Meteorological cumulative effect-based power load prediction method, apparatus and device

Abstract

The invention provides a power load prediction method, device and equipment based on a meteorological cumulative effect, and relates to the technical field of power load prediction. The method comprises the following steps: based on a meteorological cumulative effect extraction model, respectively carrying out meteorological feature extraction fusion on a meteorological sequence of a current day and meteorological sequences of a plurality of historical days to obtain meteorological cumulative effect characterization vectors of the current day and each historical day; wherein the meteorological sequence is a sequential sequence of multiple meteorological elements; calculating a similarity score between the current day and each historical day based on each meteorological cumulative effect representation vector and a trained similarity calculation model, and selecting a similar day from the plurality of historical days based on the similarity scores; and constructing a support set corresponding to the current day based on the meteorological cumulative effect representation vector of each similar day and the power load, and inputting the trained meta-learning prediction model to obtain a load prediction value of the prediction day. The method can improve the load prediction precision in a complex and changeable meteorological environment.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting technology, and in particular to a power load forecasting method, apparatus and equipment based on the cumulative effect of meteorological conditions. Background Technology

[0002] With the large-scale integration of uncertain elements such as distributed power sources and electric vehicles into the power grid, load uncertainty has increased dramatically, posing a more severe challenge to the safe and economical operation of the power grid. Load forecasting is a crucial component of the power grid dispatch center's energy management system, providing key information for dispatch decision-makers and effectively ensuring the safe and economical operation of the power grid. Meteorological factors are a significant influencing factor on load changes; selecting similar days based on meteorological factors can effectively reflect load similarity.

[0003] Current methods for forecasting load on similar days mainly include load forecasting based on similar days, load forecasting based on BP neural networks, and forecasting based on fuzzy clustering of similar days. However, the meteorological factors modeled in these methods are too simple, and they often use linear weighting or fixed window averaging to process meteorological factors. They cannot capture nonlinear lag effects and complex interactions between meteorological elements. Furthermore, traditional methods often use simple algorithms such as Euclidean distance and DTW to calculate similarity 1, which has high computational complexity and is difficult to process large-scale data, resulting in poor power load forecasting performance. Summary of the Invention

[0004] This invention provides a method, apparatus, and equipment for predicting power load based on the cumulative effect of meteorological conditions, in order to solve the problem of poor power load prediction performance.

[0005] In a first aspect, embodiments of the present invention provide a power load forecasting method based on meteorological cumulative effects, comprising: Based on the meteorological cumulative effect extraction model, meteorological features are extracted and fused from the meteorological sequence of the current day and the meteorological sequences of multiple historical days to obtain the meteorological cumulative effect representation vectors of the current day and each historical day; where the meteorological sequence is a time series sequence of multiple meteorological elements; Based on the cumulative meteorological effect representation vectors and the trained similarity calculation model, the similarity score between the current day and each historical day is calculated, and similar days are selected from multiple historical days based on the similarity score. Based on the cumulative meteorological effect representation vectors of various similar days and the power load, a support set corresponding to the current day is constructed and input into the trained meta-learning prediction model to obtain the load prediction value for the prediction day.

[0006] In one possible implementation, the meteorological cumulative effect extraction model includes a temporal convolutional network, a bidirectional long short-term memory network, and an attention module. Based on the meteorological cumulative effect extraction model, meteorological features are extracted and fused from the current day's meteorological sequence and the meteorological sequences of multiple historical days to obtain the meteorological cumulative effect representation vectors for the current day and each historical day, including: For each meteorological sequence, multi-scale meteorological features of the meteorological sequence are extracted based on a temporal convolutional network; For each multi-scale meteorological feature, the temporal dependencies of the multi-scale meteorological feature are extracted based on a bidirectional long short-term memory network to obtain the hidden state vector; For each hidden state vector, attention weighting is applied to the hidden state vector based on the attention module to obtain the meteorological cumulative effect representation vector for the current day and each historical day.

[0007] In one possible implementation, based on the various meteorological cumulative effect representation vectors and a trained similarity calculation model, a similarity score is calculated between the current day and each historical day, including: Based on the trained similarity calculation model, feature extraction and fusion are performed on the various meteorological cumulative effect representation vectors to obtain the meteorological embedding vectors for the current day and multiple historical days. Calculate the cosine similarity between the meteorological embedding vector of the current day and the meteorological embedding vector of the first historical day, and convert the cosine similarity into a similarity score between the current day and the first historical day using the inverse cosine function; where the first historical day is any historical day.

[0008] In one possible implementation, the similarity calculation model includes a gated temporal convolutional network, a Transformer encoder, and a graph attention layer. Based on the trained similarity calculation model, feature extraction and fusion are performed on the various meteorological cumulative effect representation vectors to obtain meteorological embedding vectors for the current day and multiple historical days, including: The local patterns and short-term dependencies of the first meteorological cumulative effect representation vector are extracted based on a gated temporal convolutional network to obtain the local feature vector of the first meteorological cumulative effect representation vector; wherein, the first meteorological cumulative effect representation vector is any meteorological cumulative effect representation vector. Based on the Transformer encoder, the long-range global dependency of the first meteorological cumulative effect representation vector is extracted, and the global feature vector of the first meteorological cumulative effect representation vector is obtained. Based on the graph attention layer, spatial context information features of the first meteorological cumulative effect representation vector are fused to obtain the spatial feature vector of the first meteorological cumulative effect representation vector. The local feature vector, global feature vector, and spatial feature vector are concatenated and then attention pooled to obtain the meteorological embedding vector of the first meteorological cumulative effect characterization vector.

