Power load probability prediction method, system and device based on full map weather grid and similar day prior dynamic correction and medium

Abstract

The application discloses a kind of power load probability prediction method, system, equipment and medium based on full map weather grid and similar day prior dynamic correction, it is related to power load prediction technical field, the application obtains multiple-source input data first, according to the historical similar day search result in multiple-source input data, constructs similar day prior feature tensor, and carries out full map weather feature processing, obtains exogenous weather feature sequence;Baseline load prediction result is obtained by using time series main stem prediction network prediction, then similar day dynamic correction is carried out by gate network, finally the point prediction value after correction is input quantile prediction head, and the prediction value and prediction interval of power load prediction are obtained.The application increases dynamic correction module on the basis of time series main stem prediction network, converts similar day prior feature into rectification direction, considers experience prior and deep main stem prediction ability, and improves the robustness of load prediction.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting technology, specifically to a power load probability forecasting method, system, equipment, and medium based on full-map meteorological grids and similar day prior dynamic correction. Background Technology

[0002] With the high proportion of renewable energy integration, the expansion of electricity market transactions, and the frequent occurrence of extreme weather events, the volatility and uncertainty of electricity load have significantly increased. Power grid companies and electricity market participants have placed higher demands on the accuracy, stability, and interpretability of short-term, day-ahead, and ultra-short-term load forecasts. Existing load forecasting technologies can be broadly categorized as follows: 1) Statistical methods, such as autoregressive models, moving average models, and Kalman filter models. These methods are simple in structure, but have limited adaptability to complex nonlinear relationships, multi-source heterogeneous data, and sudden weather changes.

[0003] 2) Methods based on traditional machine learning, such as support vector machines, random forests, and gradient boosting trees. These methods can utilize multi-dimensional features, but they are difficult to fully characterize the long-term dependence of power load and the spatial distribution information of meteorological fields.

[0004] 3) Deep learning-based methods, such as recurrent neural networks, long short-term memory networks, convolutional neural networks, and Transformer-like models. These methods have significant advantages in complex temporal modeling and are gradually becoming mainstream technologies.

[0005] However, the above-mentioned load forecasting techniques have at least the following problems: 1) The use of meteorological information is relatively crude. Many methods only use scalar features such as temperature and humidity from a small number of stations, failing to effectively utilize the spatial structural features of the meteorological grid across the entire region, resulting in insufficient characterization of the impact of meteorological disturbances on the load.

[0006] 2) Although the similar day method is widely used in the power industry, most schemes directly use the similar day load curve as the prediction result or fixed weight baseline, lacking dynamic assessment of the reliability of the similar day, and making it difficult to adaptively adjust in scenarios of abnormal weather, holiday disturbances or load mechanism changes.

[0007] 3) There is a lack of stable and controllable coupling mechanism between the deep learning backbone prediction model and the prior knowledge of similar days. If the input is directly spliced ​​together or the prior knowledge is forcibly superimposed, it is easy to cause unstable model training, aggravated error propagation, and even the prior knowledge may mislead the main model.

[0008] 4) Existing point forecasting methods are difficult to output forecast intervals at different confidence levels simultaneously, and cannot directly support the quantification of power market risks and the verification of dispatch security. Summary of the Invention

[0009] To address one or more shortcomings of the existing technologies, this invention provides a method, system, device, and medium for probabilistic prediction of power load based on full-map meteorological grids and prior dynamic correction of similar days. It comprehensively utilizes historical load time series, time stamp information, full-map meteorological grid information, and historical prior curves of similar days to jointly predict the load value and its quantile interval for the future prediction period, thereby improving prediction accuracy and interval prediction capability.

[0010] To achieve the above objectives, the present invention adopts one or more of the following technical solutions: Firstly, a probabilistic power load prediction method based on a full-map meteorological grid and similar-day prior dynamic correction is provided for predicting the power load of a target area within a future prediction window, wherein the future prediction window is a preset prediction time interval. The method includes the following steps: The system acquires multi-source input data, including historical power load sequences, time stamp information, meteorological raster data of the target area, and historical similar day retrieval results. The historical power load sequence consists of historical load values ​​of the target area arranged in chronological order. The time stamp information includes time period features corresponding to each sampling time of the historical power load sequence and time period features corresponding to each prediction time within the future prediction window. The meteorological raster data of the target area consists of multivariate meteorological time-series data with a spatial raster structure within the target area. The historical similar day retrieval results include load curves and similarity scores corresponding to multiple historical similar days. The historical similar days are similar samples retrieved based on the fusion vector of meteorological and time features of the future prediction window, and the similarity score is the cosine similarity between the fusion vector of the future prediction window and the fusion vector of the corresponding historical similar day. Based on the historical similar day retrieval results, a similar day prior feature tensor is constructed; The meteorological raster sequence corresponding to the future prediction window is extracted from the meteorological raster data of the target area. A meteorological convolutional coding network is used to perform convolutional coding on the meteorological raster frames at each time step in the meteorological raster sequence. The encoded meteorological feature vectors are arranged in chronological order to generate the exogenous meteorological feature sequence corresponding to the future prediction window. The historical power load sequence, the time stamp information, and the exogenous meteorological feature sequence are input into the time-series backbone prediction network to obtain the baseline load prediction result. The baseline load prediction result includes the point prediction value of the target area at each prediction time within the future prediction window. The time-series backbone prediction network is a deep learning network with time-series modeling capabilities, used to learn the nonlinear relationship between historical load and meteorological features. Based on the prior feature tensor of similar days and the baseline load prediction results, the dynamic correction module generates the corrected point prediction values ​​for each prediction time. The dynamic correction module generates dynamic correction coefficients through a gating network based on the prior feature tensor of similar days and the baseline load prediction results, and performs residual correction on the baseline load prediction results using the dynamic correction coefficients. The corrected point prediction values ​​at each prediction time are input into the quantile prediction head to obtain the power load probability prediction result. The power load probability prediction result includes the load prediction interval corresponding to each prediction time within the future prediction window. The quantile prediction head is obtained through supervised training with the quantile loss function and is used to output the load prediction values ​​of multiple quantile points to form the load prediction interval.

