Cross-scene adaptive flexible load dynamic interval prediction method
By employing multimodal scene recognition and cross-scene adaptive migration methods, the problems of scene adaptability and interval reliability in flexible load forecasting are solved, achieving high-precision and reliable cross-scene forecasting and reducing engineering application costs.
Patent Information
- Application Number
- CN202511612871.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-06
- Publication Date
- 2026-02-27
AI Technical Summary
Existing technologies have poor scenario adaptability in flexible load forecasting, low reliability of interval forecasting, and difficulty in cross-scenario knowledge transfer, resulting in insufficient forecasting accuracy and reliability.
We employ a multimodal scene recognition, dynamic interval prediction, and cross-scene adaptive transfer approach. By fusing features through a cross-attention mechanism, we use an improved K-shape clustering algorithm to automatically identify scenes. We construct a quantile regression model based on Transformer, introduce exponential stack mapping and adaptive bandwidth kernel density estimation, and combine reinforcement learning to adjust feature weights to achieve cross-scene knowledge transfer.
It achieves cross-scenario adaptive capability, reduces prediction error, improves inter-regional reliability, shortens model training time and data requirements, and meets the needs of power grid dispatching.
Smart Images
Figure CN121584533A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system load forecasting technology, and in particular to a flexible load dynamic range forecasting method that is adaptive across scenarios. Background Technology
[0002] With the advancement of new power system construction, the proportion of flexible loads (such as smart homes, industrial controllable loads, and commercial air conditioning) is increasing year by year. Their load characteristics are influenced by multiple factors, including scenario type (industrial / commercial / residential), weather conditions, and user behavior, exhibiting significant nonlinearity, randomness, and scenario-specific differences. Existing technologies have the following shortcomings: 1. Poor scene adaptability: Traditional prediction methods are mostly targeted at a single specific scene (such as transformer area, building), rely on manual classification of scene types (such as "daily / cold wave"), lack the ability to adaptively identify dynamic features of multiple scenes, and the prediction accuracy across scenes is significantly reduced (for example, when an industrial scene model is applied to a commercial scene, the MAPE error increases by 15%-20%).
[0003] 2. Low reliability of interval prediction: The quantile regression method is commonly used for dynamic interval prediction, but there is a quantile crossover problem (different quantile prediction values are disordered), which leads to an imbalance between interval coverage (PICP) and width (PINAW). In extreme scenarios, PICP is often below 90%, which is difficult to support power grid dispatching decisions.
[0004] 3. Difficulty in cross-scenario knowledge transfer: The distribution of load data varies greatly in different scenarios (e.g., industrial load cycles are 8-hour production shifts, while commercial load is affected by pedestrian flow and exhibits a "bi-peak" characteristic). Traditional models do not consider the distribution differences, resulting in weak generalization ability when transferring across scenarios. This requires retraining the model, which is time-consuming and inefficient.
[0005] Therefore, there is an urgent need for a flexible load forecasting method that can automatically identify scene types, dynamically adjust forecast intervals, and achieve cross-scene knowledge transfer in order to improve forecast accuracy and reliability. Summary of the Invention
[0006] To address the aforementioned technical problems, this invention provides a cross-scenario adaptive flexible load dynamic range prediction method.
[0007] This invention provides a cross-scenario adaptive flexible load dynamic interval prediction method, comprising the following steps: S1: Multimodal scene recognition: Collect historical load data, meteorological data and user tag data, perform preprocessing, fuse time-series features and meteorological features through cross-attention mechanism, and use an improved K-shape clustering algorithm to perform unsupervised scene clustering on multimodal features to automatically identify three types of scenes: industrial, commercial and residential. S2: Dynamic Interval Forecasting: Based on Transformer, a quantile regression model is constructed to output load forecast values at multiple quantile levels; an exponential stack mapping is introduced to impose a monotonically increasing constraint on the quantile output to solve the quantile crossover problem; and an adaptive bandwidth kernel density estimation method is used to quantify the forecast uncertainty and generate confidence intervals. S3: Cross-scene adaptive transfer: Common features across scenes are extracted through a pre-trained Transformer encoder. Weighted maximum mean difference loss is introduced to reduce the feature distribution difference between the source scene and the target scene. Reinforcement learning is combined to dynamically adjust the feature weights to achieve cross-scene knowledge transfer.
[0008] Preferably, in step S1, the improved K-shape clustering algorithm determines the optimal number of clusters K by combining the silhouette coefficient and the elbow method, where K=3 corresponds to the three scenarios of industrial, commercial, and residential. Preferably, in step S1, the time-series features include local trend features of the load curve extracted by CNN and long-term dependency features extracted by BiLSTM.
[0009] Preferably, in step S2, the exponent stack mapping normalizes the original quantile output using the Softmax function and forces it to monotonically increase, specifically as follows: in, For the first The final output of the quantiles, σ is the softmax function. This is the original output for the k-th quantile.
[0010] Preferably, in step S2, the adaptive bandwidth kernel density estimation dynamically adjusts the bandwidth parameters according to the scene stability, and the bandwidth is increased by 20% under extreme weather conditions.
