Multi-building short-term load forecasting method, device, equipment and readable storage medium

Abstract

The application relates to a multi-building short-term load prediction method, comprising the following steps: obtaining public event natural language data, building metadata of a plurality of target buildings and structured load prediction multi-source basic data; performing hour-level alignment processing and robust preprocessing on the load prediction multi-source basic data and the building metadata to obtain aligned preprocessing data; performing load prediction related text retrieval based on the public event natural language data to obtain retrieval text; inputting the retrieval text, the public event natural language data and a prediction time into a preset analysis model to analyze event records and generate expert weight priori; generating expert fusion weight according to the expert weight priori, the aligned preprocessing data and the event records through a context encoder and a horizontal dynamic gate mechanism; and generating load prediction values of the plurality of target buildings according to the expert fusion weight and a candidate load prediction sequence generated based on a preset scene expert library. The method can improve the accuracy of building load prediction.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, computer equipment, computer-readable storage medium, and computer program product for short-term load forecasting of multiple buildings. Background Technology

[0002] Short-term building load forecasting is a crucial foundation for energy-efficient building operation, model predictive control, demand response, and campus-level energy management. Existing building load forecasting technologies typically utilize structured data such as historical loads, meteorological variables, and calendar variables for modeling. With the development of sequence models such as BiLSTM, DLinear, and PatchTST, forecasting performance for regular time periods has improved. In the actual operation of urban building clusters, building electricity demand is influenced not only by structured variables such as temperature, humidity, and weekdays, but also by external events presented in natural language text form, such as holiday schedule adjustments, weather warnings, municipal announcements, commercial activities, and public events.

[0003] However, existing solutions typically rely on structured inputs such as historical loads, meteorological observations, and calendar variables, and use statistical models, machine learning models, or deep sequence models to output future load values. They fail to fully utilize natural language event information from weather warnings, holiday notices, municipal announcements, and announcements of large-scale events, resulting in low accuracy in building load forecasting. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, computer equipment, computer-readable storage medium, and computer program product for short-term building load forecasting that can improve the accuracy of building load forecasting, in response to the above-mentioned technical problems.

[0005] Firstly, this application provides a method for short-term load forecasting of multiple buildings, including:

[0006] Acquire unstructured natural language data of public events, as well as building metadata for multiple target buildings and structured multi-source basic data for load forecasting;

[0007] Hourly alignment and robust preprocessing are performed on multi-source basic data and building metadata for load forecasting to obtain aligned preprocessed data. Load forecasting-related text retrieval is then performed based on natural language data of public events to obtain the retrieved text.

[0008] The retrieved text, natural language data of public events, and predicted time are input into a preset parsing model to parse the event records and generate expert weight priors based on the event records.

[0009] By using a context encoder and a horizontal dynamic gating mechanism, expert fusion weights are generated based on expert weight priors, aligned preprocessed data, and event records.

[0010] Obtain multiple candidate load prediction sequences generated based on a preset scenario expert database, and generate load prediction values ​​for multiple target buildings based on the expert fusion weights and each candidate load prediction sequence.

[0011] In one embodiment, the multi-source load forecasting base data includes: meteorological data, calendar data, and historical load data; hourly alignment and robust preprocessing are performed on the multi-source load forecasting base data and building metadata to obtain aligned preprocessed data, including:

[0012] At least a portion of the meteorological data that is not evenly spaced after being sorted by timestamp is resampled to the hourly grid, and at least a portion of the meteorological data with missing meteorological values ​​is interpolated and filled to obtain usable meteorological data.

[0013] Hourly time stamps are constructed on the calendar data, and the calendar data with constructed hourly time stamps, historical load data, and available meteorological data are stitched together to form hourly aligned data;

[0014] Robust preprocessing is performed on the abnormal data in historical load data and building metadata to obtain robust preprocessed data;

[0015] By combining hourly aligned data with robust preprocessed data, aligned preprocessed data is obtained.

[0016] In one embodiment, load prediction-related text retrieval is performed based on public event natural language data to obtain the retrieved text, including:

[0017] The correlation assessment results are obtained by performing target building correlation assessment, prediction time correlation assessment, and load change correlation assessment on the natural language data of public events.

[0018] Based on the relevance assessment results, the natural language data of public events is filtered to obtain the retrieval text.

[0019] In one embodiment, generating expert weight priors based on event logs includes:

[0020] Event records are encoded into event vectors, and event vectors are aggregated into a unified event representation using a preset weighted pooling method.

[0021] By using a preset activation function, the unified event representation is processed into expert weighted priors.

[0022] In one embodiment, expert fusion weights are generated based on expert weight priors, aligned preprocessed data, and event records, using a context encoder and a horizontal dynamic gating mechanism, including:

[0023] The context encoder extracts the context representation from the aligned preprocessed data.