[0011] Preferably, the step of constructing a priori feature tensor for similar days based on the historical similar day retrieval results specifically includes: The similarity scores are subjected to exponential normalization to obtain the weights; Based on the weights, the multiple similar daily load curves are weighted and summed to obtain the weighted similar daily prior mean curve. The weighted similar day prior mean curve is concatenated with the multiple similar day load curves according to time steps to generate the similar day prior feature tensor.

[0012] Preferably, the dynamic correction module generates dynamic correction coefficients based on the prior feature tensor of similar days and the baseline load prediction result through a gating network, and performs residual correction on the baseline load prediction result using the dynamic correction coefficients, specifically including: Based on the prior feature tensor of similar days and the baseline load prediction result, a prior mean term, a prediction bias term, and a prior dispersion term are calculated; wherein, the prior mean term is the weighted prior mean curve of similar days, the prediction bias term is the difference between the prior mean term and the baseline load prediction result, and the prior dispersion term is used to characterize the degree of dispersion of each similar day load curve in the prior feature tensor of similar days at the corresponding time step; The prior mean, prediction bias, and prior dispersion terms are concatenated to construct the gating network input vector; The input vector of the gated network is input into the gated network to obtain the dynamic correction coefficients; The dynamic correction coefficient is used to perform residual correction on the baseline load prediction results to obtain the corrected point prediction values ​​at each prediction time.

[0013] Preferably, the step of using the dynamic correction coefficient to perform residual correction on the baseline load prediction result to obtain the corrected point prediction values ​​at each prediction time specifically involves: The product of the dynamic correction coefficient and the prediction deviation term is added to the baseline load prediction result to obtain the corrected point prediction value at each prediction time.

[0014] Preferably, the method further includes a step of training a prediction model for performing the power load probability prediction method, wherein the training employs a two-stage fine-tuning strategy: The first stage is the warm-up and fine-tuning stage, in which the meteorological convolutional coding network and the temporal backbone prediction network are frozen, and only the dynamic correction module and the quantile prediction head are trained. The second stage is the joint fine-tuning stage, in which all network parameters are unfrozen, and the meteorological convolutional coding network and the temporal backbone prediction network are fine-tuned using a first learning rate, while the dynamic correction module and the quantile prediction head are further optimized using a second learning rate; wherein, the second learning rate is greater than the first learning rate.

[0015] Preferably, the corrected point prediction values ​​for each prediction time are input into the quantile prediction head to obtain the power load probability prediction result, specifically including: For each prediction time step within the future prediction window, the corrected point prediction values ​​at each prediction time are used as input, and the load prediction values ​​at each quantile are calculated through linear mapping. The quantile prediction head is trained under supervision using a quantile loss function to determine the load prediction value corresponding to the cumulative probability at each quantile based on the point prediction value at each prediction time after correction. The quantile loss function is expressed as:

[0016] Where, q m y is a quantile; t true This represents the actual load value; y t,qm For quantile q m The predicted value below; According to the order of probability values ​​of quantiles from low to high, the load prediction values ​​corresponding to each quantile are respectively assigned to the load prediction results of the low quantile, median and high quantile.

[0017] Preferably, the similar date search results are obtained in the following ways: Historical weather windows are encoded and fused with time cycle features to construct a similar day retrieval database; After encoding the meteorological window corresponding to the future prediction window and fusing it with the time cycle features, the top K historical samples with the highest similarity are searched in the similar day retrieval library as similar samples, and the corresponding historical load curves and similarity scores are generated as the historical similar day retrieval results.

[0018] Secondly, a power load probability prediction system based on full-map meteorological grids and prior dynamic correction of similar days is provided, specifically including: The data acquisition module is used to acquire multi-source input data, including historical power load sequences, time stamp information, meteorological raster data of the target area, and historical similar day retrieval results. The historical power load sequence is the historical load value of the target area arranged in chronological order. The time stamp information includes the time period characteristics corresponding to each sampling time of the historical power load sequence and the time period characteristics corresponding to each prediction time within the future prediction window. The meteorological raster data of the target area is multivariate meteorological time series data with a spatial raster structure within the target area. The historical similar day retrieval results include load curves and similarity scores corresponding to multiple historical similar days. The historical similar days are similar samples retrieved based on the fusion vector of meteorological and time features of the future prediction window. The similarity score is the cosine similarity between the fusion vector of the future prediction window and the fusion vector of the corresponding historical similar day. The prior feature construction module is used to construct a prior feature tensor for similar days based on the historical similar day retrieval results. The meteorological feature processing module is used to extract the meteorological grid sequence corresponding to the future prediction window from the meteorological grid data of the target area, use a meteorological convolutional coding network to perform convolutional coding on the meteorological grid frames at each time step in the meteorological grid sequence, and arrange the encoded meteorological feature vectors in chronological order to generate the exogenous meteorological feature sequence corresponding to the future prediction window. The baseline load forecasting module is used to input the historical power load sequence, the time stamp information, and the exogenous meteorological feature sequence into the time-series backbone forecasting network to obtain the baseline load forecasting result. The baseline load forecasting result includes the point forecast values ​​of the target area at each forecast time within the future forecast window. The time-series backbone forecasting network is a deep learning network with time-series modeling capabilities, used to learn the nonlinear relationship between historical load and meteorological features. The dynamic correction module generates corrected point prediction values ​​for each prediction time based on the prior feature tensor of similar days and the baseline load prediction results; the dynamic correction module generates dynamic correction coefficients through a gating network based on the prior feature tensor of similar days and the baseline load prediction results, and performs residual correction on the baseline load prediction results using the dynamic correction coefficients. The load probability prediction module inputs the corrected point prediction values ​​at each prediction time into the quantile prediction head to obtain the power load probability prediction result. The power load probability prediction result includes the load prediction interval corresponding to each prediction time within the future prediction window. The quantile prediction head is obtained through supervised training with a quantile loss function and is used to output the load prediction values ​​of multiple quantile points to form the load prediction interval.

[0019] On the other hand, the present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, it can implement the steps of the power load probability prediction method based on full-map meteorological grid and similar day prior dynamic correction as described above.

[0020] On the other hand, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it can implement the steps of the power load probability prediction method based on full-map meteorological grid and similar day prior dynamic correction as described above.