[0009] In one possible implementation, before calculating the similarity score between the current day and each historical day based on the various meteorological cumulative effect representation vectors and the trained similarity calculation model, the following is also included: Anchor samples, positive samples, and negative samples are constructed based on the cumulative meteorological effect representation vector of historical days. The initial similarity calculation model is then trained based on the triplet loss function to obtain the trained similarity calculation model.

[0010] In one possible implementation, before constructing the support set corresponding to the current day based on the cumulative meteorological effect representation vectors of various similar days and the electricity load, and inputting it into the trained meta-learning prediction model to obtain the load forecast value for the prediction day, the following steps are also included: By using a sliding window, the cumulative meteorological effect representation vector of K consecutive historical days is extracted as the context, and the power load of the next day after K consecutive historical days is used as the prediction target. The initial meta-learning prediction model is then trained offline to obtain the trained meta-learning prediction model; where K is a preset value.

[0011] In one possible implementation, similar days are selected from multiple historical days based on similarity scores, including: Sort multiple historical days in descending order of similarity, and select the top K historical days as similar days.

[0012] Secondly, embodiments of the present invention provide a power load forecasting device based on meteorological cumulative effects, comprising: The extraction module is used to extract and fuse meteorological features from the current day's meteorological sequence and the meteorological sequences of multiple historical days based on the meteorological cumulative effect extraction model, so as to obtain the meteorological cumulative effect representation vectors of the current day and each historical day; wherein, the meteorological sequence is a time series sequence of multiple meteorological elements; The selection module is used to calculate the similarity score between the current day and each historical day based on the cumulative meteorological effect representation vectors and the trained similarity calculation model, and select similar days from multiple historical days based on the similarity scores. The prediction module is used to construct the support set corresponding to the current day based on the cumulative meteorological effect representation vector of each similar day and the power load, and input it into the trained meta-learning prediction model to obtain the load prediction value for the prediction day.

[0013] Thirdly, embodiments of the present invention provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect or any possible implementation thereof.

[0014] The power load forecasting method, apparatus, and equipment based on meteorological cumulative effects provided in this invention effectively capture the complex temporal dependencies and nonlinear cumulative effects of multiple meteorological elements through meteorological cumulative effect representation vectors extracted by multi-level neural networks. This overcomes the dependence of traditional methods on instantaneous meteorological states and provides a high-information-density feature foundation for similarity calculation. The similarity calculation model, integrating local, global, and spatial feature extraction, combined with metric learning training, can accurately quantify the similarity of meteorological patterns. This ensures that the selected similar days are highly consistent with the forecast days in terms of meteorological-load correlation, and the quality of the constructed support set is significantly better than methods based on simple Euclidean distance or calendar rules. A meta-learning prediction model is introduced, utilizing the high-quality support set for rapid task adaptation. This mechanism enables the model to learn load response patterns under specific current meteorological patterns from small samples, greatly enhancing its generalization ability to historically unseen or rare meteorological conditions. It effectively solves the prediction bottleneck of data-driven models in concept drift and cold-start scenarios, ultimately significantly improving load forecasting accuracy, assisting in grid optimization scheduling, reducing reserve capacity costs, and enhancing responsiveness to meteorological changes to ensure power supply stability. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the implementation of a power load forecasting method based on meteorological cumulative effects according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the implementation of a power load forecasting method based on meteorological cumulative effects, provided in another embodiment of the present invention. Figure 3 This is a comparison chart of the load forecasting performance of a power load forecasting method based on meteorological cumulative effects provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a power load forecasting device based on meteorological cumulative effects provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0016] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0017] See Figure 1 The document illustrates a flowchart of the implementation of the power load forecasting method based on meteorological cumulative effects provided in an embodiment of the present invention, which is described in detail below: Step 101: Based on the meteorological cumulative effect extraction model, meteorological features are extracted and fused from the meteorological sequence of the current day and the meteorological sequences of multiple historical days to obtain the meteorological cumulative effect representation vectors of the current day and each historical day; wherein, the meteorological sequence is a time series sequence of multiple meteorological elements.

[0018] In this embodiment, the existing meteorological similarity daily load prediction methods mainly include: (1) Load forecasting method based on similar days: The correlation between meteorological factors and load is analyzed by Pearson coefficient and maximum information coefficient. Meteorological factors that are strongly correlated with load are selected. Then, similar days are selected based on meteorological factors to form a meteorological similar day set. Finally, a TCN-LSTM forecasting model is constructed for forecasting.

[0019] (2) Load forecasting method based on BP neural network: The correlation between maximum power load and various meteorological factors is analyzed, and a forecasting model is established based on the BP neural network algorithm. This type of method considers the cumulative effect of meteorological factors on meteorological load (the study found that the correlation is strongest when the cumulative effect lasts for 2 days).

[0020] (3) Prediction method based on fuzzy clustering for similar days: meteorological factors are subdivided into temperature, air pressure, wind speed, cloudy and rainy conditions, which together with weekday type and date type constitute the influencing factors of similar days. A fuzzy coefficient feature mapping table is established through fuzzy rules, and fuzzy clustering method is used for classification.

[0021] These methods have the following shortcomings: (1) The modeling of meteorological cumulative effects is too simple: existing methods mostly use linear weighting or fixed window averaging (such as 2-day cumulative) to deal with meteorological cumulative effects, which cannot capture nonlinear lag effects and complex interactions between meteorological elements.

[0022] (2) The similarity calculation method is inefficient: Traditional methods often use simple algorithms such as Euclidean distance and DTW to calculate the similarity 1, which has high computational complexity and is difficult to process large-scale data.