[0024] Based on contextual representation, prior expert weights, and unified event representation generated from event records, expert fusion weights are generated through a horizontal dynamic gating mechanism.

[0025] In one embodiment, multiple candidate load prediction sequences are generated in the following manner:

[0026] Acquire multiple expert models and construct a pre-defined expert database based on these models;

[0027] The aligned preprocessed data is input into a preset scenario expert library to obtain multiple candidate load prediction sequences.

[0028] Secondly, this application also provides a multi-building short-term load forecasting device, comprising:

[0029] The multi-source data acquisition module is used to acquire unstructured natural language data of public events, as well as building metadata of multiple target buildings and structured multi-source basic data for load forecasting.

[0030] The preprocessing and text retrieval module is used to perform hourly alignment and robust preprocessing on multi-source basic data and building metadata for load forecasting to obtain aligned preprocessed data, and to perform load forecasting-related text retrieval based on public event natural language data to obtain the retrieved text.

[0031] The weight prior generation module is used to input the retrieved text, public event natural language data, and prediction time into the preset parsing model, parse out the event records, and generate expert weight priors based on the event records.

[0032] The expert fusion weight generation module is used to generate expert fusion weights based on expert weight priors, aligned preprocessed data, and event records through a context encoder and a horizontal dynamic gating mechanism.

[0033] The load forecasting module is used to acquire multiple candidate load forecasting sequences generated based on a preset scenario expert database, and generate load forecasting values ​​for multiple target buildings based on the expert fusion weights and each candidate load forecasting sequence.

[0034] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above embodiments.

[0035] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above embodiments.

[0036] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps in the above embodiments.

[0037] The aforementioned method, apparatus, computer equipment, computer-readable storage medium, and computer program product for short-term load forecasting of multiple buildings first acquire unstructured natural language data of public events, as well as building metadata and structured multi-source basic data for load forecasting for each of the multiple target buildings. Hourly-level alignment and robust preprocessing are then performed on the multi-source basic data and building metadata to obtain aligned preprocessed data. Load forecasting-related text retrieval is then performed based on the public event natural language data to obtain the retrieval text. Next, the retrieval text, public event natural language data, and forecast time are input into a preset parsing model to parse event records. Expert weight priors are generated based on these event records. Further, through a context encoder and a horizontal dynamic gating mechanism, expert fusion weights are generated based on the expert weight priors, aligned preprocessed data, and event records. Finally, based on the expert fusion weights and multiple candidate load forecast sequences generated from a preset scenario expert database, load forecast values ​​for multiple target buildings are generated. The entire process can effectively integrate public event natural language data, building metadata, and structured load forecasting multi-source basic data. By using a pre-set scenario expert database and horizontal dynamic gating, it can improve the prediction accuracy, robustness, and interpretability of building loads under various scenarios, thereby effectively improving the accuracy of building load forecasting. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a diagram illustrating the application environment of a multi-building short-term load forecasting method in one embodiment.

[0040] Figure 2 This is a flowchart illustrating a method for short-term load forecasting of multiple buildings in one embodiment;

[0041] Figure 3 This is a flowchart illustrating the implementation of hour-level alignment and robust preprocessing in one embodiment;

[0042] Figure 4 This is a schematic diagram of the architecture of a short-term multi-building load forecasting system based on a large language model multi-source event perception and dynamic expert fusion in one embodiment.

[0043] Figure 5 This is a structural block diagram of a multi-building short-term load forecasting device in one embodiment;

[0044] Figure 6 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0046] The multi-building short-term load forecasting method provided in this application embodiment can be applied to, for example... Figure 1 In the application environment shown, data acquisition terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or placed on a cloud or other network server. Server 104 can acquire unstructured public event natural language data, building metadata for multiple target buildings, and structured multi-source load forecasting basic data, as well as multiple candidate load forecasting sequences generated based on a preset scenario expert database, through various data acquisition terminals 102. Furthermore, server 104 can perform hourly-level alignment processing and robust preprocessing on the load forecasting multi-source basic data and building metadata to obtain aligned preprocessed data, and perform load forecasting-related text retrieval based on the public event natural language data to obtain retrieval text. Then, the retrieval text, public event natural language data, and forecast time are input into a preset analytical model to parse event records and generate expert weight priors based on the event records. Finally, server 104 can generate expert fusion weights based on expert weight priors, aligned preprocessed data, and event records through context encoder and horizontal dynamic gating mechanism. Then, based on the expert fusion weights and each candidate load prediction sequence, load prediction values ​​for multiple target buildings are generated.