[0021] By adopting the above technical solution, the beneficial effects of the present invention are as follows: 1. This invention utilizes full-map meteorological raster data to generate exogenous meteorological feature sequences, making full use of the spatial structure information of the full-map meteorological raster. Compared with methods that only use a small number of meteorological scalar features, it can more accurately characterize the impact of meteorological disturbances on load changes. Furthermore, it uses a time-series backbone prediction network to obtain the baseline load prediction results, and then uses the constructed similar day prior feature tensor for dynamic correction. Unlike traditional techniques that directly replace the deep model output with similar day results as a hard baseline, this invention adds a similar day dynamic correction module to the time-series backbone prediction network, transforming the similar day prior features into "correction directions," taking into account both empirical priors and deep backbone prediction capabilities, and improving the robustness of load prediction.

[0022] 2. The power load forecasting method of this invention can directly output multiple accurate quantile forecast values, meeting the needs of power market bidding, power grid security verification, and uncertainty analysis for interval forecasting.

[0023] 3. This invention constructs a gated network input vector by simultaneously introducing prior mean, prediction bias and prior dispersion, thereby achieving dynamic perception and adaptive correction of reliability on similar days and improving adaptability to different operating conditions.

[0024] 4. This invention obtains a better combination of structure and training parameters between the full-map meteorological convolutional feature extraction module and the temporal backbone network through an automatic hyperparameter optimization mechanism, providing a higher quality basic backbone model for subsequent similar day dynamic correction; and by performing gradient blocking on the bias term, it reduces the adverse gradient interference of the newly added similar day dynamic correction module on the temporal backbone prediction network, thereby improving the stability of joint training. Attached Figure Description

[0025] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0026] Figure 1 This is a schematic diagram of the power load probability prediction method based on full-map meteorological grid and similar day prior dynamic correction in one or more embodiments of the present invention; Figure 2 This is a schematic diagram of the overall structure of the prediction model used in one or more embodiments of the present invention; Figure 3 This is a schematic diagram illustrating the process of generating similar day prior feature tensors in one or more embodiments of the present invention; Figure 4 This is a schematic diagram of the dynamic correction principle of the gating network in one or more embodiments of the present invention; Figure 5 This is a schematic diagram of an ablation experiment using the prediction method in one or more embodiments of the present invention. Detailed Implementation

[0027] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0028] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0029] Example 1 In one typical embodiment of this application, a power load probability prediction method based on a full-map meteorological grid and similar-day prior dynamic correction is provided. This method is used to predict the power load of a target area within a future prediction window, where the future prediction window is a preset prediction time interval. (Refer to...) Figures 1-5 The method is executed through a prediction model, which is a TimeXer load probability prediction model that integrates meteorological data and similar day data. Its structure includes: The full-map meteorological convolutional feature extraction module is used to encode the meteorological raster sequence within the future prediction window into a meteorological exogenous feature sequence; The TimeXer time series backbone network is used to output benchmark forecasts of future loads based on historical loads and exogenous meteorological characteristics. A similar-day dynamic gating network is used to generate dynamic correction coefficients based on the prior mean, prediction bias, and prior dispersion of similar days. The quantile prediction header is used to output results for multiple quantiles.

[0030] The specific steps of the power load forecasting method are as follows: Step S1: Obtain multi-source input data, which includes historical power load sequences, time stamp information, meteorological raster data of the target area, and historical similar day search results.

[0031] Specifically, the following data is collected and organized: historical power load sequences, which are historical load values ​​arranged in chronological order within the target area; time stamp information corresponding to historical load periods, specifically including time period characteristics corresponding to each sampling time in the historical power load sequence, and time period characteristics corresponding to each prediction time within the future prediction window; meteorological raster data of the target area, which are multivariate meteorological time series data with spatial raster structure within the target area, which may include one or more of temperature, humidity, wind speed, air pressure, precipitation, and irradiance; exogenous time stamp information corresponding to the future prediction period; and historical similar day retrieval results, including load curves and similarity scores corresponding to multiple historical similar days. Among them, historical similar days are similar samples retrieved based on the fusion vector of meteorological and time features of the future prediction window, and the similarity score is the cosine similarity between the fusion vector of the future prediction window and the fusion vector of the corresponding historical similar day.

[0032] Specifically, the historical similar day retrieval results include the load curves and similarity scores corresponding to multiple historical similar days. The historical similar days are similar samples retrieved based on the fusion vector of meteorological and temporal features of the future prediction window. The similarity score is the cosine similarity between the fusion vector of the future prediction window and the fusion vector of the corresponding historical similar day.

[0033] In this embodiment, after encoding the meteorological window corresponding to the time to be predicted, the top K historical samples with the highest similarity are searched in the similar day retrieval database as similar samples, thus obtaining the Top-K historical similar days. Corresponding historical load curves and similarity scores are then generated for the Top-K historical similar days, resulting in the Top-K similar day load curves and their corresponding similarity scores. Specifically, this includes: The basic dataset was collected and organized, including four core data categories: historical power load sequences, historical load matching timestamp sequences, regional full-map meteorological raster sequences, and target start time information corresponding to future forecast periods. Among them, the historical power load sequences and meteorological raster sequences are long-term time-series monitoring data, which can be exported in batches from the power grid dispatch database and the regional meteorological numerical forecasting system.

[0034] Basic attributes are defined for the various types of data collected. The meteorological raster sequence includes at least one meteorological variable channel from temperature, humidity, air pressure, wind speed, precipitation, and irradiance. In a preferred embodiment, the meteorological raster sequence is a multidimensional variable. The meteorological sequence time granularity supports fixed step sizes such as 1 hour and 15 minutes, while the load sequence time granularity supports fixed step sizes such as 15 minutes, 30 minutes, and 1 hour, which can be flexibly switched according to short-term or ultra-short-term load forecasting needs of the power grid. The length of the future forecast period is denoted as pred_len. In this embodiment, pred_len is preferably 96, suitable for 24-hour load forecasting scenarios with a 15-minute granularity.

[0035] Using the target start time of the future forecast period as the global alignment anchor point, based on the time granularity of the meteorological sequence, the system outputs the raw data of the standardized meteorological process window, including the historical meteorological process window of the preset length or the meteorological window of the forecast sample, from the subsequent continuous meteorological data sequence according to the time alignment rules.