[0023] (3) Insufficient generalization ability of machine learning models: Existing neural network models (such as BP neural network and TCN-LSTM) have poor generalization and stability in scenarios with sparse or volatile data, and lack the ability to adapt to small sample learning.

[0024] (4) Insufficient fusion of multi-source heterogeneous data: Existing methods ignore the cross-influence and spatiotemporal heterogeneity between meteorological elements and fail to effectively integrate multi-scale meteorological characteristics and historical load patterns.

[0025] Therefore, the problem this solution aims to solve is how to improve the accuracy and generalization ability of power load forecasting under extreme weather and small sample scenarios, specifically including: (1) How to more accurately model the nonlinear cumulative effects and lag effects of meteorological elements (2) How to efficiently calculate the similarity of consecutive multi-day meteorological sequences and capture their dynamic patterns (3) How to improve the adaptability of prediction models in scenarios with sparse or volatile data? (4) How to effectively integrate multi-source heterogeneous data (meteorological data, historical load, time factors, etc.) To address the aforementioned technical problems, this invention proposes a power load forecasting method based on meteorological cumulative effects and deep learning, the process of which is as follows: Figure 2 As shown, the main technological innovations include the following: (1) Dynamic weighting model of meteorological cumulative effect: A nonlinear cumulative model is constructed using LSTM-Attention, and a temporal convolutional network (TCN) is used to capture the multi-scale cumulative effect. The cross-term influence analysis between meteorological elements is also included. This part corresponds to step 101.

[0026] (2) Pattern matching algorithm for continuous time series: a meteorological sequence embedding space is constructed using deep metric learning, a contrastive loss function is used to optimize the similarity metric, and a time-aligned graph neural network is added for matching. This part corresponds to step 102.

[0027] (3) Multi-dimensional similarity composite evaluation mechanism: Construct a rapid adaptation framework based on Model-Agnostic Meta Learning (MAML), use a neural process model to quantify uncertainty, and add an interpretable AI module to provide decision-making basis. This part corresponds to step 103.

[0028] In this embodiment, before step 101, data acquisition and preprocessing are performed, specifically including the following steps: (1) Access the distributed meteorological monitoring network data stream and collect diverse meteorological elements such as temperature, humidity, wind speed, air pressure, and precipitation. (2) Obtain historical load data of the power grid (3) Obtain date type information (weekday / weekend / holiday) (4) Perform outlier detection and cleaning on the original data, and use time series interpolation to handle missing values. (5) Standardize the data for each dimension to eliminate the influence of units. Then, in step 101, the collected meteorological sequences and historical daily meteorological sequences are processed for cumulative meteorological effects, specifically including: (1) Extracting multi-scale meteorological features using a temporal convolutional network (TCN): A three-layer dilated convolutional structure was used, with dilation coefficients set to 1, 2, and 4 respectively, and the number of convolutional kernels increasing from 32 to 64 to 128. (2) Introducing a bidirectional long short-term memory network (BiLSTM) to capture temporal dependencies: The hidden layer dimension is set to 64, and the bidirectional structure allows the model to consider contextual information at the same time. (3) Dynamic weighting of integrated attention mechanism: The importance weights of each time step are calculated through a two-layer fully connected network and normalized using the softmax function. (4) The output is the weighted and fused meteorological cumulative effect feature vector. Step 102: Based on the cumulative meteorological effect representation vectors and the trained similarity calculation model, calculate the similarity score between the current day and each historical day, and select similar days from multiple historical days based on the similarity scores.

[0029] In this embodiment, the similarity calculation model can calculate the similarity of continuous sequences, and the specific implementation steps include: (1) Construct a deep metric learning network to map the meteorological cumulative effect representation vector to a low-dimensional embedding space. (2) Using a gated temporal convolutional network (GatedTCN) as the basic feature extractor to enhance nonlinear representation capabilities. (3) Introduce a graph attention network layer to model the spatial topological relationships between different meteorological stations. (4) Use a temporal pooling layer to aggregate sequence information and generate a fixed-length sequence representation vector. (5) Calculate the similarity between sequences using cosine similarity and output the similarity score. Step 103: Construct the support set corresponding to the current day based on the cumulative meteorological effect representation vector of each similar day and the power load, and input it into the trained meta-learning prediction model to obtain the load prediction value for the prediction day.

[0030] In this embodiment, the main contents of the meta-learning prediction model include: (1) Constructing a few-shot adaptation mechanism based on the meta-learning (MAML) framework (2) The support set contains recent historical data and is used for rapid model parameter adjustment. (3) The query set contains the meteorological data to be predicted and is used to generate the final load forecast. (4) Integrate uncertainty quantification components and output the confidence interval of the prediction results. This invention utilizes a multi-level neural network to extract meteorological cumulative effect representation vectors, effectively capturing the complex temporal dependencies and nonlinear cumulative effects of multiple meteorological elements. This overcomes the dependence of traditional methods on instantaneous meteorological states, providing a high-information-density feature foundation for similarity calculation. The integrated similarity calculation model, combining local, global, and spatial feature extraction with metric learning training, can accurately quantify the similarity of meteorological patterns. This ensures that the selected similar days are highly consistent with the forecast days in terms of meteorological-load correlation, and the quality of the constructed support set is significantly better than methods based on simple Euclidean distance or calendar rules. A meta-learning prediction model is introduced, utilizing the high-quality support set for rapid task adaptation. This mechanism enables the model to learn load response patterns under specific current meteorological patterns from small samples, greatly enhancing its generalization ability to historically unseen or rare meteorological conditions. It effectively solves the prediction bottleneck of data-driven models in concept drift and cold-start scenarios, ultimately significantly improving load forecast accuracy, assisting in grid optimization scheduling, reducing reserve capacity costs, and enhancing responsiveness to meteorological changes, thus ensuring power supply stability.