[0047] The short-term multi-building load forecasting involves predicting the power load of multiple buildings within a building complex over a relatively short period. This embodiment uses 48 hours of historical data to predict the load for the next 24 hours by default. The data acquisition terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and IoT devices. The server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0048] In one exemplary embodiment, such as Figure 2As shown, a method for short-term load forecasting of multiple buildings is provided, which is then applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps 202 to 210. Wherein:

[0049] Step 202: Obtain unstructured public event natural language data, as well as building metadata and structured load forecast multi-source basic data for each target building.

[0050] Unstructured public event natural language data includes weather warnings, holiday notices, municipal announcements, and public events. This data can originate from traffic announcements, business activity platforms, news summaries, property notices, equipment alarm records, or publicly available event summaries on social media. This embodiment enhances the utilization of unstructured public event natural language data. Structured load forecasting multi-source basic data includes meteorological data, calendar data, and historical load data. Building metadata describes the static attributes of a building, including building type, secondary type, total building area, air-conditioned area, and cooling system attributes.

[0051] Optionally, the server can acquire natural language data of public events, as well as meteorological data, calendar data, historical load data, and building metadata of multiple target buildings from the Internet or other channels, or by communicating with various data acquisition terminals, in order to input the subsequent multi-source event perception and identify the event semantics that will affect the building load from multiple sources of text such as weather warnings, holiday notices, municipal announcements, and public event announcements.

[0052] For example, for a group of buildings including multiple target buildings, the server can collect or receive the first... Historical load data series consisting of historical load data of target buildings , To predict the time, The input window length can be set to 24, 48, 72 hours, etc., according to business requirements. The server can also collect or receive structured exogenous feature sequences. Structured exogenous feature sequences are time-aligned numerical or categorical features other than historical loads, including meteorological and calendar data. Meteorological data includes meteorological variables such as temperature, humidity, wind speed, visibility, rainfall, air pressure, and air quality, while calendar data includes calendar variables. Simultaneously, the server can also collect or receive building metadata sequences composed of building metadata. Building metadata includes building type, secondary type, total building area, air-conditioned area, cooling system attributes, etc.

[0053] Step 204: Perform hourly alignment and robust preprocessing on the multi-source basic data of load forecasting and building metadata to obtain aligned preprocessed data, and perform load forecasting-related text retrieval based on public event natural language data to obtain the retrieved text.

[0054] The core objective of hourly alignment is to normalize messy and irregular timestamps onto a unified, standardized hourly time grid. This can be used to eliminate time discrepancies between multi-source data, normalizing them to the same hourly time scale. Robust preprocessing aims to resist anomalous interference and preserve the true signal, thus enabling the extraction of stable statistical features even when bad or outlier values ​​exist in the data. The hourly alignment and robust preprocessing steps in this process, such as hourly weather alignment, load spike pruning, short invalid segment interpolation, and long invalid segment masking, can reduce the impact of actual data anomalies on training and inference, improve data quality robustness, and enhance prediction accuracy under event-driven disturbances.

[0055] Optionally, the server can resample at least a portion of the meteorological data that is not evenly spaced after being sorted by timestamp to an hourly grid, and perform interpolation and imputation processing on at least a portion of the meteorological data with missing meteorological values ​​to obtain usable meteorological data. Simultaneously, hourly time stamps are constructed on the calendar data, and the calendar data with constructed hourly time stamps, historical load data, and usable meteorological data are concatenated to form hourly aligned data. Further, the server can perform robust preprocessing on anomalous data in the historical load data and building metadata to obtain robust preprocessed data, thereby aggregating the hourly aligned data and the robust preprocessed data to obtain aligned preprocessed data.

[0056] Step 206 involves inputting the retrieved text, public event natural language data, and predicted time into a preset parsing model to parse the event records and generate expert weight priors based on the event records.

[0057] Directly using a pre-defined parsing model to output load prediction values ​​can lead to insufficient numerical stability and controllability. Therefore, in this embodiment, the pre-defined parsing model is only used for event record parsing and expert weight prior generation, and does not directly output load prediction values ​​for multiple target buildings, thus improving the stability of numerical prediction. Specifically, the pre-defined parsing model can be a Large Language Model (LLM, a model capable of understanding and generating natural language text). The LLM only handles event understanding and expert prior generation, and does not directly output load values. Numerical prediction is completed by a trainable load prediction network and gating fusion, thereby reducing the impact of numerical illusions or unstable outputs from the LLM on the results. The LLM can also be replaced with other natural language understanding models, domain-fine-tuned language models, rule template parsers, named entity recognition models, or text classification models, as long as they can output structured event records and expert weight priors.

[0058] Among them, the expert weight prior is the initial tendency distribution of each expert model in the preset scenario expert library, calculated from the event records, which is used to guide the gating fusion model in learning the weights. The preset scenario expert library is a set of models consisting of multiple load prediction models (i.e., expert models) trained for different operating scenarios. The event records are the structured results of the preset parsing model (such as a large language model) for the output of the retrieved text, and contain at least the following information: event category, direction of impact, intensity of impact, affected period, expected duration, confidence level, and expert weight prior.

