Power load prediction method based on multi-city meteorological feature fusion and festival embedding

Abstract

The invention discloses a power load prediction method based on multi-city meteorological feature fusion and festival embedding, and the method comprises the steps: collecting and preprocessing multi-source heterogeneous data, and achieving the deep feature extraction and fusion of the heterogeneous data through constructing a power load prediction model of multi-city meteorological collaborative fusion and multi-stage festival information embedding; and constructing a weighted composite loss function containing historical reconstruction loss and future prediction loss to train the model, and extracting a future time period load prediction value output by the model as a final prediction result. According to the invention, the power utilization mode difference of different festival stages can be accurately captured, and the festival and holiday prediction error is reduced; the spatial aggregation effect of the regional load is effectively mined, and the prediction robustness in extreme weather is improved; the calculation amount is reduced, and weather irrelevant features are also discarded; smoothness and consistency of hidden state expression are enhanced, and the basic reasoning ability of the model is guaranteed through a prediction task.

Description

Technical Field

[0001] This invention relates to the field of power system load forecasting technology, and in particular to a power load forecasting method based on the fusion of meteorological characteristics of multiple cities and the embedding of holidays. Background Technology

[0002] Power load forecasting is one of the core technologies for power grid dispatching and energy optimization, and its accuracy directly affects the economy and security of power grid operation. The main shortcomings of existing technologies include: simplistic handling of holiday factors: traditional methods only use 0 / 1 to mark statutory holidays, failing to effectively distinguish the characteristics of different stages such as before, during, and after holidays, and lacking refined coding for different holiday types such as weekends and work-short-day adjustments, leading to a significant increase in prediction errors during holidays; insufficient utilization of multi-city data: regional power grids typically contain load and meteorological data from multiple cities. Existing technologies often use simple splicing or independent modeling, failing to adaptively learn the spatial correlation and dynamic weights between different cities, limiting the model's ability to capture the overall trend of regional load; heterogeneous time series length matching problem: meteorological forecasts (such as temperature and irradiance) and load data differ in collection frequency and effective duration (e.g., meteorological data corresponding to the effective forecast duration usually does not affect load observation). Existing technologies often use a uniform alignment method, causing information loss or introducing noise; limited single-task learning objectives: traditional prediction models only focus on the prediction accuracy of future periods, neglecting the implicit feature reconstruction capability of historical periods, resulting in insufficient learning of time series dynamic characteristics. Therefore, it is necessary to propose a power load prediction method to improve the effect and accuracy of holiday power load prediction. Summary of the Invention

[0003] Purpose of the invention: The purpose of this invention is to provide a power load forecasting method based on the fusion of meteorological characteristics of multiple cities and the embedding of holidays, so as to improve the forecasting effect and accuracy of holiday power load.

[0004] Technical solution: The power load forecasting method based on multi-city meteorological feature fusion and holiday embedding described in this invention includes the following steps:

[0005] S1: Multi-source heterogeneous data acquisition and preprocessing;

[0006] S2: By constructing a power load forecasting model that integrates meteorological data from multiple cities and embeds multi-level holiday information, we can achieve deep feature extraction and fusion of heterogeneous data.

[0007] S3: Construct a weighted composite loss function that includes historical reconstruction loss and future prediction loss to train the model, and extract the future load forecast values ​​output by the model as the final prediction result.

[0008] Furthermore, the multi-source heterogeneous data in step S1 includes recent meteorological data, load data, and date-labeled data from multiple cities in the regional power grid.

[0009] Further, step S1 includes performing multi-level coupled classification coding on the date tag data based on holiday type, holiday period, and weekday; the holiday type coding includes weekdays, weekends, statutory holidays, and adjusted workdays; the holiday period coding includes weekdays, before holidays, during holidays, and after holidays; the weekday coding includes Monday to Sunday; the three codes are combined in the channel dimension to form date tag features.

[0010] Furthermore, step S1 also includes feature extraction of recent meteorological data, using the following extraction strategy:

[0011] For meteorological characteristics, extract the most recent Using 1 collection point, construct a time-series data training sample tensor:

[0012]

[0013]

[0014]

[0015] in, Indicates batch size; Indicates the length of the meteorological sequence; Indicates the load and classification coding sequence length, historical data. Data to be predicted at each time point At each time point, all data points to be predicted before being input into the model during training are set to 0. ; This represents weather data for multiple cities; dimension descriptions will be omitted in subsequent steps. Dimension.

[0016] Furthermore, the power load prediction model in step S2 includes an adaptive multi-city meteorological feature fusion module, a multi-scale and multi-level holiday feature embedding module, a feature enhancement and fusion module, a Transformer encoder module, and a downstream task output module.