[0031] In one possible implementation, the meteorological cumulative effect extraction model includes a temporal convolutional network, a bidirectional long short-term memory network, and an attention module. Based on the meteorological cumulative effect extraction model, meteorological features are extracted and fused from the current day's meteorological sequence and the meteorological sequences of multiple historical days to obtain the meteorological cumulative effect representation vectors for the current day and each historical day, including: For each meteorological sequence, multi-scale meteorological features of the meteorological sequence are extracted based on a temporal convolutional network; For each multi-scale meteorological feature, the temporal dependencies of the multi-scale meteorological feature are extracted based on a bidirectional long short-term memory network to obtain the hidden state vector; For each hidden state vector, attention weighting is applied to the hidden state vector based on the attention module to obtain the meteorological cumulative effect representation vector for the current day and each historical day.

[0032] In this embodiment, the calculation process for the cumulative meteorological effect includes: (1) Input meteorological sequence data of length T, containing M meteorological elements.

[0033] (2) Multi-scale feature extraction using TCN network: The first layer uses 32 convolutional kernels with a kernel size of 3 and an expansion coefficient of 1; the second layer uses 64 convolutional kernels with an expansion coefficient of 2; and the third layer uses 128 convolutional kernels with an expansion coefficient of 4.

[0034] (3) The TCN output features are fed into a bidirectional LSTM network, and the sequences are processed in the forward and reverse directions respectively. The hidden state dimension is 64.

[0035] (4) Attention mechanism calculates the weights at each time step: the LSTM output is mapped to 64 dimensions through a fully connected layer, and then the attention score is generated through an unbiased linear layer.

[0036] (5) Use the softmax function to normalize the attention weights and perform weighted summation on the LSTM output to obtain the meteorological cumulative effect representation vector.

[0037] In one possible implementation, based on the various meteorological cumulative effect representation vectors and a trained similarity calculation model, a similarity score is calculated between the current day and each historical day, including: Based on the trained similarity calculation model, feature extraction and fusion are performed on the various meteorological cumulative effect representation vectors to obtain the meteorological embedding vectors for the current day and multiple historical days. Calculate the cosine similarity between the meteorological embedding vector of the current day and the meteorological embedding vector of the first historical day, and convert the cosine similarity into a similarity score between the current day and the first historical day using the inverse cosine function; where the first historical day is any historical day.

[0038] In this embodiment, the similarity score calculation process includes: (1) Input the current meteorological sequence and historical sequence into the encoder network with shared parameters.

[0039] (2) The encoder contains a GatedTCN layer: using a gating mechanism to control the information flow and enhance gradient propagation.

[0040] (3) Graph attention layer processes multi-site data: constructs a site relationship graph and aggregates neighbor node information through an attention mechanism.

[0041] (4) The temporal pooling layer uses adaptive max pooling to convert variable-length sequences into fixed-dimensional embedding vectors.

[0042] (5) Calculate the cosine similarity between the two embedding vectors and convert it into a similarity score (range 0-1) using the inverse cosine function. In one possible implementation, the similarity calculation model includes a gated temporal convolutional network, a Transformer encoder, and a graph attention layer. Based on the trained similarity calculation model, feature extraction and fusion are performed on the various meteorological cumulative effect representation vectors to obtain meteorological embedding vectors for the current day and multiple historical days, including: The local patterns and short-term dependencies of the first meteorological cumulative effect representation vector are extracted based on a gated temporal convolutional network to obtain the local feature vector of the first meteorological cumulative effect representation vector; wherein, the first meteorological cumulative effect representation vector is any meteorological cumulative effect representation vector. Based on the Transformer encoder, the long-range global dependency of the first meteorological cumulative effect representation vector is extracted, and the global feature vector of the first meteorological cumulative effect representation vector is obtained. Based on the graph attention layer, spatial context information features of the first meteorological cumulative effect representation vector are fused to obtain the spatial feature vector of the first meteorological cumulative effect representation vector. The local feature vector, global feature vector, and spatial feature vector are concatenated and then attention pooled to obtain the meteorological embedding vector of the first meteorological cumulative effect characterization vector.

[0043] In this embodiment, the similarity calculation model adopts a triple encoder hybrid architecture, which integrates the advantages of temporal convolution, self-attention and graph neural networks to simultaneously capture local dependencies, global dependencies and spatial dependencies in meteorological data.

[0044] (1) Input: Shape: (Batch_Size, Sequence_Length, Num_Features) Meaning: A batch of multivariate meteorological sequences. Num_Features includes temperature, humidity, wind speed, air pressure, precipitation, etc.

[0045] (2) Feature Extraction Backbone: The parallel processing architecture contains three sub-encoders, each processing different aspects of information: a) Local Feature Encoder (based on Gated TCN): Objective: To efficiently capture local patterns and short-term dependencies (such as diurnal temperature variation and short-term strong winds) in sequences.

[0046] Structure: 3-4 stacked temporal convolutional layers using gated activation units. Unlike the ReLU activation of standard TCN, the gating mechanism (similar to GRU gating) can dynamically control the information flow, enhance the model's ability to select important features, and improve gradient propagation.

[0047] Output: A high-level feature sequence that preserves the time dimension of the input (Batch_Size, Seq_Len, Hidden_Dim=64).

[0048] b) Globally dependent encoder (based on Transformer Encoder): Objective: To capture long-range global dependencies between any two time points in a sequence (such as the cumulative effect of a high temperature a week ago on the current load).