[0059] Optionally, the server can input building metadata and prediction times from retrieved text, public event natural language data, and other relevant data into a pre-defined parsing model (such as a Large Language Model, LLM). Additionally, a unified prompt template can also be input into the pre-defined parsing model to limit its output to structured event records only, excluding future load forecasts. Based on this, the event records parsed and output by the pre-defined parsing model can be obtained. These event records are then encoded into event vectors, and a pre-defined weighted pooling method is used to aggregate the event vectors into a unified event representation. Finally, a pre-defined activation function is used to process the unified event representation into expert weight priors. These steps, based on retrieved text, public event natural language data, and prediction times, fully utilize natural language event information from weather warnings, holiday notices, municipal announcements, and large-scale event announcements to parse event records and generate expert weight priors, rather than simply setting fixed expert weight priors or relying solely on numerical features to determine them. This effectively reduces prediction errors caused by holidays, weather warnings, and public events.

[0060] For example, the server can retrieve text, building metadata, and predicted time. And a unified prompt template for inputting a large language model (LLM). For the first... Search text Large Language Model (LLM) can output event logs. As shown in formula (1):

[0061] (1)

[0062] In formula (1), Indicates the first Building metadata of the target building. This represents the model processing function for a Large Language Model (LLM). Event categories in the event log can be constrained to a predefined set: normal state, holiday-related state, high-temperature state, extreme weather state, public event state, and operational anomaly state.

[0063] Step 208: Through the context encoder and horizontal dynamic gating mechanism, expert fusion weights are generated based on expert weight priors, aligned preprocessed data, and event records.

[0064] Optionally, the server can extract contextual representations from the aligned preprocessed data using a context encoder, then concatenate the contextual representations, expert weight priors, and a unified event representation generated based on event records. This concatenation is then input into a multilayer perceptron to obtain the output expert score matrix. The expert score matrix is ​​then normalized to obtain the horizontal dynamic fusion weights, i.e., the expert fusion weights. Through these steps, building metadata is injected into the text retrieval, event parsing, and expert fusion processes. This fully utilizes building metadata such as building type, building area, and air conditioning system to conditionally process text retrieval, event parsing, and expert selection, improving cross-building generalization capabilities. This allows the same external event to have differentiated impacts depending on the building type, such as office buildings or shopping malls, avoiding biases caused by a single shared load template and improving the adaptability to multi-building heterogeneity.

[0065] It should be noted that the expert fusion weights are not fixed, but are calculated in real time based on the input real-time data, varying with each prediction time and target building, thus exhibiting adaptability. The context encoder can be a deep learning model capable of deeply mining the inherent patterns in time-series data, capturing the dependencies, periodic fluctuations, and recent trends of historical loads along the time axis. The context representation is a high-dimensional dense vector output by the context encoder. The multilayer perceptron can be a feedforward neural network with multiple hidden layers and nonlinear activation functions. In this embodiment, it can serve as a scoring decision-maker, concatenating the context representation, expert weight priors, and unified event representation along the feature dimension to merge them into a longer comprehensive feature vector, enabling comprehensive reasoning based on the above three types of information simultaneously. The expert score matrix is ​​the scoring vector output by the multilayer perceptron. The values ​​in this matrix represent the preference strength that the model tends to assign to each expert model under comprehensive information; higher values ​​indicate greater applicability.

[0066] Step 210: Obtain multiple candidate load prediction sequences generated based on a preset scenario expert database, and generate load prediction values ​​for multiple target buildings according to the expert fusion weights and each candidate load prediction sequence.

[0067] The pre-defined scenario expert database includes at least the following expert models: general experts, holiday experts, high-temperature experts, extreme weather experts, and public event experts. The categories of expert models can be expanded to include typhoon experts, rainstorm experts, traffic congestion experts, sports event experts, and operational anomaly experts, depending on the specific region and business requirements.

[0068] Optionally, the server can obtain multiple candidate load prediction sequences generated based on a preset scenario expert database, and generate load prediction values ​​for multiple target buildings according to the expert fusion weights and each candidate load prediction sequence, as shown in formula (2):

[0069] For the With a prediction step size, the candidate load prediction sequences of the expert models are fused according to the expert fusion weights to obtain the load prediction values ​​for each of the target buildings:

[0070] (2)

[0071] In formula (2), Indicates the first One prediction step size, Indicates the first The target building Indicates the predicted time. This indicates the total number of expert models in the preset scenario expert library. This can be represented as: the first in the preset scene expert database. The expert model for the first Expert fusion weighting for generating target buildings. This can be represented as: the first in the preset scene expert database. The expert model for the first Candidate load prediction sequences generated for the target building. It can be expressed as: for the first The load forecast values ​​generated for the target building. This processing method allows different expert models to dominate at different times in the future. For example, holiday experts can have increased weight during weekday rest periods, high-temperature experts can have increased weight during the afternoon load peak period, and public event experts can have increased weight during the start or end of commercial activities.

