the weighting is adjusted only for price peak and trough periods, while the initial weighting remains for the flat period. In addition, adjustments will be made for low-sensitivity users, with the weight of non-core loads being reduced only when the volatility exceeds 0.2 during peak price periods.

[0058] Based on the user sensitivity level S and the preprocessed price data, the initial weights are adjusted according to the principles of peak suppression and trough incentive, and then normalization is performed. 1. Adjustment for highly sensitive users (S=1.0, example: a shopping mall) Adjust the formula: Peak segment (i∈T_peak): w i1 =w i0 ×(1-V_peak_corr×0.8×S)=w i0 ×(1-0.2376×0.8×1.0)=w i0 ×0.810 (down 19.0%) Valley segment (i∈T_valley): w i1 =w i0 ×(1+V_valley_corr×0.6×S)=w i0 ×(1+0.08×1.08×0.6×1.0)=w i0 ×1.0518 (up 5.18%) Flat segment (i∈T_flat): w i1 =w i0 ×(1-V_flat_corr×0.2×S)=w i0 ×(1+0.11×1.05×0.2×1.0)=w i0 ×0.9769 (down 2.31%) Constraint application: Adjusted hourly weight 0.03 ≤ w i1 ≤0.18, the initial weight w in the example at hour 18. 18 =0.07, adjusted w 181 =0.07×0.810=0.0567 (meets the constraints).

[0059] 2. Adjustment for moderately sensitive users (S=0.5, example: a machine processing plant) Adjust the formula: Peak segment: w i1 =w i0 ×(1-V_peak_corr×0.5×S)=w i0 ×(1-0.2376×0.5×0.5)=w i0 ×0.9406 (down 5.94%) Valley segment: w i1 =w i0 ×(1+V_valley_corr×0.3×S)=w i0 ×(1+0.08×1.08×0.3×0.5)=w i0 ×1.0130 (up 1.30%) Average value segment: w i1 =w i0 (No adjustment); Constraint application: The adjusted weight of adjustable load is ≤30%. In the example, the total hourly weight of adjustable load is 0.28 (which meets the constraint).

[0060] 3. Adjustment for low-sensitivity users (S=0.2, example: a chemical company) Adjust the formula: Peak range (V_peak_corr=0.2376>0.2): w i1 =w i0 ×(1-V_peak_corr×0.3×S)=w i0 ×(1-0.2376×0.3×0.2)=w i0 ×0.9869 (down 1.31%) Valley segment, average segment: w i1 =w i0 (No adjustment); Constraint Application: The adjusted core production load weight ratio is ≥80%, and in the example, the total hourly weight corresponding to the core load is 0.85 (meets the constraint).

[0061] 4. Adjusted weight normalization Calculate the adjusted weighted sum: Sum_w1 = sum(w i1 In the example, the highly sensitive user Sum_w1=1.02, normalization is performed: w i _adjust=w i1 / Sum_w1; Verification after normalization: ΣW_adjust=1, hourly weight 0.01≤w i _adjust≤0.2, in the example, the weight of the 18th hour after normalization is 0.0567 / 1.02≈0.0556 (meets the constraint).

[0062] Preferably, the industry business rule verification and correction in this embodiment specifically includes: Hourly weighted traversal verification: Match the corresponding business rules based on the user's industry tag, iterate through the 24-hour weights of W_adjust, and mark violations: Example 1 (Chemical company, low-sensitivity user): The total weight from 22:00 to 6:00 at night = 0.28 (<0.3), and the weight of the second hour = 0.02 (<0.025), which is marked as insufficient weight at night and too low weight in a single hour; Example 2 (Shopping mall, highly sensitive users): The total weight of daytime 8:00-22:00 = 0.68 (<0.7), which is marked as insufficient daytime weight.

[0063] Correction of violations (minimum adjustment principle): Example 1 (Chemical Enterprise) Correction: The required additional weight is calculated as Δ = 0.3 - 0.28 = 0.02. Weights are extracted from the daytime average period (12:00-14:00, non-critical production period), with a weight of 0.06 for the 12th hour and 0.055 for the 13th hour, and 0.01 is extracted from each. Transferred to nighttime violation hours: The weight of the 2nd hour changed from 0.02 to 0.03, and the weight of the 3rd hour changed from 0.022 to 0.027. After the adjustment, the total weight of the nighttime violation is 0.30, indicating no violation. Example 2 (Shopping Mall) Correction: The required additional weight is calculated as Δ = 0.7 - 0.68 = 0.02. Weights are extracted from the nighttime off-peak period (2:00-4:00, when there is no electricity demand for businesses), with a weight of 0.04 for the 3rd hour and a total weight of 0.02. The weighting was shifted to the 9th hour, which had the lowest daytime weight (0.028 → 0.048). After the adjustment, the total daytime weight was 0.70, which was not a violation.