[0049] Structure: A simplified Transformer encoder layer (layers 1-2) containing a multi-head self-attention mechanism and a feedforward network. The self-attention mechanism allows the model to directly compute the relationship weights between all time steps in the sequence, regardless of their distance.

[0050] Output: A feature sequence (Batch_Size, Seq_Len, Hidden_Dim) containing global context information.

[0051] c) Spatial Relationship Encoder (based on Graph Attention Network - GAT): Objective: When input data comes from multiple geographically distributed weather stations, this encoder explicitly models the spatial topological relationships between stations (such as the mutual influence of neighboring stations).

[0052] Structure: The meteorological data at each time step is viewed as a "graph," where nodes are weather stations, node features are the meteorological data of each station, and edges represent the connections between stations (such as geographical distance and network topology). A two-layer graph attention network is used for processing, and information about neighboring nodes is aggregated according to attention weights.

[0053] Output: Each time step yields features (Batch_Size, Num_Stations, Seq_Len, Hidden_Dim) incorporating spatial context information, which can then be reconstructed back into sequence format.

[0054] (3) Feature Fusion: The output sequences of the three encoders are concatenated to form a powerful feature sequence (Batch_Size, Seq_Len, Fused_Dim) that integrates local, global, and spatial information.

[0055] (4) Temporal Aggregation Layer: Objective: To aggregate variable-length sequence features into a fixed-length vector representation.

[0056] Method: Attention pooling is used instead of simple average or max pooling. This layer learns an attention weight vector, assigning importance scores to each time step in the sequence, and then performing a weighted sum. This allows the model to focus on the most discriminative key time segments (such as periods of extreme weather).

[0057] Output: Fixed-size embedding vectors (Batch_Size, Embedding_Dim=128).

[0058] (5) Embedding Vector Output: The final output 128-dimensional vector is the "digital fingerprint" of the input meteorological sequence, which is directly sent to the subsequent similarity calculation module.

[0059] In one possible implementation, before calculating the similarity score between the current day and each historical day based on the various meteorological cumulative effect representation vectors and the trained similarity calculation model, the following is also included: Anchor samples, positive samples, and negative samples are constructed based on the cumulative meteorological effect representation vector of historical days. The initial similarity calculation model is then trained based on the triplet loss function to obtain the trained similarity calculation model.

[0060] In this embodiment, the similarity calculation model adopts the contrastive learning paradigm, and the training objective (loss function) is the triplet loss.

[0061] enter: Anchor (anchor sample): A meteorological sequence.

[0062] Positive samples: meteorological sequences similar to the Anchor, such as those from dates with similar historical load patterns.

[0063] Negative samples: meteorological sequences that are not similar to Anchor.

[0064] Objective: Through training, ensure that the distance (D_ap) between the Anchor and the Positive in the embedding space is much smaller than the distance (D_an) between the Anchor and the Negative. The loss function is: Loss = max(D_ap - D_an + margin, 0).

[0065] Effect: By minimizing this loss, the network learns to "bring similar sequences closer" and "push dissimilar sequences further apart," thereby forming an embedding space with good discriminative power.

[0066] In one possible implementation, before constructing the support set corresponding to the current day based on the cumulative meteorological effect representation vectors of various similar days and the electricity load, and inputting it into the trained meta-learning prediction model to obtain the load forecast value for the prediction day, the following steps are also included: By using a sliding window, the cumulative meteorological effect representation vector of K consecutive historical days is extracted as the context, and the power load of the next day after K consecutive historical days is used as the prediction target. The initial meta-learning prediction model is then trained offline to obtain the trained meta-learning prediction model; where K is a preset value.

[0067] In this embodiment, the meta-learning prediction process includes: (1) Construct a support set from historical data (usually containing data from the most recent 7-14 days).

[0068] (2) Use an inner loop in MAML to perform parameter adaptation on the support set in a small number of steps (usually 3-5 steps). (3) Based on the adapted parameters, perform forward calculations on the query set (date to be predicted). (4) Output load forecast and 95% confidence interval (5) It also provides interpretability analysis, showing the ranking of the contributions of each meteorological element. The specific design concept of this model is as follows: This model has learned a "rapid learning" ability by being trained on a large number of different weather models (tasks). When encountering a new weather model (such as a sudden cold wave), it can quickly adjust its parameters and make accurate predictions using only a small amount of recent data (support set).

[0069] Traditional deep learning models often perform poorly when encountering rare extreme weather patterns. Meta-learning, by teaching models "how to learn," greatly improves their generalization ability and robustness in small sample sizes and non-stationary environments.

[0070] Algorithm selection: Model-independent learning (MAML). Its biggest advantage is its independence from specific model architectures, allowing it to be combined with any feature extraction network (TCN, LSTM, Transformer). This algorithm finds initial model parameters that are sensitive to all tasks. When faced with a new task, it can achieve optimal performance by starting from this initial point and performing only a few gradient descent steps.

[0071] The training process of MAML consists of two loops: the inner loop and the outer loop.

[0072] Phase 1: Meta-Training - Conducted offline The goal at this stage is to find that "universal" initial parameter θ for the model.

[0073] Meta-training phase data construction (Task Construction): A task (Task T_i) simulates a complete prediction scenario: Support Set: Randomly selects data from K consecutive days (e.g., 7 days) in history as context.

[0074] Query Set: The data from the day following the support set, used as the prediction target.

[0075] By using a sliding window, thousands of such tasks can be constructed from historical data for training meta-learning models.