[0072] For example, by using steps 202 to 210, short-term multi-building load forecasting based on multi-source event perception and dynamic expert fusion using a large language model can be achieved, which is applicable to the following application scenarios:

[0073] 1. Office building scenario: When retrieving texts such as holiday work schedule adjustments, typhoon warnings, or high temperature warnings, the system combines the office building metadata to analyze the effect of the event on the workload during office hours, and increases the weight of holiday experts, high temperature experts, or extreme weather experts in the relevant hours.

[0074] 2. Shopping mall scenario: When large-scale commercial events, holiday traffic changes, or extreme weather warnings are detected, the system combines commercial building metadata to analyze the impact of the events on afternoon and evening load peaks, and increases the prediction weight of the corresponding prediction step size through public event experts or high temperature experts.

[0075] 3. Unified prediction scenario for urban building clusters: The server trains a global model for multiple buildings in a unified manner, while retaining cross-building heterogeneity through building metadata, enabling the system to improve its generalization ability on individual buildings with limited sample size by leveraging building cluster data.

[0076] The aforementioned multi-building short-term load forecasting method first acquires unstructured public event natural language data, as well as building metadata and structured multi-source load forecasting base data for each target building. Hourly alignment and robust preprocessing are then performed on the multi-source load forecasting base data and building metadata to obtain aligned preprocessed data. Load forecasting-related text retrieval is then performed based on the public event natural language data to obtain retrieval text. Next, the retrieval text, public event natural language data, and forecast time are input into a preset parsing model to parse event records. Expert weight priors are generated based on these event records. Further, through a context encoder and a horizontal dynamic gating mechanism, expert fusion weights are generated based on the expert weight priors, aligned preprocessed data, and event records. Then, based on these expert fusion weights and multiple candidate load forecasting sequences generated from a preset scenario expert database, load forecast values ​​for multiple target buildings are generated. This entire process effectively integrates public event natural language data, building metadata, and structured multi-source load forecasting base data. By using a preset scenario expert database and horizontal dynamic gating, the prediction accuracy, robustness, and interpretability of building loads under various scenarios are improved, thereby effectively enhancing the accuracy of building load forecasting.

[0077] In one exemplary embodiment, such as Figure 3 The diagram illustrates the process of implementing hourly alignment and robust preprocessing. Specifically, the multi-source load forecasting base data includes meteorological data, calendar data, and historical load data. Hourly alignment and robust preprocessing are performed on the multi-source load forecasting base data and building metadata to obtain aligned preprocessed data, including:

[0078] Step 302: Resample at least a portion of the meteorological data that is not evenly spaced after being sorted by timestamp to the hourly grid, and perform interpolation and filling processing on at least a portion of the meteorological data with missing meteorological values ​​to obtain usable meteorological data.

[0079] The hourly grid refers to dividing the time axis into hourly intervals to form a regular time series. Non-uniform interval resampling to the hourly grid aims to: unify data frequency, align data from different times and sources; improve time resolution by refining the data to hourly intervals; and ensure data consistency. Interpolation and padding methods include, but are not limited to: linear interpolation, spline interpolation, time proximity padding, and spatial interpolation.

[0080] Optionally, the server can sort the meteorological data in the multi-source basic data of load forecasting by timestamp, then resample at least a portion of the non-equal interval meteorological data to the hourly grid, and perform time-based interpolation processing on at least a portion of the meteorological data with missing meteorological values, and perform forward and backward filling processing when necessary, so that the meteorological data sequence corresponds one-to-one with the hourly historical load data sequence to obtain usable meteorological data.

[0081] Step 304: Construct hourly time stamps for the calendar data, and then stitch together the calendar data with constructed hourly time stamps, historical load data, and available meteorological data to form hourly aligned data.

[0082] Optionally, the server can construct features such as hour markers, weekday markers, year-end day number markers, and weekend markers for calendar variables in the calendar data, and concatenate the calendar data with constructed hour-level time markers with historical load data and available meteorological data on the same hour index to form a structured input sequence, i.e., hour-level aligned data.

[0083] For example, constructing an hour-level time stamp may specifically include: (1) extracting the "hour" part of the time, such as 0 to 23, to construct an hour stamp; (2) extracting the "day of the week" of the time, such as 0 to 6, where 0 represents Monday and 6 represents Sunday, to construct a day of the week stamp; (3) extracting the sequence number of each day in the year, such as 1 to 366, to construct a day sequence number stamp within the year; and (4) extracting binary features, such as 1 for weekends and 0 for weekdays, to construct a weekend stamp.