[0011] Preferably, in step S3, the expression for the weighted maximum mean difference loss is: Where S and T are the source scene and target scene datasets, respectively. Let H be the feature mapping function, and H be the reproducing kernel Hilbert space.
[0012] Preferably, in step S3, a near-end strategy optimization algorithm is used to dynamically adjust the feature weights under different scenarios.
[0013] Preferably, the loss function of the quantile regression model used in step S2 is the weighted quantile loss: in, , This is the actual load value. To predict quantiles.
[0014] Compared with related technologies, the cross-scenario adaptive flexible load dynamic interval prediction method provided by this invention has the following beneficial effects: This invention provides a cross-scenario adaptive flexible load dynamic interval prediction method. The cross-scenario adaptive capability enables automatic identification of industrial, commercial and residential scenarios, and the cross-scenario prediction MAPE error is reduced to less than 3%, which is better than the traditional model (8%-12%). Improved reliability of the interval: Quantile non-crossing constraints keep PICP stable above 95% and PINAW reduced to 5%-8%, meeting the grid dispatching requirements for "narrow and reliable" intervals; Knowledge transfer efficiency: Through domain-adaptive transfer learning, the training time for new scenario models is reduced by 60%, the data requirement is reduced by 50%, and the cost of engineering applications is reduced. Attached Figure Description
[0015] Figure 1 The structural architecture diagram of the cross-scenario adaptive flexible load dynamic interval prediction system provided by the present invention; Figure 2 The flowchart of the cross-scenario adaptive flexible load dynamic interval prediction method provided by the present invention is shown below. Figure 3 This is a schematic diagram of the quantile non-crossing constraint provided by the present invention; Figure 4 A comparison chart of cross-scene prediction accuracy provided by this invention. Detailed Implementation
[0016] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0017] In the specific implementation process, such as Figures 1-4 As shown, a flexible load dynamic range prediction method that adapts across scenarios includes the following steps: S1: Multimodal scene recognition: Collect historical load data, meteorological data and user tag data, perform preprocessing, fuse time-series features and meteorological features through cross-attention mechanism, and use an improved K-shape clustering algorithm to perform unsupervised scene clustering on multimodal features to automatically identify three types of scenes: industrial, commercial and residential. S2: Dynamic Interval Forecasting: Based on Transformer, a quantile regression model is constructed to output load forecast values at multiple quantile levels; an exponential stack mapping is introduced to impose a monotonically increasing constraint on the quantile output to solve the quantile crossover problem; and an adaptive bandwidth kernel density estimation method is used to quantify the forecast uncertainty and generate confidence intervals. S3: Cross-scene adaptive transfer: Common features across scenes are extracted through a pre-trained Transformer encoder. Weighted maximum mean difference loss is introduced to reduce the feature distribution difference between the source scene and the target scene. Reinforcement learning is combined to dynamically adjust the feature weights to achieve cross-scene knowledge transfer.
[0018] In step S1, the improved K-shape clustering algorithm determines the optimal number of clusters K by combining the silhouette coefficient and the elbow method, where K=3 corresponds to the three scenarios of industrial, commercial, and residential. In step S1, the time-series features include local trend features of the load curve extracted by CNN and long-term dependency features extracted by BiLSTM.
[0019] In step S2, the exponential stack mapping normalizes the original quantile output using the Softmax function and forces it to monotonically increase, specifically as follows: in, For the first The final output of the quantiles, For the Softmax function, This is the original output for the k-th quantile.
[0020] In step S2, the adaptive bandwidth kernel density estimation dynamically adjusts the bandwidth parameters based on scene stability, increasing the bandwidth by 20% under extreme weather conditions.
[0021] In step S3, the expression for the weighted maximum mean difference loss is: Where S and T are the source scene and target scene datasets, respectively. Let H be the feature mapping function, and H be the reproducing kernel Hilbert space.
[0022] In step S3, a near-end strategy optimization algorithm is used to dynamically adjust the feature weights under different scenarios.
[0023] The loss function of the quantile regression model used in step S2 is the weighted quantile loss: in, , This is the actual load value. To predict quantiles.
[0024] Taking data from a provincial power grid from 2018 to 2022 as an example, the following multi-source data was collected:
[0025] Data preprocessing includes: filling missing values using linear interpolation, and employing... Outliers are removed based on criteria, with a sampling frequency of 15 minutes per point.
[0026] Model training: 1. Multimodal scene recognition An improved K-shape clustering algorithm was adopted, with the number of clusters K=3, achieving a silhouette coefficient of 0.82 and a scene classification accuracy of 96%. Feature fusion employs a cross-attention mechanism, with a weight ratio of 7:3 for temporal features and meteorological features. Temporal feature extraction: Local trend features of the load curve are extracted using CNN, and long-term dependencies are extracted using BiLSTM; Meteorological feature weighting: The weights of factors such as temperature and humidity are dynamically adjusted using an attention mechanism, and the weight of air conditioning load is increased to 0.6 during hot weather.