[0076] Phase Two: Meta-Testing / Inference - Online Prediction Once the model is deployed, the process for predicting tomorrow is as follows: (1) Constructing the Support Set: Obtain the real weather and load data for the most recent K days (consistent with K during training) D_support = {(x_t, y_t)}_{t=1}^K.

[0077] This K-day data defines a new task that the model may not have seen during training.

[0078] (2) Inner-Loop Adaptation: Load the pre-trained initial parameters θ* Using D_support as input, perform 1 to 5 steps of gradient descent (the exact same process as the meta-training phase) to obtain the adapted parameters φ for the latest weather model.

[0079] This step involves very little computation and is fast, enabling rapid online adaptation.

[0080] (3) Prediction: Get tomorrow's forecast weather data x_tomorrow.

[0081] The adapted parameters φ and x_tomorrow are used for forward propagation to obtain the final load forecast value y_hat_tomorrow = model(φ, x_tomorrow).

[0082] (4) Uncertainty Quantification: To achieve probabilistic prediction, we employ an ensemble approach.

[0083] During the meta-training phase, multiple meta-learning models (with different random seeds) are trained.

[0084] During prediction, steps (2) and (3) are run multiple times to obtain a set of predicted values ​​{y_hat_tomorrow_1, ...,y_hat_tomorrow_N}.

[0085] Calculate the mean of this set of predicted values ​​as the final point prediction, and calculate its standard deviation or quantile as the prediction uncertainty (confidence interval).

[0086] The final output of this module is a probability prediction distribution, typically represented as: Point Forecast: The expected (mean) final load.

[0087] Prediction Interval: For example, [y_hat_lower, y_hat_upper] represents a 95% probability that the actual load value will fall within this interval. This provides crucial risk information for grid dispatching. Details of data processing and feature engineering for the training samples of the meta-learning prediction model include: (1) Data standardization: The Min-Max standardization method is used to scale each feature dimension to the [0,1] interval.

[0088] (2) Feature selection: The importance of features is evaluated using the Pearson correlation coefficient and the maximum information coefficient, and meteorological elements with high correlation to load are retained.

[0089] (3) Sequence construction: Training samples are generated using a sliding window method, with the window size set to 7-14 days, and the prediction target is the load value on the first day after the window.

[0090] (4) Sample balancing: To address the problem of scarce extreme weather samples, oversampling and data augmentation techniques are used to expand the training set.

[0091] Model training and optimization strategies include: (1) Loss function design: combine mean squared error (MSE) and mean absolute error (MAE) to balance prediction accuracy and robustness.

[0092] (2) Optimization algorithm: The Adam optimizer is used, the initial learning rate is set to 0.001, and the learning rate decay strategy is applied.

[0093] (3) Regularization techniques: Use Dropout (ratio 0.2-0.5) and weight decay to prevent overfitting.

[0094] (4) Training strategy: phased training, first pre-training the feature extraction module, and then fine-tuning the entire system end-to-end.

[0095] (5) Early stopping mechanism: Monitor the performance of the validation set and terminate training early when the performance no longer improves.

[0096] In one possible implementation, similar days are selected from multiple historical days based on similarity scores, including: Sort multiple historical days in descending order of similarity, and select the top K historical days as similar days.

[0097] In this embodiment, the value of K can be set according to the actual situation.

[0098] In one specific embodiment, the method provided by the present invention is used to perform power load forecasting, resulting in Table 1 and... Figure 3 The comparison data shown is for the effects.

[0099]

[0100] As can be seen from the above embodiments, the method provided by the present invention brings the following technical effects: (1) Prediction accuracy is significantly improved The prediction error was significantly reduced due to the use of more accurate nonlinear cumulative effect modeling and deep metric learning.

[0101] (2) The accuracy of extreme weather forecasts has been greatly improved. Thanks to the meta-learning framework and the ability to adapt quickly to small samples, the model performs better in extreme weather conditions.

[0102] (3) Optimization of computational efficiency The application of deep metric learning and graph neural networks reduces computational complexity. Although the amount of computation increases, the improvement in prediction accuracy is more significant.

[0103] (4) Enhanced generalization ability The meta-learning framework enables the model to quickly adapt to new scenarios and sparse data situations. In sparse data scenarios, the prediction error is reduced by more than 35% compared to traditional methods.

[0104] (5) Improved interpretability Attention mechanisms and interpretable AI modules provide a basis for decision-making and can visualize the contribution of various meteorological elements.

[0105] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0106] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.

[0107] Figure 4 A schematic diagram of the power load forecasting device based on meteorological cumulative effects provided in an embodiment of the present invention is shown. For ease of explanation, only the parts related to the embodiment of the present invention are shown, and are described in detail below: like Figure 4 As shown, the power load forecasting device 4 based on meteorological cumulative effects includes: Extraction module 41 is used to extract and fuse meteorological features from the meteorological sequence of the current day and the meteorological sequences of multiple historical days based on the meteorological cumulative effect extraction model, so as to obtain the meteorological cumulative effect representation vector of the current day and each historical day; wherein, the meteorological sequence is a time series sequence of multiple meteorological elements; Module 42 is selected to calculate the similarity score between the current day and each historical day based on the cumulative meteorological effect representation vectors and the trained similarity calculation model, and to select similar days from multiple historical days based on the similarity scores. The prediction module 43 is used to construct the support set corresponding to the current day based on the meteorological cumulative effect representation vector of each similar day and the power load, and input it into the trained meta-learning prediction model to obtain the load prediction value for the prediction day.