[0017] Furthermore, the construction of the adaptive multi-city meteorological feature fusion module includes:

[0018] For each meteorological feature (e.g., temperature and humidity), perform cross-city dynamic weighted fusion, and Input of similar characteristics from cities The module generates dynamic weights for weather characteristics of each city through learnable linear transformations. ;right application Function along city dimension Normalization yields dynamic scores for each city. ; Calculate the feature representation after weighted fusion ; calculate separately After analyzing individual meteorological features, they are concatenated to obtain the dynamic weighted meteorological features. .

[0019] Furthermore, the construction of the multi-scale and multi-level festival feature embedding module includes:

[0020] The preprocessed sequence is then processed. Division: Settings The length and step size are used to divide the dynamically weighted meteorological features, load features, and codes; temporal compression coding is applied to the coding features; a one-dimensional max pooling layer is used to... The internal sequence is compressed into a scalar, and then... The layer is mapped to a high-dimensional space, resulting in a dimension of Embedded vectors; for meteorological and load characteristics, using Divide the time segment and transform it using a linear projection layer to generate segments with the same dimensions. The characteristics are represented.

[0021] Furthermore, the construction of the feature enhancement and fusion module includes:

[0022] Load target augmentation vector: For load data, holiday embeddings, official holiday period embeddings, and weekday embeddings are fused to generate an augmentation target vector. Meteorological features are concatenated with the load target enhancement vector to form the encoder input. Through a multi-level dynamic embedding mechanism, holiday types, vacation periods, and weekday information are deeply integrated into the load feature space.

[0023] The construction of the Transformer encoder module and the downstream task output module includes:

[0024] Furthermore, Input to layer The encoder, with each layer containing a multi-head self-attention mechanism, a feedforward neural network, and residual connections and normalization, outputs features after encoding. Input the downstream task output header and retrieve The end Dimensional features , After expansion, it undergoes a linear transformation to return to the predicted target. .

[0025] Furthermore, the weighted composite loss function in step S3 is expressed as:

[0026]

[0027]

[0028]

[0029] in, The mean square error for the historical period. The mean square error for the future forecast period; and Adjustable parameters

[0030] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0031] (1) Refined holiday modeling: Through a multi-level dynamic embedding mechanism, the holiday type, holiday period, and weekday information are deeply integrated into the load feature space through the holiday embedding module, so that the model can accurately capture the differences in electricity consumption patterns during different holiday periods and reduce the prediction error of holidays.

[0032] (2) Adaptive spatial fusion: The CityFusion module learns the nonlinear correlation between cities with the same meteorological characteristics, automatically identifies the dominant city and the following city, effectively explores the spatial clustering effect of regional load, and improves the prediction robustness under extreme weather conditions;

[0033] (3) Efficient alignment of heterogeneous time series: Based on the patch-based time series compression and selective splicing strategy, a non-aligned prediction method for meteorological and load characteristics is proposed without losing key meteorological information, which reduces the amount of computation and also eliminates meteorological irrelevant features;

[0034] (4) Multi-task joint optimization: By constraining the model’s learning ability on temporal dynamic characteristics through historical reconstruction tasks, the smoothness and consistency of hidden state expression are enhanced, and the basic reasoning ability of the model is guaranteed by prediction tasks.

[0035] (5) End-to-end scalable architecture: All components are differentiable modules, supporting gradient backhaul, facilitating access to larger-scale regional power grid data, and can be adapted to ultra-short-term, short-term and medium-to-long-term prediction tasks through hyperparameter adjustment. Attached Figure Description

[0036] Figure 1 This is a flowchart of the method of the present invention;

[0037] Figure 2 This is a schematic diagram of the specific structure of the present invention;

[0038] Figure 3This is a schematic diagram of the CityFusion module structure;

[0039] Figure 4 This is a schematic diagram of the module structure embedded for holidays. Detailed Implementation

[0040] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0041] like Figure 1 As shown, the power load forecasting method based on multi-city meteorological feature fusion and holiday embedding of the present invention includes the following steps:

[0042] S1: Multi-source heterogeneous data acquisition and preprocessing; specifically including the following:

[0043] The regional power grid load forecasting involves diverse and heterogeneous data, including M types of meteorological data (temperature, humidity, wind speed, etc.) from N cities, provincial load data, and date-labeled data.

[0044] Structured meteorological data and load data can be exported from the real-time power grid database; for date-labeled data, this invention innovatively designs a multi-level classification coding mechanism:

[0045] Holiday type coding: distinguishes between U-type tags such as weekdays, statutory holidays, and workdays with adjusted working hours;

[0046] Holiday phase coding: Refine the V-type time period labels such as weekdays, before holidays, during holidays, and after holidays;

[0047] Weekday encoding: Retains the periodic information of Monday to Sunday (w-class).