[0064] Final verification: Iterate through the corrected weights W_final again to confirm that there are no violations and that ΣW_final=1 (error ≤ 0.001). Output: Final 24-hour time-of-use power consumption weights W_final=[w0_final,w1_final,...,w 23 [_final], simultaneously generate weight adjustment and compliance reports.

[0065] This embodiment achieves dynamic optimization of initial weights through a complete process of price signal preprocessing, differentiated weight adjustment, and business rule verification and correction. K-means clustering accurately divides price periods, volatility calculation quantifies the intensity of price fluctuations, and a differentiated adjustment formula based on user sensitivity levels adapts to different user types. The principle of minimum adjustment corrects violations to ensure compliance. The final output weights not only conform to the price orientation of the electricity spot market but also meet industry production constraints, fully adapting to the market-based reporting needs of electricity sales companies.

[0066] Example 2, as Figure 2 As shown, the present invention also provides a multi-dimensional attribute time-of-use electricity consumption clustering analysis and prediction system for electricity users, the prediction system comprising: The data acquisition unit collects basic data, electricity spot market data, and business data from users to be predicted. The data preprocessing unit extracts multidimensional predictive input features, including time-series features, market features, and business features, from the collected data. The first processing unit calls the pre-trained typical electricity consumption pattern clustering templates, calculates the similarity between the user to be predicted and each clustering template based on the multi-dimensional prediction input features, and determines the target cluster and the corresponding typical electricity consumption curve. The second processing unit inputs the multi-dimensional prediction input features and target clustering information into the multi-branch collaborative prediction model to make predictions and obtain the initial 24-hour time-of-use electricity weights of the users to be predicted. The third processing unit dynamically adjusts the initial 24-hour time-of-use electricity weights by combining the electricity spot market price signals with the price sensitivity levels of the users to be predicted, and simultaneously executes industry business rule verification and correction. The output unit outputs optimized time-of-use electricity weights at multiple time scales, including day-ahead, intraday, and real-time, and simultaneously generates a hierarchical explanation report containing feature contributions and business compliance.

[0067] Preferably, the first processing unit further includes: a typical electricity consumption pattern clustering template, the typical electricity consumption pattern clustering template including: Dataset module: Collects multi-source training data from the past 3 years and divides it into clustering training set and validation set according to proportion; The clustering module performs triple clustering training on the training set; the triple clustering includes a first-level spatiotemporal clustering, a second-level density clustering, and a third-level business clustering. The output module uses the median sequence of each cluster as a template for the typical power curve of that cluster.

[0068] Preferably, the second processing unit further includes: a multi-branch collaborative prediction model, the multi-branch collaborative prediction model comprising: Input layer: Aggregates all features required by the model and integrates multi-source data according to branch requirements; Parallel branching layer: includes time-series branch, attribute branch, and business constraint branch; among them, a Transformer encoder with time position encoding is used as the time-series branch, inputting the 24-hour electricity weight sequence and the hourly electricity spot price sequence of the user to be predicted for the past 7 days, to calculate the time-series feature vector; a LightGBM model is used as the attribute branch, inputting the user's basic attribute features, market attribute features, and target clustering information, to calculate the attribute feature vector; a 3-layer MLP network is used as the business constraint branch, inputting industry production standard features and business feature vector, to obtain the business feature vector and business compliance pre-score; Fusion layer: The three-branch outputs are weighted and fused through an attention mechanism to generate a fused feature vector, which is then fed into the fully connected layer for output. Output layer: Each sub-layer outputs the power weight for one hour, ultimately yielding a 24-dimensional time-division power weight prediction result.

[0069] Example 3: This invention also provides a multi-dimensional attribute electricity user time-of-use consumption clustering analysis and prediction model. The prediction model is a typical electricity consumption pattern clustering template or a multi-branch collaborative prediction model in the multi-dimensional attribute electricity user time-of-use consumption clustering analysis and prediction method described in the first aspect. Although exemplary embodiments have been described herein with reference to the accompanying drawings, it should be understood that the above exemplary embodiments are merely illustrative and are not intended to limit the scope of the invention thereto. Various changes and modifications can be made therein by those skilled in the art without departing from the scope and spirit of the invention. All such changes and modifications are intended to be included within the scope of the invention as claimed in the appended claims.