[0084] Step 306: Perform robust preprocessing on the historical load data and the abnormal data in the building metadata to obtain robust preprocessed data.

[0085] Optionally, for historical load data, the server can use Hampel filtering (a robust anomaly detection method based on local median and dispersion to identify outliers) to detect abnormal load spikes, and then perform pruning using a building-level high quantile threshold; it can also interpolate short missing or invalid segments, and generate training masks for long zero-value or long invalid segments to prevent anomaly segments from directly participating in loss calculations. For building metadata, the server can normalize each target building based on preset training set statistics, and fill in missing numerical building metadata according to the building type median; for categorical building metadata, it can use one-hot vectors or learnable embedding vectors. Through these methods, robust preprocessed data is obtained, improving the stability of subsequent model prediction processes and the reliability of prediction results.

[0086] Step 308: Combine hourly aligned data and robust preprocessed data to obtain aligned preprocessed data.

[0087] In this embodiment, by performing hourly alignment and robust preprocessing on the multi-source load forecasting base data, the chaotic and multi-source load forecasting base data can be processed into regular, aligned preprocessed data containing rich and effective information, which is beneficial to the accuracy and robustness of subsequent building load forecasting.

[0088] In one embodiment, load prediction-related text retrieval is performed based on public event natural language data to obtain the retrieved text, including:

[0089] The correlation assessment results are obtained by performing target building correlation assessment, prediction time correlation assessment, and load change correlation assessment on the natural language data of public events.

[0090] Based on the relevance assessment results, the natural language data of public events is filtered to obtain the retrieval text.

[0091] Specifically, the server can use public event natural language data to predict the timing of events. Constructing a candidate common text set within the previous scrolling time window The sources of natural language data for public events include weather warnings, public holiday notices, municipal announcements, and announcements of major events. Furthermore, as shown in formula (3), the server can target candidate public text sets... Conduct target building correlation assessment, forecast time correlation assessment, and load change correlation assessment, and then... Candidate texts in the candidate public text set Sorting yields the correlation assessment results:

[0092] (3)

[0093] In formula (3), Used for time-relevance assessment during prediction, representing candidate text. Text release time and prediction time Timeliness relevance; Used for target building relevance assessment, representing candidate text. The text involves locations and the first Spatial relevance of the target building's location or its surrounding area; It is also used for target building relevance assessment, representing candidate texts. The text content and the first Building metadata of the target building The degree of matching between building type, business model, or usage attributes; Used for load change correlation assessment, representing candidate text. The correlation between text keywords and semantic content and load changes; The scoring weights can be set manually or adjusted through validation set optimization, learning a ranking model, or reinforcement learning strategies.

[0094] Furthermore, the server can select candidate texts based on the relevance assessment results. The candidate texts are sorted (e.g., by score from highest to lowest). Filter the data and retain the top-M texts to form the search results. It transforms the originally variable-length and complex natural language data into standardized search texts of a fixed size (M items).

[0095] In this embodiment, key information can be actively retrieved on demand, irrelevant public information can be filtered, and target building relevance assessment, prediction time relevance assessment, and load change relevance assessment can be introduced. This allows different text inputs to be generated for the same city-level event on office buildings and commercial buildings, improving the intelligent perception capability of the "differential impact" of the same city-level event on different target buildings and improving the accuracy of building load prediction.

[0096] In some embodiments, generating expert weight priors based on event logs includes:

[0097] Event records are encoded into event vectors, and event vectors are aggregated into a unified event representation using a preset weighted pooling method.

[0098] By using a preset activation function, the unified event representation is processed into expert weighted priors.

[0099] Specifically, the server can encode event records into event vectors and aggregate the event vectors into a unified event representation using a preset weighted pooling method (such as weighted average, attention pooling, or confidence-weighted pooling).

[0100] (4)

[0101] In formula (4), Indicates that the event is logged Encoded event vectors, Indicates the first The target building at the predicted time The corresponding unified event representation, This indicates weighted pooling processing. This represents the total number of target buildings. Event vectors can be aggregated using methods such as weighted average, max pooling, attention pooling, confidence-weighted pooling, or graph attention aggregation.

[0102] Furthermore, the server can use the preset activation function shown in formula (5). To unify event representation Processed as expert weight prior :

[0103] (5)

[0104] In formula (5), , Indicates the expert library for preset scenarios The number of expert models in China and These are trainable or preset parameters. Indicates the first The target building at the predicted time The corresponding expert weight priors serve only as guiding signals for expert fusion weights and do not directly replace the numerical prediction model. It is a normalization function that can transform a set of real numbers into a non-negative probability distribution that sums to 1.