[0108] In one possible implementation, the meteorological cumulative effect extraction model includes a temporal convolutional network, a bidirectional long short-term memory network, and an attention module; the extraction module 41 is specifically used for: For each meteorological sequence, multi-scale meteorological features of the meteorological sequence are extracted based on a temporal convolutional network; For each multi-scale meteorological feature, the temporal dependencies of the multi-scale meteorological feature are extracted based on a bidirectional long short-term memory network to obtain the hidden state vector; For each hidden state vector, attention weighting is applied to the hidden state vector based on the attention module to obtain the meteorological cumulative effect representation vector for the current day and each historical day.

[0109] In one possible implementation, module 42 is specifically used for: Based on the trained similarity calculation model, feature extraction and fusion are performed on the various meteorological cumulative effect representation vectors to obtain the meteorological embedding vectors for the current day and multiple historical days. Calculate the cosine similarity between the meteorological embedding vector of the current day and the meteorological embedding vector of the first historical day, and convert the cosine similarity into a similarity score between the current day and the first historical day using the inverse cosine function; where the first historical day is any historical day.

[0110] In one possible implementation, the similarity calculation model includes a gated temporal convolutional network, a Transformer encoder, and a graph attention layer; module 42 is specifically used for: The local patterns and short-term dependencies of the first meteorological cumulative effect representation vector are extracted based on a gated temporal convolutional network to obtain the local feature vector of the first meteorological cumulative effect representation vector; wherein, the first meteorological cumulative effect representation vector is any meteorological cumulative effect representation vector. Based on the Transformer encoder, the long-range global dependency of the first meteorological cumulative effect representation vector is extracted, and the global feature vector of the first meteorological cumulative effect representation vector is obtained. Based on the graph attention layer, spatial context information features of the first meteorological cumulative effect representation vector are fused to obtain the spatial feature vector of the first meteorological cumulative effect representation vector. The local feature vector, global feature vector, and spatial feature vector are concatenated and then attention pooled to obtain the meteorological embedding vector of the first meteorological cumulative effect characterization vector.

[0111] In one possible implementation, the selection module 42 is also used for: Before calculating the similarity score between the current day and each historical day based on the meteorological cumulative effect representation vectors and the trained similarity calculation model, anchor samples, positive samples, and negative samples are constructed based on the meteorological cumulative effect representation vectors of historical days. The initial similarity calculation model is then trained based on the triplet loss function to obtain the trained similarity calculation model.

[0112] In one possible implementation, the prediction module 43 is also used for: Before constructing the support set corresponding to the current day based on the cumulative meteorological effect representation vectors of various similar days and the power load, and inputting it into the trained meta-learning prediction model to obtain the load prediction value for the prediction day, the cumulative meteorological effect representation vectors of K consecutive historical days are extracted as context through a sliding window, and the power load of the next day after K consecutive historical days is taken as the prediction target. The initial meta-learning prediction model is then trained offline to obtain the trained meta-learning prediction model; where K is a preset value.

[0113] In one possible implementation, module 42 is specifically used for: Sort multiple historical days in descending order of similarity, and select the top K historical days as similar days.

[0114] This invention utilizes a multi-level neural network to extract meteorological cumulative effect representation vectors, effectively capturing the complex temporal dependencies and nonlinear cumulative effects of multiple meteorological elements. This overcomes the dependence of traditional methods on instantaneous meteorological states, providing a high-information-density feature foundation for similarity calculation. The integrated similarity calculation model, combining local, global, and spatial feature extraction with metric learning training, can accurately quantify the similarity of meteorological patterns. This ensures that the selected similar days are highly consistent with the forecast days in terms of meteorological-load correlation, and the quality of the constructed support set is significantly better than methods based on simple Euclidean distance or calendar rules. A meta-learning prediction model is introduced, utilizing the high-quality support set for rapid task adaptation. This mechanism enables the model to learn load response patterns under specific current meteorological patterns from small samples, greatly enhancing its generalization ability to historically unseen or rare meteorological conditions. It effectively solves the prediction bottleneck of data-driven models in concept drift and cold-start scenarios, ultimately significantly improving load forecast accuracy, assisting in grid optimization scheduling, reducing reserve capacity costs, and enhancing responsiveness to meteorological changes, thus ensuring power supply stability.

[0115] Figure 5 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. For example... Figure 5 As shown, the electronic device 5 of this embodiment includes a processor 50 and a memory 51. The memory 51 stores a computer program 52. When the processor 50 executes the computer program 52, it implements the steps in the various method embodiments described above. Alternatively, when the processor 50 executes the computer program 52, it implements the functions of each module / unit in the various device embodiments described above.

[0116] For example, computer program 52 may be divided into one or more modules / units, which are stored in memory 51 and executed by processor 50 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 52 in electronic device 5.

[0117] Electronic device 5 may include, but is not limited to, processor 50 and memory 51. Those skilled in the art will understand that... Figure 5 This is merely an example of electronic device 5 and does not constitute a limitation on electronic device 5. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device 5 may also include input / output devices, network access devices, buses, etc.

[0118] For the sake of simplicity and clarity, only the above-described functional modules / units are used as examples. In practical applications, the functions described above can be assigned to different functional modules / units as needed. These modules / units can be implemented in hardware, software, or a combination of both.

[0119] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0120] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A power load forecasting method based on meteorological cumulative effects, characterized in that, include: Based on the meteorological cumulative effect extraction model, meteorological features are extracted and fused from the meteorological sequence of the current day and the meteorological sequences of multiple historical days to obtain the meteorological cumulative effect representation vectors of the current day and each historical day; wherein, the meteorological sequence is a time series sequence of multiple meteorological elements; Based on the cumulative meteorological effect representation vectors and the trained similarity calculation model, the similarity score between the current day and each historical day is calculated, and similar days are selected from the multiple historical days based on the similarity score. Based on the cumulative meteorological effect representation vectors of various similar days and the power load, a support set corresponding to the current day is constructed and input into the trained meta-learning prediction model to obtain the load prediction value for the prediction day.