[0036] Among them, the time alignment rules include: when the time granularity of the meteorological sequence is completely consistent with the time granularity of the load sequence, the precise alignment method is adopted, and the target start time is directly used as the starting point to synchronously extract continuous meteorological grids of equal duration; When the time granularity of the meteorological series is inconsistent with that of the load series, the starting time of the load window is used as the reference. The most recent meteorological sampling point on the meteorological time axis, no later than this reference, is selected as the starting point of the meteorological window. From this starting point, a continuous meteorological raster sequence that can completely cover the date to be predicted is extracted to complete the time alignment between the meteorological window and the load forecast window. The time granularity of the meteorological series is the data frequency, generally 15 minutes or 1 hour.

[0037] After differential preprocessing of the original standardized meteorological process window data output earlier, it is input into a pre-trained spatiotemporal autoencoder network to extract the meteorological latent vector.

[0038] In this study, a ConvLSTM-AE convolutional recurrent autoencoder was selected as the original spatiotemporal autoencoder network with historical meteorological process windows as input. The spatiotemporal features of the historical meteorological process windows were extracted and compressed into latent vectors at the encoding end. The input historical meteorological process windows were reconstructed at the decoding end. The initial autoencoder network was trained with the goal of minimizing the reconstruction loss function to obtain a pre-trained spatiotemporal autoencoder network.

[0039] Using the pre-trained spatiotemporal autoencoder network described above, we obtain: Let the first The meteorological window corresponding to each historical sample is The encoded historical meteorological potential vector is represented as:

[0040] For the weather query window corresponding to the future forecast period The query meteorological potential vector is obtained. :

[0041] Obtain the time period vector of the future period to be predicted: Multi-dimensional periodic features are extracted from the start timestamps of the period to be predicted and historical samples to construct a time period vector. Optional periodic features include: day of week, day of month, day of year, week of year, and month ID. Any one or more of these features can be selected individually or in combination.

[0042] The above numerical periodic features are normalized to their maximum and minimum values ​​and mapped to the standard interval [0,1]. Then, a periodic mapping function Cyc( ) is used, which includes both sine and cosine cyclic mapping. The scaled periodic features are periodically encoded to eliminate abrupt changes in time boundaries and generate a standardized time period vector, including the historical sample time period vector t. i With the query sample time period vector t q , respectively represented as:

[0043]

[0044] in, Represents a periodic mapping function. .

[0045] The potential meteorological vector of the query sample (i.e., the period to be predicted) is weighted and fused with the time period vector to generate a query feature vector for similar day retrieval, specifically including: Historical meteorological potential vector z i With querying the potential meteorological vector z q L2 normalization is performed on the historical sample time period vector t. i With the query sample time period vector t q Perform L2 normalization separately.

[0046] A time weighting coefficient α is introduced to weight the normalized time period vector, which is used to control the contribution ratio of the time period in the similarity measurement. In this embodiment, α=1.22 is preferred to amplify the weight of time period similarity and avoid retrieving abnormal samples with similar weather but mismatched weekdays / seasons / weeks.

[0047] The normalized meteorological potential vector and the weighted time period vector are concatenated using channel concatenation to obtain the concatenated fused vector. This concatenated vector is then subjected to L2 normalization again to finally obtain the fused vector of historical samples. Or the fusion vector of the future prediction window , where the fusion vector This is the query feature vector used for similarity retrieval. The fused vector of historical samples. Query feature vector They are represented as follows:

[0048]

[0049] in, This indicates the L2 normalization operation. This is the time weighting coefficient.

[0050] Based on the query feature vector, a similar day retrieval is performed in a pre-built historical similar day retrieval database to obtain the top-K historical similar day samples, along with the corresponding historical load curve and similarity score for each sample. The historical similar day retrieval database uses the fusion vector of the historical sample as the search key and the corresponding load process curve as the search value, while simultaneously storing metadata including the historical sample's start timestamp, time slot identifier, meteorological source identifier, and search parameters. The historical similar day retrieval database can be constructed offline using the IndexFlatIP inner product nearest neighbor index, where the vector inner product is equivalent to cosine similarity.

[0051] Specifically, the historical similar day search database can be built offline in the following way: Based on a sliding time window `build_stride`, all historical load monitoring data is extracted. In this embodiment, `build_stride=12` is preferred (adapting to a 15-minute time granularity, with a step size corresponding to 3 hours). Each extracted segment is used as an independent historical sample, and each historical sample corresponds to a load process curve. The length is exactly equal to the downstream load forecast length pred_len.

[0052] Iterate through all historical samples, and generate historical meteorological potential vectors for each historical sample using the aforementioned method. Historical sample time period vector t i The fusion vector of normalized historical samples .

[0053] Fusion vector of historical samples As a search key, the load process curve corresponding to the historical sample The data is used as the retrieval value and metadata is stored synchronously. The metadata includes the historical sample start timestamp, time slot identifier, meteorological source identifier, retrieval parameters, etc.

[0054] An offline retrieval database is established using an inner product nearest neighbor search index or an equivalent similarity retrieval structure. This embodiment preferably uses the IndexFlatIP inner product nearest neighbor index to construct the historical similarity day retrieval database (DB), with the database structure defined as follows:

[0055] in, For the first The fusion vector of historical samples, For the first The load process curve corresponding to each historical sample For the first Metadata of a historical sample This represents the total number of historical samples.

[0056] Since all fused vectors are L2 normalized, the vector inner product is equivalent to cosine similarity, thus achieving similarity calculation and significantly reducing the cost of similarity calculation. The retrieval library in this embodiment supports persistent offline storage, achieving decoupling between offline library construction and online querying.

[0057] In the prediction phase, historical candidate samples that meet preset retrieval constraints are selected from the constructed historical similar day retrieval database. In this embodiment, the preset retrieval constraints include: the timestamp of the historical candidate sample is earlier than the current query time, and the time slot identifier of the historical candidate sample is the same as the time slot identifier of the query sample. For all obtained historical candidate samples, the dot product of the query feature vector and the fusion vector corresponding to each historical candidate sample is calculated to obtain the cosine similarity; the historical candidate samples are sorted in descending order of cosine similarity scores; based on the sorting results, the top K candidate historical samples, i.e., the top-K historical similar day samples, are selected, and the top-K historical load curves and their similarity scores corresponding to the top-K historical similar day samples are obtained; where K is a positive integer greater than or equal to 1.

[0058] In this embodiment, the retrieval process can be represented as follows:

[0059] in, This indicates a Top-K search operation based on FAISS. This represents the set of indexes of similar samples returned, where K is the number of historical similar day samples.