[0027] 2. Dynamic Interval Prediction The quantile regression model uses a 6-layer Transformer architecture with a hidden dimension of 512. Training parameters: number of training epochs = 200, batch size = 64, learning rate = 0.001; Output quantile levels: 5%, 25%, 50%, 75%, 95%; The loss function uses weighted quantile loss: in, The quantile loss function, At the quantile level, This is the actual load value. To predict quantiles.
[0028] Quantile non-crossing constraint: Introduce exponential stack mapping to force quantiles to monotonically increase. in, For the Softmax function, For the raw output of the k-th quantile, ensure .
[0029] Uncertainty quantification: Adaptive bandwidth kernel density estimation is adopted, and the bandwidth is dynamically adjusted according to the stability of the scenario. Under extreme weather conditions, the bandwidth is increased by 20%.
[0030] 3. Cross-scene migration Shared feature extraction: A pre-trained Transformer encoder is used to extract common features across scenes, the underlying parameters are frozen, and only the scene classification head and prediction head are fine-tuned; Domain Adaptive Loss: Introducing a weighted maximum mean difference loss to reduce the distribution difference between the source and target scenes. Where S and T are the source scene and target scene datasets, respectively. Let H be the feature mapping function, and H be the reproducing kernel Hilbert space; Dynamic weight adjustment: The PPO algorithm is used to optimize the weight of scene features in real time. The weight of the production cycle feature in the industrial scene is set to 0.7, and the weight of the pedestrian flow feature in the commercial scene is set to 0.6.
[0031] To verify the effectiveness of the method of this invention, the prediction performance of the traditional LSTM model and the method of this invention were compared on the same dataset:
[0032] In terms of interval prediction performance, the method of this invention also shows significant advantages:
[0033] Regarding cross-scenario transfer efficiency, when a model trained in the source scenario (industrial) is transferred to a commercial scenario: WMMD loss decreased from 0.42 to 0.15; After 10 rounds of fine-tuning, the MAPE was 2.8%, while the traditional model required 50 rounds of retraining with a MAPE of 5.7%. The training time for the new scenario model is reduced by 60%, and the data requirement is reduced by 50%. The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.
Claims
1. A flexible load dynamic interval prediction method that adapts across scenarios, characterized in that, Includes the following steps: S1: Multimodal scene recognition: Collect historical load data, meteorological data and user tag data, perform preprocessing, fuse time-series features and meteorological features through cross-attention mechanism, and use an improved K-shape clustering algorithm to perform unsupervised scene clustering on multimodal features to automatically identify three types of scenes: industrial, commercial and residential. S2: Dynamic Interval Forecasting: Based on Transformer, a quantile regression model is constructed to output load forecast values at multiple quantile levels; an exponential stack mapping is introduced to impose a monotonically increasing constraint on the quantile output and solve the quantile crossover problem; An adaptive bandwidth kernel density estimation method is used to quantify the prediction uncertainty and generate confidence intervals; S3: Cross-scene adaptive transfer: Common features across scenes are extracted through a pre-trained Transformer encoder. Weighted maximum mean difference loss is introduced to reduce the feature distribution difference between the source scene and the target scene. Reinforcement learning is combined to dynamically adjust the feature weights to achieve cross-scene knowledge transfer.
2. The method for cross-scenario adaptive flexible load dynamic interval prediction according to claim 1, characterized in that: In step S1, the improved K-shape clustering algorithm determines the optimal number of clusters K by combining the contour coefficient and the elbow method, where K=3 corresponds to the three scenarios of industrial, commercial and residential.
3. The method for predicting flexible load dynamic intervals across scenarios according to claim 1, characterized in that: In step S1, the time-series features include local trend features of the load curve extracted by CNN and long-term dependency features extracted by BiLSTM.
4. The method for predicting flexible load dynamic intervals across scenarios according to claim 1, characterized in that: In step S2, the exponential stack mapping normalizes the original quantile output using the Softmax function and forces it to monotonically increase, specifically as follows: in, For the first The final output of the quantiles, For the Softmax function, This is the original output for the k-th quantile.
5. The method for predicting flexible load dynamic intervals across scenarios according to claim 1, characterized in that: In step S2, the adaptive bandwidth kernel density estimation dynamically adjusts the bandwidth parameters based on scene stability, increasing the bandwidth by 20% under extreme weather conditions.
6. The method for predicting flexible load dynamic intervals across scenarios according to claim 1, characterized in that: In step S3, the expression for the weighted maximum mean difference loss is: Where S and T are the source scene and target scene datasets, respectively, φ is the feature mapping function, and H is the regeneration kernel Hilbert space.
7. The method for predicting flexible load dynamic intervals across scenarios according to claim 1, characterized in that: In step S3, a near-end strategy optimization algorithm is used to dynamically adjust the feature weights under different scenarios.
8. The method for predicting flexible load dynamic intervals across scenarios according to claim 1, characterized in that: The loss function of the quantile regression model used in step S2 is the weighted quantile loss: in, , This is the actual load value. To predict quantiles.