2. The power load forecasting method based on meteorological cumulative effects according to claim 1, characterized in that, The meteorological cumulative effect extraction model includes a temporal convolutional network, a bidirectional long short-term memory network, and an attention module. Based on this model, meteorological features are extracted and fused from the current day's meteorological sequence and multiple historical day meteorological sequences to obtain meteorological cumulative effect representation vectors for the current day and each historical day, including: For each meteorological sequence, multi-scale meteorological features of the meteorological sequence are extracted based on a temporal convolutional network; For each multi-scale meteorological feature, the temporal dependencies of the multi-scale meteorological feature are extracted based on a bidirectional long short-term memory network to obtain the hidden state vector; For each hidden state vector, attention weighting is applied to the hidden state vector based on the attention module to obtain the meteorological cumulative effect representation vector for the current day and each historical day.

3. The power load forecasting method based on meteorological cumulative effects according to claim 1, characterized in that, The similarity score between the current day and each historical day is calculated based on the cumulative meteorological effect representation vectors and the trained similarity calculation model, including: Based on the trained similarity calculation model, feature extraction and fusion are performed on each meteorological cumulative effect representation vector to obtain the meteorological embedding vectors of the current day and the multiple historical days. Calculate the cosine similarity between the meteorological embedding vector of the current day and the meteorological embedding vector of the first historical day, and convert the cosine similarity into a similarity score between the current day and the first historical day using an inverse cosine function; wherein, the first historical day is any historical day.

4. The power load forecasting method based on meteorological cumulative effects according to claim 2, characterized in that, The similarity calculation model includes a gated temporal convolutional network, a Transformer encoder, and a graph attention layer. Based on the trained similarity calculation model, feature extraction and fusion are performed on each meteorological cumulative effect representation vector to obtain meteorological embedding vectors for the current day and the multiple historical days, including: The local patterns and short-term dependencies of the first meteorological cumulative effect representation vector are extracted based on a gated temporal convolutional network to obtain the local feature vector of the first meteorological cumulative effect representation vector; wherein, the first meteorological cumulative effect representation vector is any meteorological cumulative effect representation vector. Based on the Transformer encoder, the long-range global dependency of the first meteorological cumulative effect representation vector is extracted to obtain the global feature vector of the first meteorological cumulative effect representation vector. Based on the graph attention layer, spatial context information features of the first meteorological cumulative effect representation vector are fused to obtain the spatial feature vector of the first meteorological cumulative effect representation vector. The local feature vector, the global feature vector, and the spatial feature vector are concatenated and then attention pooled to obtain the meteorological embedding vector of the first meteorological cumulative effect characterization vector.

5. The power load forecasting method based on meteorological cumulative effects according to claim 3, characterized in that, Before calculating the similarity score between the current day and each historical day based on the cumulative meteorological effect representation vectors and the trained similarity calculation model, the method further includes: Anchor samples, positive samples, and negative samples are constructed based on the cumulative meteorological effect representation vector of historical days. The initial similarity calculation model is then trained based on the triplet loss function to obtain the trained similarity calculation model.

6. The power load forecasting method based on meteorological cumulative effects according to claim 1, characterized in that, Before constructing the support set corresponding to the current day based on the cumulative effect representation vector of meteorological events on each similar day and the power load, and inputting it into the trained meta-learning prediction model to obtain the load forecast value for the prediction day, the following steps are also included: By using a sliding window, the cumulative meteorological effect representation vector of K consecutive historical days is extracted as the context, and the power load of the next day after the K consecutive historical days is used as the prediction target. The initial meta-learning prediction model is then trained offline to obtain the trained meta-learning prediction model; where K is a preset value.

7. The power load forecasting method based on meteorological cumulative effects according to claim 6, characterized in that, The selection of similar days from the plurality of historical days based on similarity scores includes: The historical days are sorted in descending order of similarity, and the top K historical days are selected as similar days.

8. A power load forecasting device based on meteorological cumulative effects, characterized in that, include: The extraction module is used to extract and fuse meteorological features from the meteorological sequence of the current day and the meteorological sequences of multiple historical days based on the meteorological cumulative effect extraction model, so as to obtain the meteorological cumulative effect representation vectors of the current day and each historical day; wherein, the meteorological sequence is a time series sequence of multiple meteorological elements; The selection module is used to calculate the similarity score between the current day and each historical day based on each meteorological cumulative effect representation vector and the trained similarity calculation model, and select similar days from the multiple historical days based on the similarity score; The prediction module is used to construct the support set corresponding to the current day based on the cumulative meteorological effect representation vector of each similar day and the power load, and input it into the trained meta-learning prediction model to obtain the load prediction value for the prediction day.

9. The power load forecasting device based on meteorological cumulative effect according to claim 8, characterized in that, The meteorological cumulative effect extraction model includes a temporal convolutional network, a bidirectional long short-term memory network, and an attention module; the extraction module is specifically used for: For each meteorological sequence, multi-scale meteorological features of the meteorological sequence are extracted based on a temporal convolutional network; For each multi-scale meteorological feature, the temporal dependencies of the multi-scale meteorological feature are extracted based on a bidirectional long short-term memory network to obtain the hidden state vector; For each hidden state vector, attention weighting is applied to the hidden state vector based on the attention module to obtain the meteorological cumulative effect representation vector for the current day and each historical day.

10. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 7.

Images