[0060] For the candidate set that satisfies the time constraint The similarity score can be expressed as:

[0061] in, This represents the dot product of vectors. Because... and All values ​​have been normalized, therefore the above inner product can be equivalently regarded as cosine similarity.

[0062] Furthermore, this embodiment is geared towards the short-term load forecasting task of the regional power grid. The input data of the entire model can be defined as 5 types of structured variables, namely: (1) historical load sequence load_x; (2) time mark x_mark_enc that matches the historical load; (3) exogenous time mark x_exo_mark corresponding to the future forecast period; (4) weather_x, the weather grid sequence of the future window, which can also be retrieved in batches from the meteorological database through the weather frame index weather_x_index; (5) similar day prior feature tensor generated by retrieval.

[0063] Step S2: Based on the historical similar day retrieval results, construct the prior feature tensor of similar days.

[0064] Specifically, let the top-K similar daily load curves obtained from the retrieval be: The corresponding similarity score is: .

[0065] The similarity scores are normalized exponentially to obtain the weight w shown in equation (1). i : (1) Where i = 1, 2, ..., K; j = 1, 2, ..., K.

[0066] Based on the aforementioned weights, a softmax weighted summation is performed on multiple similar daily load curves to obtain the weighted prior mean curve for similar days. : (2) The weighted prior mean curve of similar days is concatenated in parallel with the previously obtained Top-K original load curves of similar days according to time steps to generate a prior feature tensor P of similar days with multiple channels: (3) The first channel is the weighted prior mean curve of similar days, and the second to (K+1) channels are the load curves of each similar day.

[0067] In the preferred configuration of this embodiment, the number of historical similar days retrieved is Top-K=3, that is, the prior feature tensor P of similar days is broken by the weighted prior mean curve of similar days and the load curve of 3 similar days. The corresponding generated prior feature tensor of similar days has a dimension of [Batch, pred_len, TopK+1] and a total number of channels of 4. For example, when the prediction duration is 1 day and the time granularity is 15 minutes, the tensor size is [Batch, 96, 4].

[0068] Step S3: Extract the meteorological raster sequence corresponding to the future prediction window from the meteorological raster data of the target area. Use a meteorological convolutional coding network to perform convolutional coding on the meteorological raster frames at each time step in the meteorological raster sequence, and arrange the encoded meteorological feature vectors in chronological order to generate the exogenous meteorological feature sequence corresponding to the future prediction window. This step is performed by the full-map meteorological convolutional feature extraction module.

[0069] Specifically, the meteorological raster sequence corresponding to the future forecast window can be extracted from the meteorological raster data of the target area. This can be done by inputting the complete meteorological raster sequence [B, T, C, H, W] (batch, time step, channel, height, width) from a unified meteorological frame library, or by inputting [B, T] (batch, time step) as an index from the unified meteorological frame library, thus obtaining the meteorological raster sequence corresponding to the future forecast window.

[0070] In this embodiment, the meteorological convolutional coding network employs a CNN full-image convolutional feature extractor. The full-image convolutional feature extractor performs convolutional coding on the meteorological raster frames at each time step, outputting the corresponding meteorological feature vectors. These encoded meteorological feature vectors from each time step are then arranged in chronological order to form an exogenous meteorological feature sequence. To reduce memory usage, a block-based coding method can be used to process the meteorological frames in batches. In this embodiment, the original input dimensions of the meteorological raster are [B, T, C, H, W] (batch, time step, meteorological channel, region height, region width). The meteorological channel includes six types of meteorological features: temperature, humidity, wind speed, air pressure, precipitation, and irradiance.

[0071] Step S4: Input the historical power load sequence, time stamp information and exogenous meteorological feature sequence into the time series backbone prediction network to obtain the baseline load prediction result; the baseline load prediction result includes the point prediction value of the target area at each prediction time within the future prediction window; the time series backbone prediction network is a deep learning network with time series modeling capability, used to learn the nonlinear relationship between historical load and meteorological features.

[0072] Specifically, the historical power load sequence, time stamp information, and the exogenous meteorological feature sequence obtained in step S3 are input into the time-series backbone prediction network, which outputs the baseline load prediction result for the future prediction period. In this embodiment, the time-series backbone prediction network is preferably the TimeXer network. Of course, in other embodiments, it can also be replaced by other deep networks with time-series modeling capabilities, as long as they can meet the requirement of learning the nonlinear relationship between historical load and meteorological features.

[0073] In this embodiment, the baseline load forecast result output by the time-series backbone forecasting network is assumed to be the future load forecast sequence. : (4) Step S5: Based on the prior feature tensor of similar days and the baseline load prediction results, the dynamic correction module generates the corrected point prediction values ​​for each prediction time. The dynamic correction module generates dynamic correction coefficients through a gating network based on the prior feature tensor of similar days and the baseline load prediction results, and performs residual correction on the baseline load prediction results using the dynamic correction coefficients.

[0074] Specifically, for each prediction time step within the future prediction window, the load prediction sequence obtained in step S4 is used. The prior mean term is constructed. Prediction deviation term With the prior discreteness term Three types of features. Among them, the prior mean term. The weighted prior_mean curve of similar days obtained in step S2 by weighting using Softmax is... t Prediction bias term Defined as the difference between the prior mean term and the baseline load forecast result, i.e.: Prior discreteness term To characterize the dispersion of the similar day load curves in the 2nd to (K+1th)th channels of the similar day prior feature tensor at that time step, this embodiment uses the standard deviation, i.e.: .

[0075] The above three types of features are concatenated to construct the input vector of the gating network. : (5) In a preferred embodiment of this application, the prediction deviation term Gradient blocking is performed when inputting to the gating network, so that the model training tool PyTorch no longer records the gradient changes of gap_t. This avoids the gating network directly interfering with the basic prediction results of the backbone network during backpropagation, thereby reducing gradient competition and improving the stability of joint training.

[0076] Gated input vector The data is fed into the gate control network to obtain the dynamic correction coefficients at the time step level. (6) in, The Sigmoid activation function is used. The value range is (0,1).

[0077] According to equation (7), the obtained dynamic correction coefficients are used to perform residual correction on the main prediction results to obtain the corrected point prediction values ​​at each prediction time. : (7) Right now: (8) Therefore, we can conclude that: 1) When When the value is close to 0, the model primarily relies on the time series backbone prediction. 2) When When the value is close to 1, the model tends to use the correction direction based on prior similar days more often; 3) When When the magnitude is large, the gating network can learn to reduce the magnitude of prior corrections in order to reduce the misleading effects of unreliable priors.