[0070] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the statement "comprising a…" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0071] Although the description of the invention has been given in conjunction with the specific embodiments described above, it will be apparent to those skilled in the art that many substitutions, modifications, and variations can be made based on the foregoing. Therefore, all such substitutions, modifications, and variations are included within the spirit and scope of the appended claims.

Claims

1. A method for clustering and predicting time-of-use electricity consumption of multi-dimensional attribute electricity users, characterized in that, include: Collect basic data, electricity spot market data and business data of users to be predicted, and extract multi-dimensional prediction input features containing time series features, market features and business features; Call the pre-trained typical electricity consumption pattern cluster templates, calculate the similarity between the user to be predicted and each cluster template based on the multi-dimensional prediction input features, and determine the target cluster and the corresponding typical electricity consumption curve; The multidimensional prediction input features and target clustering information are input into the multi-branch collaborative prediction model to make predictions and obtain the initial 24-hour time-of-use electricity weights of the users to be predicted. By combining electricity spot market price signals with the price sensitivity levels of users to be predicted, the initial 24-hour time-of-use electricity weights are dynamically adjusted, and industry business rules are simultaneously verified and corrected. The optimized time-of-use electricity weights are output at multiple time scales, including day-ahead, intraday, and real-time, and a hierarchical explanatory report containing feature contributions and business compliance is generated simultaneously.

2. The method for multi-dimensional attribute time-of-use electricity consumption clustering analysis and prediction for electricity users according to claim 1, characterized in that, The typical electricity consumption pattern clustering template includes: The training data collected from multiple sources over the past three years were divided into clustered training sets and validation sets according to a certain ratio. The training set is trained using triple clustering; the triple clustering includes a first-level spatiotemporal clustering, a second-level density clustering, and a third-level business clustering. The median sequence of each cluster is used as the template for the typical charge curve of that cluster; Among them, the application status of each template in the real-time monitoring template database is triggered to update immediately when any of the following conditions are met: the actual user sample ratio corresponding to any template exceeds 15% of the change in user samples when the template is generated; or in the prediction application of any template, the actual electricity consumption data for 15 consecutive days exceeds 12% of the template's MAPE.

3. The method for multi-dimensional attribute time-of-use electricity consumption clustering analysis and prediction for electricity users according to claim 2, characterized in that, The typical electricity consumption pattern clustering template also includes: The first spatiotemporal clustering uses a spatiotemporal density clustering algorithm, taking the feature sequence of a single user for 7 consecutive days as one spatiotemporal sample. The key parameters of ST-DBSCAN are determined through the K-distance graph, and all spatiotemporal samples in the clustering training set are clustered to output a time-continuous electricity consumption pattern cluster. The second density clustering uses the DBSCAN algorithm. Taking the pattern clusters generated by the first clustering as units, for each sample in each cluster, DBSCAN secondary clustering is performed on each time-series pattern cluster according to the dual criteria of electricity consumption pattern similarity and price sensitivity coefficient similarity, and the subdivided pattern clusters are output. In addition to the third level of business clustering, for each of the subdivided pattern clusters output by the second level, we verify whether the samples within the cluster conform to the business constraints of the corresponding industry and remove cross-industry mis-clustered samples that do not conform to the constraints. Finally, based on the business constraint verification results, sub-clusters with a similarity of ≥0.9 are merged, and 10-15 typical electricity consumption pattern clusters are finally determined.

4. The method for multi-dimensional attribute time-of-use electricity consumption clustering analysis and prediction for electricity users according to claim 3, characterized in that, The multi-branch collaborative prediction model includes: A Transformer encoder with time position encoding is used as the time series branch. The input consists of a 24-hour electricity weight sequence of the user to be predicted over the past 7 days and an hourly electricity spot price sequence over the past 7 days. The time series feature vector is calculated. The LightGBM model is used as the attribute branch. The user's basic attribute features, market attribute features and target clustering information are input to calculate the attribute feature vector. A 3-layer MLP network is used as the business constraint branch. The industry production standard features and business feature vectors are input to obtain the business feature vector and the business compliance pre-score. The three branch outputs are weighted and fused using an attention mechanism to generate a fused feature vector that is then connected to the fully connected layer. Each sub-layer outputs an hourly power weight, resulting in a 24-dimensional time-sharing power weight prediction result.