[0048] The meteorological feature extraction strategy is as follows: for meteorological features, only the most recent ones are extracted. One collection point to address the issue that the effective forecast duration of meteorological data is usually shorter than the load observation duration;

[0049] Constructing a training sample tensor for time-series data:

[0050]

[0051]

[0052]

[0053] in, Indicates batch size; Indicates the length of the meteorological sequence; Indicates the load and classification coding sequence length, historical data. Data to be predicted at each time point At each time point, before inputting the data into the model during training, all data points to be predicted are set to 0 to prevent information leakage. ; It represents weather data for various cities.

[0054] S2: By constructing a power load forecasting model that integrates meteorological data from multiple cities and embeds multi-level holiday information, we can achieve deep feature extraction and fusion of heterogeneous data.

[0055] like Figure 2 As shown, a power load forecasting model is constructed, which mainly includes the following modules: an adaptive multi-city meteorological feature fusion module; a multi-scale and multi-level holiday feature embedding module; a feature enhancement and fusion module; and a Transformer encoder and downstream task output module.

[0056] S2-1: Adaptive multi-city meteorological feature fusion module;

[0057] For each weather feature (such as temperature, humidity, etc.), a cross-city dynamic weighted fusion is performed. This invention proposes... Modules, such as Figure 3 As shown, its workflow is as follows:

[0058] Will Similar characteristics of cities, such as temperature enter The module generates dynamic weights for weather characteristics of each city through learnable linear transformations. ;right application Function along city dimension Normalization yields dynamic scores for each city. ; Calculate the feature representation after weighted fusion ; calculate separately After analyzing individual meteorological features, they are concatenated to obtain the dynamic weighted meteorological features. .

[0059] The specific calculation formula is as follows:

[0060]

[0061]

[0062]

[0063] in, This module can automatically identify leading and following cities, effectively explore the spatial clustering effect of regional loads, and improve the forecast robustness under extreme weather conditions.

[0064] S2-2: Multi-scale and multi-level festival feature embedding module;

[0065] The preprocessed sequence is then processed. Division: Settings The length and step size are used to divide the dynamically weighted meteorological features, load features, and codes; temporal compression embedding is applied to the coding features (holiday type, official holiday period, weekday), such as... Figure 4 As shown, the specific operation is as follows:

[0066] Using a one-dimensional max pooling layer The internal sequence is compressed into a scalar, and then... The layer is mapped to a high-dimensional space, resulting in a dimension of Embedded vectors; for meteorological and load characteristics, using Divide the time segment and transform it using a linear projection layer to generate segments with the same dimensions. The characteristics are represented.

[0067] The dimensional transformation method is as follows:

[0068]

[0069]

[0070]

[0071]

[0072] in, .

[0073] S2-3: Feature enhancement and fusion module;

[0074] Load target augmentation vector: For load data, holiday embeddings, official holiday period embeddings, and weekday embeddings are fused to generate an augmentation target vector. Meteorological features are concatenated with the load target enhancement vector to form the encoder input. .

[0075] The calculation formula is as follows:

[0076]

[0077]

[0078] in, By embedding holiday information at multiple levels, the types of holidays, holiday periods, and weekday information are deeply integrated into the load feature space, enabling the model to accurately capture the differences in electricity consumption patterns during different holiday periods.

[0079] S2-4: Transformer encoder and downstream task output module;

[0080] Will Input to layer The encoder, with each layer containing a multi-head self-attention mechanism; a feedforward neural network; residual connections and normalization; after encoding, it outputs features. Input downstream task output header: retrieve The end Dimensional features . After linear transformation, the expansion returns to the predicted target. .

[0081] S3: Construct a weighted composite loss function that includes historical reconstruction loss and future prediction loss to train the model, and extract the future load forecast values ​​output by the model as the final prediction result.

[0082] Construct a weighted composite loss function:

[0083]

[0084]

[0085]

[0086] in, The mean square error for the historical period. The mean square error for the future forecast period; and This is an adjustable parameter.

[0087] The training and testing process includes: calculating the mean squared error (MSE) for historical periods and the mean squared error for future predicted periods separately; combining the historical and future MSEs using a weighted method as the total loss during the training phase; and using only the future MSE as the evaluation metric during the validation phase. In the testing phase, the predicted load values ​​for future periods are directly selected as the final prediction result. The optimizer performs end-to-end training during the training phase. Stable model, in the validation phase Only keep To ensure the accuracy of inference, the load forecast value for the future period is directly selected as the final forecast result during the testing phase.