[0078] For example, in this embodiment, the gated network adopts a single hidden layer structure with a hidden layer dimension of 32; at the same time, the output layer bias of the gated network is initialized by inverse solution, and the initial gate coefficient is preset so that the correction weights are close to 0 in the early stage of training, and the temporal backbone prediction results are retained first.

[0079] Step S6: Input the corrected point prediction values ​​at each prediction time into the quantile prediction head to obtain the power load probability prediction result. The power load probability prediction result includes the load prediction interval corresponding to each prediction time within the future prediction window. The quantile prediction head is obtained through supervised training of the quantile loss function and is used to output the load prediction values ​​of multiple quantile points to form the load prediction interval.

[0080] Specifically, let the corrected point prediction values ​​for each prediction time output in step S5 be... The resulting point prediction sequence is as follows:

[0081] In this embodiment, the set of quantiles is defined as follows:

[0082] Where M is the number of quantiles.

[0083] For each prediction time step t, the corrected point prediction values ​​at each prediction time are... Input the quantile prediction header to obtain the output value at each quantile: (9) in:

[0084] This is the weight matrix for the quantile predictor head. For bias vectors, This indicates that time step t is at the quantile. The predicted value is below.

[0085] The quantile loss function uses the deviation between the actual load value and the predicted value at the corresponding quantile as the optimization objective, and is used to adjust the network parameters of the quantile prediction head. For any quantile... Its quantile loss function is defined as: (10) Where, q m y is a quantile; t true This represents the actual load value; y t,qm For quantile q m The predicted value is below.

[0086] Under the supervision of this quantile loss function, the weights in the linear layer and bias The cumulative probability is optimized to fit the target conditional distribution. Through this mechanism, the model predicts point values. As a distribution center reference, the statistical characteristics of load deviation from the center at different confidence levels are learned, thereby improving the output results. satisfy The probability requirement.

[0087] Furthermore, the quantile output for all prediction periods can be expressed as:

[0088] It can also be written equivalently as:

[0089] in, This represents the quantile prediction head mapping function.

[0090] In this embodiment, when the Q set of quantiles is taken as When the time is right, the corresponding outputs are the lower quantile predicted value, the median predicted value, and the upper quantile predicted value, respectively, that is:

[0091] Among them, the 0.5 quantile output This can be used as a median prediction result; the combination of 0.1 and 0.9 quantiles forms a 90% confidence load prediction interval. It can be used as a prediction confidence interval, and at the same time, the low quantile, median and high quantile load prediction results are obtained.

[0092] In addition, to improve prediction accuracy and engineering stability, in the preferred embodiment of this application, the prediction model used in the power load prediction method of this embodiment is subjected to joint training of automatic hyperparameter optimization and similar day correction.

[0093] Specifically, before introducing the similarity-day dynamic correction module involved in step S5, a hyperparameter search space is constructed for the combined backbone network that already includes the full-map meteorological convolutional feature extraction module and the temporal backbone prediction module. An automatic hyperparameter optimization framework based on Bayesian optimization is used for offline optimization to obtain a better backbone network configuration and pre-trained weights, thereby improving the starting point quality of subsequent joint fine-tuning. The automatic hyperparameter optimization framework is preferably implemented using the Optuna third-party Python library, but is not limited to this specific tool.

[0094] The hyperparameter search objects include, but are not limited to, one or more of the following parameters: meteorological feature dimension; temporal backbone network hidden dimension; number of attention heads; number of encoding layers; feedforward network dimension; dropout parameter; patch length; batch size; learning rate; number of similar days K.

[0095] For example, the engineering values ​​can be as follows: the TimeXer hidden dimension of the time-series backbone is 512, the number of coding layers is 2, the dropout coefficient dimension is 0.1, the patch length dimension is 96, and the batch size is 32.

[0096] In this embodiment, the hyperparameter vector is set as follows:

[0097] For each group of candidate hyperparameters Train the combined backbone network and compute the objective function value on the validation set:

[0098] in, For the loss function value on the validation set, quantile loss, mean squared error, mean absolute error, or a combination thereof can be used.

[0099] The automatic hyperparameter optimization process can be represented as:

[0100] in, For the pre-defined hyperparameter search space, The optimal hyperparameter configuration obtained from the search.

[0101] After obtaining the optimal hyperparameter configuration, the optimal configuration file, optimal hyperparameter file, optimal backbone model weight file, and optimal experimental result record file are saved to obtain the optimal backbone network foundation. Subsequently, a similar daily dynamic correction module and a quantile prediction head are introduced on the basis of the optimal backbone network to perform joint training on the prediction model used in this embodiment. The specific training process adopts a two-stage fine-tuning strategy: The first stage is the warm-up and fine-tuning stage, which includes: The meteorological convolutional coding network and the temporal backbone prediction network were frozen, and only the dynamic gating network and quantile prediction head involved in the newly added similar daily dynamic correction module were trained. A relatively large third learning rate was used to enable the two newly added modules to quickly converge to a reasonable parameter range.

[0102] The second stage is the joint fine-tuning stage, which includes: unfreezing all network parameters included in the hyperparameter search object; performing protective fine-tuning on the meteorological convolutional coding network and the temporal backbone prediction network using a relatively low first learning rate; and continuing optimization on the dynamic gating network and the quantile prediction head using a relatively higher second learning rate. Specifically, the third learning rate is greater than the second learning rate, and the second learning rate is greater than the first learning rate.

[0103] Through the training method, this embodiment can effectively avoid the addition of modules from damaging the feature extraction capability of the pre-trained backbone network, thus ensuring the model's convergence stability and prediction accuracy.

[0104] In a preferred embodiment, during the training of the prediction model, the bias of the output layer of the dynamic gating network is initialized by inverse kinematics according to the preset initial gating coefficients. This allows the model to use a small proportion of similar day corrections in the early stages of training, prioritizing the retention of the backbone prediction capability. As training progresses, the model gradually learns a better correction magnitude, further improving the convergence effect.

[0105] As a preferred engineering configuration, the number of training rounds in the two stages of this embodiment is set as follows: 10 rounds of warm-up and fine-tuning iteration in the first stage, and 15 rounds of joint fine-tuning iteration across the entire network in the second stage.