5. The method for multi-dimensional attribute time-of-use electricity consumption clustering analysis and prediction according to claim 4, characterized in that, The method involves dynamically adjusting the initial 24-hour time-of-use electricity weights by combining electricity spot market price signals with the price sensitivity levels of the users to be predicted, while simultaneously performing industry business rule verification and correction, including: The K-means clustering algorithm is used to cluster the day-ahead price curve for the forecast date into time periods; the clustering results include price peak periods, price average periods, and price trough periods; price volatility is calculated separately for each type of time period. The initial 24-hour time-of-use electricity weights are dynamically adjusted and normalized based on user sensitivity levels and price volatility. The normalized weights are then iterated hourly according to industry business rules, and any violations are corrected according to the principle of minimum adjustment until there are no violations, at which point the final 24-hour time-of-use electricity weights are output.

6. The method for multi-dimensional attribute time-of-use electricity consumption clustering analysis and prediction for electricity users according to claim 5, characterized in that, The dynamic adjustment of the initial 24-hour time-of-use electricity weights based on user sensitivity levels and price volatility includes: For highly sensitive users, adjustments are made based on price peaks, averages, and troughs. The initial 24-hour time-of-use electricity weights are dynamically adjusted using a multiplication factor, taking into account price volatility and sensitivity coefficients. For users with moderate sensitivity, the weighting is adjusted only for price peak and trough periods, while the weighting remains the same for the flat period. In addition, adjustments will be made for low-sensitivity users, with the weight of non-core loads being reduced only when the volatility exceeds 0.2 during peak price periods.

7. A prediction system for a multi-dimensional attribute time-of-use electricity consumption clustering analysis and prediction method according to any one of claims 1-6, characterized in that, The prediction system includes: The data acquisition unit collects basic data, electricity spot market data, and business data from users to be predicted. The data preprocessing unit extracts multidimensional predictive input features, including time-series features, market features, and business features, from the collected data. The first processing unit calls the pre-trained typical electricity consumption pattern clustering templates, calculates the similarity between the user to be predicted and each clustering template based on the multi-dimensional prediction input features, and determines the target cluster and the corresponding typical electricity consumption curve. The second processing unit inputs the multi-dimensional prediction input features and target clustering information into the multi-branch collaborative prediction model to make predictions and obtain the initial 24-hour time-of-use electricity weights of the users to be predicted. The third processing unit dynamically adjusts the initial 24-hour time-of-use electricity weights by combining the electricity spot market price signals with the price sensitivity levels of the users to be predicted, and simultaneously executes industry business rule verification and correction. The output unit outputs optimized time-of-use electricity weights at multiple time scales, including day-ahead, intraday, and real-time, and simultaneously generates a hierarchical explanation report containing feature contributions and business compliance.

8. The prediction system according to claim 7, characterized in that, The first processing unit further includes: a typical electricity consumption pattern clustering template, wherein the typical electricity consumption pattern clustering template includes: Dataset module: Collects multi-source training data from the past 3 years and divides it into clustering training set and validation set according to proportion; The clustering module performs triple clustering training on the training set; the triple clustering includes a first-level spatiotemporal clustering, a second-level density clustering, and a third-level business clustering. The output module uses the median sequence of each cluster as a template for the typical power curve of that cluster.

9. The prediction system according to claim 8, characterized in that, The second processing unit further includes: a multi-branch collaborative prediction model, wherein the multi-branch collaborative prediction model includes: Input layer: Aggregates all features required by the model and integrates multi-source data according to branch requirements; Parallel branching layer: includes time-series branch, attribute branch, and business constraint branch; among them, a Transformer encoder with time position encoding is used as the time-series branch, inputting the 24-hour electricity weight sequence and the hourly electricity spot price sequence of the user to be predicted for the past 7 days, to calculate the time-series feature vector; a LightGBM model is used as the attribute branch, inputting the user's basic attribute features, market attribute features, and target clustering information, to calculate the attribute feature vector; a 3-layer MLP network is used as the business constraint branch, inputting industry production standard features and business feature vector, to obtain the business feature vector and business compliance pre-score; Fusion layer: The three-branch outputs are weighted and fused through an attention mechanism to generate a fused feature vector, which is then fed into the fully connected layer for output. Output layer: Each sub-layer outputs the power weight for one hour, ultimately yielding a 24-dimensional time-division power weight prediction result.

10. A multi-dimensional attribute electricity user time-of-use electricity clustering analysis and prediction model, wherein the prediction model is a typical electricity consumption pattern clustering template or a multi-branch collaborative prediction model in the multi-dimensional attribute electricity user time-of-use electricity clustering analysis and prediction method as described in any one of claims 1-6.

Images