Claims

1. A power load forecasting method based on multi-city meteorological feature fusion and holiday embedding, characterized in that, Includes the following steps: S1: Multi-source heterogeneous data acquisition and preprocessing; S2: By constructing a power load forecasting model that integrates meteorological data from multiple cities and embeds multi-level holiday information, we can achieve deep feature extraction and fusion of heterogeneous data. S3: Construct a weighted composite loss function that includes historical reconstruction loss and future prediction loss to train the model, and extract the future load forecast values ​​output by the model as the final prediction result.

2. The power load forecasting method based on multi-city meteorological feature fusion and holiday embedding as described in claim 1, characterized in that, The multi-source heterogeneous data in step S1 includes recent meteorological data, load data, and date-labeled data from multiple cities in the regional power grid.

3. The power load forecasting method based on multi-city meteorological feature fusion and holiday embedding as described in claim 1, characterized in that, Step S1 includes performing multi-level coupled classification coding of date tag data based on holiday type, holiday period, and weekday; holiday type coding includes weekdays, weekends, statutory holidays, and adjusted workdays; holiday period coding includes weekdays, before holidays, during holidays, and after holidays; weekday coding includes Monday to Sunday; the three codes are combined in the channel dimension to form date tag features.

4. The power load forecasting method based on multi-city meteorological feature fusion and holiday embedding as described in claim 1, characterized in that, Step S1 further includes feature extraction from recent meteorological data, using the following extraction strategy: For meteorological characteristics, extract the most recent Using 1 collection point, construct a time-series data training sample tensor: ， ， ， in, Indicates batch size; Indicates the length of the meteorological sequence; Indicates the load and classification coding sequence length, historical data. Data to be predicted at each time point At each time point, all data points to be predicted before being input into the model during training are set to 0. ; This represents weather data for multiple cities; dimension descriptions will be omitted in subsequent steps. Dimension.

5. The power load forecasting method based on multi-city meteorological feature fusion and holiday embedding as described in claim 1, characterized in that, The power load prediction model in step S2 includes an adaptive multi-city meteorological feature fusion module, a multi-scale and multi-level holiday feature embedding module, a feature enhancement and fusion module, a Transformer encoder module, and a downstream task output module.

6. The power load forecasting method based on multi-city meteorological feature fusion and holiday embedding as described in claim 5, characterized in that, The construction of the adaptive multi-city meteorological feature fusion module includes: For each temperature and humidity meteorological characteristic, a cross-city dynamic weighted fusion is performed, and Input of similar characteristics from cities The module generates dynamic weights for weather characteristics of each city through learnable linear transformations. ;right application Function along city dimension Normalization yields dynamic scores for each city. ; Calculate the feature representation after weighted fusion ; calculate separately After analyzing individual meteorological features, they are concatenated to obtain the dynamic weighted meteorological features. .

7. The power load forecasting method based on multi-city meteorological feature fusion and holiday embedding as described in claim 5, characterized in that, The construction of the multi-scale and multi-level festival feature embedding module includes: The preprocessed sequence is then processed. Division: Settings The length and step size are used to divide the dynamically weighted meteorological features, load features, and codes; temporal compression coding is applied to the coding features; a one-dimensional max pooling layer is used to... The internal sequence is compressed into a scalar, and then... The layer is mapped to a high-dimensional space, resulting in a dimension of Embedded vectors; for meteorological and load characteristics, using Divide the time segment and transform it using a linear projection layer to generate segments with the same dimensions. The characteristics are represented.

8. The power load forecasting method based on multi-city meteorological feature fusion and holiday embedding as described in claim 5, characterized in that, The construction of the feature enhancement and fusion module includes: Load target augmentation vector: For load data, holiday embeddings, official holiday period embeddings, and weekday embeddings are fused to generate an augmentation target vector. Meteorological features are concatenated with the load target enhancement vector to form the encoder input. Through a multi-level dynamic embedding mechanism, holiday types, vacation periods, and weekday information are deeply integrated into the load feature space.

9. The power load forecasting method based on multi-city meteorological feature fusion and holiday embedding as described in claim 5, characterized in that, The construction of the Transformer encoder module and the downstream task output module includes: Will Input to layer The encoder, with each layer containing a multi-head self-attention mechanism, a feedforward neural network, and residual connections and normalization, outputs features after encoding. Input the downstream task output header and retrieve The end Dimensional features , After linear transformation, the expansion returns to the predicted target. .

10. The power load forecasting method based on multi-city meteorological feature fusion and holiday embedding according to claim 1, characterized in that, The weighted composite loss function in step S3 is expressed as follows: ， ， ， in, The mean square error for the historical period. The mean square error for the future forecast period; and This is an adjustable parameter.

Images