[0106] By adopting the above scheme, the power load forecasting method provided in this embodiment can improve the load forecasting accuracy under complex weather conditions such as high temperature, cold wave, and heavy rainfall, enhance the robustness of the forecasting model under different load modes such as weekdays, holidays, and special event days, and provide point forecasting and interval forecasting results at the same time, providing support for dispatching operation, spot trading declaration and risk boundary analysis.

[0107] Example 2 In another typical embodiment of the present invention, a power load probability prediction system based on a full-map meteorological grid and similar day prior dynamic correction is provided, specifically including: The data acquisition module is used to acquire multi-source input data, including historical power load sequences, time stamp information, meteorological raster data of the target area, and historical similar day retrieval results. The historical power load sequence consists of historical load values ​​of the target area arranged in chronological order. The time stamp information includes the time period characteristics corresponding to each sampling time of the historical power load sequence, as well as the time period characteristics corresponding to each prediction time within the future prediction window. The meteorological raster data of the target area consists of multivariate meteorological time series data with a spatial raster structure within the target area. The historical similar day retrieval results include load curves and similarity scores corresponding to multiple historical similar days. The historical similar days are similar samples retrieved based on the fusion vector of meteorological and time features of the future prediction window. The similarity score is the cosine similarity between the fusion vector of the future prediction window and the fusion vector of the corresponding historical similar day. The prior feature construction module is used to construct the prior feature tensor of similar days based on the historical similar day retrieval results; The meteorological feature processing module is used to extract the meteorological raster sequence corresponding to the future prediction window from the meteorological raster data of the target area, and to perform convolutional coding on the meteorological raster frames at each time step in the meteorological raster sequence using a meteorological convolutional coding network. The encoded meteorological feature vectors are arranged in chronological order to generate the exogenous meteorological feature sequence corresponding to the future prediction window. The baseline load forecasting module is used to input historical power load sequences, time stamp information, and exogenous meteorological feature sequences into the time-series backbone forecasting network to obtain baseline load forecasting results. The baseline load forecasting results include point forecast values ​​for the target area at each forecast time within the future forecast window. The time-series backbone forecasting network is a deep learning network with time-series modeling capabilities, used to learn the nonlinear relationship between historical load and meteorological features. The dynamic correction module generates the corrected point prediction values ​​for each prediction time based on the prior feature tensor of similar days and the baseline load prediction results. The dynamic correction module generates dynamic correction coefficients through a gating network based on the prior feature tensor of similar days and the baseline load prediction results, and performs residual correction on the baseline load prediction results using the dynamic correction coefficients. The load probability prediction module inputs the corrected point prediction values ​​at each prediction time into the quantile prediction head to obtain the power load probability prediction result. The power load probability prediction result includes the load prediction interval corresponding to each prediction time within the future prediction window. The quantile prediction head is obtained through supervised training with the quantile loss function and is used to output the load prediction values ​​of multiple quantile points to form the load prediction interval.

[0108] The present invention also provides an electronic device, including a memory and a processor. The memory stores a computer program, which, when executed by the processor, can implement any step of the power load probability prediction method based on full-map meteorological grids and similar day prior dynamic correction in the embodiments.

[0109] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can implement any step of the power load probability prediction method based on full-map meteorological grids and similar day prior dynamic correction in the embodiments.

[0110] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Those skilled in the art should understand that the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A power load probability prediction method based on full-map meteorological grids and similar-day prior dynamic correction, characterized in that, To predict the power load of a target area within a future prediction window, where the future prediction window is a preset prediction time interval, the method includes the following steps: The system acquires multi-source input data, including historical power load sequences, time stamp information, meteorological raster data of the target area, and historical similar day retrieval results. The historical power load sequence consists of historical load values ​​of the target area arranged in chronological order. The time stamp information includes time period features corresponding to each sampling time of the historical power load sequence and time period features corresponding to each prediction time within the future prediction window. The meteorological raster data of the target area consists of multivariate meteorological time-series data with a spatial raster structure within the target area. The historical similar day retrieval results include load curves and similarity scores corresponding to multiple historical similar days. The historical similar days are similar samples retrieved based on the fusion vector of meteorological and time features of the future prediction window, and the similarity score is the cosine similarity between the fusion vector of the future prediction window and the fusion vector of the corresponding historical similar day. Based on the historical similar day retrieval results, a similar day prior feature tensor is constructed; The meteorological raster sequence corresponding to the future prediction window is extracted from the meteorological raster data of the target area. A meteorological convolutional coding network is used to perform convolutional coding on the meteorological raster frames at each time step in the meteorological raster sequence. The encoded meteorological feature vectors are arranged in chronological order to generate the exogenous meteorological feature sequence corresponding to the future prediction window. The historical power load sequence, the time stamp information, and the exogenous meteorological feature sequence are input into the time-series backbone prediction network to obtain the baseline load prediction result. The baseline load prediction result includes the point prediction value of the target area at each prediction time within the future prediction window. The time-series backbone prediction network is a deep learning network with time-series modeling capabilities, used to learn the nonlinear relationship between historical load and meteorological features. Based on the prior feature tensor of similar days and the baseline load prediction results, the dynamic correction module generates the corrected point prediction values ​​for each prediction time. The dynamic correction module generates dynamic correction coefficients through a gating network based on the prior feature tensor of similar days and the baseline load prediction results, and performs residual correction on the baseline load prediction results using the dynamic correction coefficients. The corrected point prediction values ​​at each prediction time are input into the quantile prediction head to obtain the power load probability prediction result. The power load probability prediction result includes the load prediction interval corresponding to each prediction time within the future prediction window. The quantile prediction head is obtained through supervised training with the quantile loss function and is used to output the load prediction values ​​of multiple quantile points to form the load prediction interval.

2. The power load probability prediction method based on full-map meteorological grids and similar day prior dynamic correction as described in claim 1, characterized in that, The construction of the prior feature tensor for similar days based on the historical similar day retrieval results specifically includes: The similarity scores corresponding to the multiple similar daily load curves are subjected to exponential normalization to obtain the weights; Based on the weights, the multiple similar daily load curves are weighted and summed to obtain the weighted similar daily prior mean curve. The weighted similar day prior mean curve is concatenated with the multiple similar day load curves according to time steps to generate the similar day prior feature tensor.