[0105] In this embodiment, corresponding expert weight priors can be generated for each target building. Combined with the previous filtering of different search texts for different target buildings, the building-level personalization of expert weight priors can be further realized, which is conducive to distinguishing the differentiated impact of the same event on different buildings such as office buildings and shopping malls.

[0106] In one possible implementation, expert fusion weights are generated based on expert weight priors, aligned preprocessed data, and event logs, using a context encoder and a horizontal dynamic gating mechanism. These weights include:

[0107] The context encoder extracts the context representation from the aligned preprocessed data.

[0108] Based on contextual representation, prior expert weights, and unified event representation generated from event records, expert fusion weights are generated through a horizontal dynamic gating mechanism.

[0109] Among them, horizontal-wise gating is used for predicting the horizon. A gating mechanism is used to calculate expert fusion weights for each future prediction step. Horizontal dynamic gating can generate expert fusion weights for each future prediction step, which can characterize the changes in the impact of an event over 24 hours. Unlike fixed-weight fusion methods, this embodiment can improve the dynamic adaptability within the prediction horizon.

[0110] Alternatively, the server can use a context encoder. From aligned preprocessed data (including historical load data sequences that have undergone hourly alignment and robust preprocessing) Structured exogenous feature sequences Building metadata sequence Extract the first The target building at the predicted time Corresponding context representation , Then, the context representation Expert weight prior Unified event representation generated based on event logs The matrix is ​​concatenated and input into a multilayer perceptron to obtain the expert score matrix. , , , This indicates the number of expert models in the preset scenario expert library. This indicates the predicted horizon, which can be set to 12, 24, 48 hours, etc., according to business needs.

[0111] Furthermore, as shown in formula (6), the server can perform predictions along the expert dimension for each prediction step. Normalization yields the horizontal dynamic fusion weights, also known as expert fusion weights:

[0112] (6)

[0113] In formula (6), Indicates the first in the predicted horizon One prediction step size, Indicates the first The target building This indicates the total number of expert models in the preset scenario expert library. It is an exponential function. Indicates: The first in the preset scene expert database The expert model for the first Expert fusion weighting for generating target buildings.

[0114] In this embodiment, a deep fusion of multi-source heterogeneous information can be achieved through a context encoder and a horizontal dynamic gating mechanism, providing a highly feasible architectural design for high-precision building load prediction.

[0115] In one exemplary embodiment, multiple candidate load prediction sequences are generated in the following manner:

[0116] Acquire multiple expert models and construct a pre-defined expert database based on these models;

[0117] The aligned preprocessed data is input into a preset scenario expert library to obtain multiple candidate load prediction sequences.

[0118] The various expert models include at least: general experts, holiday experts, high-temperature experts, extreme weather experts, and public event experts. During the training process of the expert models, general experts learn the regular load patterns in all samples; while special expert models such as holiday experts, high-temperature experts, extreme weather experts, and public event experts, which are applicable to different scenarios, learn the local load distribution of samples under their respective event categories.

[0119] Optionally, a pre-defined scenario expert library can be built on the server based on multiple expert models, and multiple candidate load prediction sequences can be output through the pre-defined scenario expert library, as shown in formula (7):

[0120] (7)

[0121] In formula (7), the first An expert model Receive historical load data sequence Structured exogenous feature sequences and building metadata sequence Output candidate prediction sequences .

[0122] For example, each expert model can use a BiLSTM numerical prediction backbone with a hidden dimension of 64, a recurrent layer of 2, and an input window of 48 hours. To ensure the diversity among expert models, each expert model can be trained using different scenario sample distributions or different sample weights, while maintaining a consistent network backbone structure. In other implementations, the backbone network of each scenario expert model can be replaced by LSTM, GRU, TCN, Transformer, DLINear, PatchTST, XGBoost, or other time-series prediction models, instead of BiLSTM (Bidirectional Long Short-Term Memory, used to extract time-dependent features from historical sequences).

[0123] In this embodiment, a preset scenario expert library containing multiple expert models can be constructed to ensure that no matter how the load curve changes, at least one expert model can provide a relatively accurate baseline prediction, laying a solid "candidate pool" foundation for subsequent fusion.

[0124] In a specific application scenario, based on the above embodiments, the following scheme can be adopted to achieve the training objective and deployment output of a short-term multi-building load forecasting system based on large language model multi-source event perception and dynamic expert fusion:

[0125] 1. The training, validation, and test sets are divided according to the building's construction time sequence to avoid time leakage caused by random window partitioning. Optionally, the first 70% of the annual sequence for each building is used for training, the middle 10% for validation, and the last 20% for testing, with training samples generated based on a sliding window.

[0126] 2. The mean absolute error used for the main prediction loss Its general mathematical form is shown in formula (8):

[0127] (8)

[0128] In formula (8), Indicating the first [horizontal measurement] in the horizon One prediction step size, Indicates the first The target building Indicates the predicted time. For the first The load forecast for the target building. For the first The actual historical load value of the target building. It can represent the total sample.