3. The power load probability prediction method based on full-map meteorological grids and similar-day prior dynamic correction as described in claim 2, characterized in that, The dynamic correction module, based on the prior feature tensor of similar days and the baseline load prediction result, generates dynamic correction coefficients through a gating network, and uses these dynamic correction coefficients to perform residual correction on the baseline load prediction result. Specifically, this includes: Based on the prior feature tensor of similar days and the baseline load prediction result, a prior mean term, a prediction bias term, and a prior dispersion term are calculated; wherein, the prior mean term is the weighted prior mean curve of similar days, the prediction bias term is the difference between the prior mean term and the baseline load prediction result, and the prior dispersion term is used to characterize the degree of dispersion of each similar day load curve in the prior feature tensor of similar days at the corresponding time step; The prior mean, prediction bias, and prior dispersion terms are concatenated to obtain the gating network input vector. The input vector of the gated network is input into the gated network to obtain the dynamic correction coefficients; The dynamic correction coefficient is used to perform residual correction on the baseline load prediction results to obtain the corrected point prediction values ​​at each prediction time.

4. The power load probability prediction method based on full-map meteorological grids and similar day prior dynamic correction as described in claim 3, characterized in that, The process of using the dynamic correction coefficient to perform residual correction on the baseline load prediction results to obtain the corrected point prediction values ​​at each prediction time is as follows: The product of the dynamic correction coefficient and the prediction deviation term is added to the baseline load prediction result to obtain the corrected point prediction value at each prediction time.

5. The power load probability prediction method based on full-map meteorological grids and similar day prior dynamic correction as described in claim 1, characterized in that, The method further includes a step of training a prediction model used to perform the power load probability prediction method, wherein the training employs a two-stage fine-tuning strategy: The first stage is the warm-up and fine-tuning stage, in which the meteorological convolutional coding network and the temporal backbone prediction network are frozen, and only the dynamic correction module and the quantile prediction head are trained. The second stage is the joint fine-tuning stage, in which all network parameters are unfrozen, and the meteorological convolutional coding network and the temporal backbone prediction network are fine-tuned using a first learning rate, while the dynamic correction module and the quantile prediction head are further optimized using a second learning rate; wherein, the second learning rate is greater than the first learning rate.

6. The power load probability prediction method based on full-map meteorological grids and similar day prior dynamic correction as described in claim 1, characterized in that, The step of inputting the corrected point prediction values ​​at each prediction time into the quantile prediction head to obtain the power load probability prediction result specifically includes: For each prediction time step within the future prediction window, the corrected point prediction values ​​at each prediction time are used as input, and the load prediction values ​​at each quantile are calculated through linear mapping. The quantile prediction head is trained under supervision using a quantile loss function to determine the load prediction value corresponding to the cumulative probability at each quantile based on the point prediction value at each prediction time after correction. The expression for the quantile loss function is: Where, q m y is a quantile; t true This represents the actual load value; y t,qm For quantile q m The predicted value below, The quantile loss function; Based on the quantile probability values ​​from low to high, the load prediction values ​​corresponding to each quantile are respectively assigned to the low quantile, median, and high quantile load prediction results.

7. The power load probability prediction method based on full-map meteorological grids and similar day prior dynamic correction as described in claim 1, characterized in that, The similar date search results are obtained through the following methods: Historical weather windows are encoded and fused with time cycle features to construct a similar day retrieval database; After encoding the meteorological window corresponding to the future prediction window and fusing it with the time cycle features, the top K historical samples with the highest similarity are searched in the similar day retrieval library as similar samples, and the corresponding historical load curves and similarity scores are generated as the historical similar day retrieval results.

8. A power load probability prediction system based on full-map meteorological grids and similar-day prior dynamic correction, characterized in that, Specifically, it includes: The data acquisition module is used to acquire multi-source input data, including historical power load sequences, time stamp information, meteorological raster data of the target area, and historical similar day retrieval results. The historical power load sequence is the historical load value of the target area arranged in chronological order. The time stamp information includes the time period characteristics corresponding to each sampling time of the historical power load sequence and the time period characteristics corresponding to each prediction time within the future prediction window. The meteorological raster data of the target area is multivariate meteorological time series data with a spatial raster structure within the target area. The historical similar day retrieval results include load curves and similarity scores corresponding to multiple historical similar days. The historical similar days are similar samples retrieved based on the fusion vector of meteorological and time features of the future prediction window. The similarity score is the cosine similarity between the fusion vector of the future prediction window and the fusion vector of the corresponding historical similar day. The prior feature construction module is used to construct a prior feature tensor for similar days based on the historical similar day retrieval results. The meteorological feature processing module is used to extract the meteorological grid sequence corresponding to the future prediction window from the meteorological grid data of the target area, use a meteorological convolutional coding network to perform convolutional coding on the meteorological grid frames at each time step in the meteorological grid sequence, and arrange the encoded meteorological feature vectors in chronological order to generate the exogenous meteorological feature sequence corresponding to the future prediction window. The baseline load forecasting module is used to input the historical power load sequence, the time stamp information, and the exogenous meteorological feature sequence into the time-series backbone forecasting network to obtain the baseline load forecasting result. The baseline load forecasting result includes the point forecast values ​​of the target area at each forecast time within the future forecast window. The time-series backbone forecasting network is a deep learning network with time-series modeling capabilities, used to learn the nonlinear relationship between historical load and meteorological features. The dynamic correction module generates corrected point prediction values ​​for each prediction time based on the prior feature tensor of similar days and the baseline load prediction results; the dynamic correction module generates dynamic correction coefficients through a gating network based on the prior feature tensor of similar days and the baseline load prediction results, and performs residual correction on the baseline load prediction results using the dynamic correction coefficients. The load probability prediction module inputs the corrected point prediction values ​​at each prediction time into the quantile prediction head to obtain the power load probability prediction result. The power load probability prediction result includes the load prediction interval corresponding to each prediction time within the future prediction window. The quantile prediction head is obtained through supervised training with a quantile loss function and is used to output the load prediction values ​​of multiple quantile points to form the load prediction interval.

9. An electronic device, characterized in that, The device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, can implement the steps of the power load probability prediction method based on full-map meteorological grids and similar day prior dynamic correction as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by a processor, can implement the steps of the power load probability prediction method based on full-map meteorological grids and similar day prior dynamic correction as described in any one of claims 1-7.

Images