[0129] 3. To enhance the consistency between the learned expert fusion weights and the prior expert weights, the average expert weights for predicting the horizon are calculated. As shown in formula (9):

[0130] (9)

[0131] In formula (9), This can be represented as: the first in the preset scene expert library The expert model for the first The expert fusion weights for generating target buildings are further introduced. Prior regularization terms are also introduced. As shown in formula (10):

[0132] (10)

[0133] In formula (10), It is a loss function used to measure the difference between two probability distributions. This represents the expert weight prior. The prior regularization term can be replaced by KL divergence with cross-entropy, mean squared error, or a temperature-scaled distance function. KL divergence is a metric that measures the difference between two probability distributions and is used to ensure that the learned expert weight prior is consistent with the LLM prior.

[0134] 4. The calculation method for the total training objective is shown in formula (11):

[0135] (11)

[0136] In formula (11), These are the prior regularization weights. After training, the short-term multi-building load forecasting system based on large language model multi-source event perception and dynamic expert fusion can receive the latest 48-hour historical load and available public text during the inference phase, and output the load forecast sequence for each target building in the next 24 hours, the expert fusion weights for each forecast step, the main event categories and event confidence levels, for use by building energy management, demand response scheduling and park energy consumption monitoring systems.

[0137] Once deployed, the short-term multi-building load forecasting system based on multi-source event perception and dynamic expert fusion using a large language model can transform natural language events such as weather warnings, holiday notices, municipal announcements, and public events into computable event signals through hourly data alignment, building-related public text retrieval, large language model event parsing, event vector and expert weight prior generation, scene expert database training, and prediction horizon-level dynamic gating fusion. This transforms these events into computable event signals, which are then used in conjunction with historical loads, meteorological variables, calendar variables, and building metadata for 24-hour proactive load forecasting. Compared to existing forecasting schemes that rely solely on structured inputs or static integration, this system does not directly output load values ​​using the large language model but instead utilizes it for event understanding and expert prior generation. Simultaneously, by driving text retrieval, event parsing, and expert weight allocation through building metadata, the same external event can generate differentiated interpretations across different building types. Furthermore, horizontal dynamic gating allows different experts to dominate at different future time periods, thus adapting to non-stationary load patterns such as holidays, extreme weather, and public events.

[0138] In a specific application scenario, using the above system can reduce MAE (Mean Absolute Error, which measures the average absolute deviation between the predicted and actual values), RMSE (Root Mean Squared Error, which measures the square root of the mean of the squared prediction error), and MAPE (Mean Absolute Percentage Error, which measures the relative prediction error) by approximately 13.86%, 13.79%, and 14.40%, respectively. Furthermore, it maintains a stable improvement in office buildings, shopping malls, and holiday-related windows, effectively achieving short-term multi-building load forecasting.

[0139] like Figure 4As shown, based on the above embodiments, a short-term multi-building load forecasting system based on large language model multi-source event perception and dynamic expert fusion is provided, which realizes load forecasting through the following process:

[0140] I. Multi-source data acquisition and hourly alignment.

[0141] The system receives unstructured natural language data of public events (i.e., a public text pool containing data from sources such as warnings, holidays, and events), as well as building metadata (including building type, area, cooling attributes, etc.) for multiple target buildings and structured multi-source load forecasting base data. This multi-source load forecasting base data includes: meteorological data (obtained from meteorological observations), calendar data (containing calendar features such as hourly, weekly, yearly, and weekend data), and historical load data (48-hour window). The multi-source load forecasting base data and building metadata undergo hourly alignment and robust preprocessing to obtain aligned preprocessed data.

[0142] II. Public text retrieval related to architecture and time.

[0143] Load prediction-related text retrieval is performed based on natural language data of public events to obtain the retrieved text.

[0144] III. Event parsing and prior generation in large language models.

[0145] The retrieved text, public event natural language data, and predicted time are input into a pre-defined parsing model to extract event records. Expert weight priors are then generated based on these event records. Through a context encoder and a horizontal dynamic gating mechanism, expert fusion weights are generated based on the expert weight priors, aligned preprocessed data, and event records.

[0146] IV. Predicting Horizon-Level Dynamic Expert Fusion.

[0147] The system acquires multiple candidate load forecast sequences generated based on a pre-defined scenario expert database, and generates load forecast values ​​for multiple target buildings based on the expert fusion weights and each candidate load forecast sequence. The pre-defined scenario expert database includes various expert models, such as general experts, holiday experts, high-temperature experts, extreme weather experts, and public event experts.

[0148] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0149] Based on the same inventive concept, this application also provides a multi-building short-term load forecasting device for implementing the multi-building short-term load forecasting method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the multi-building short-term load forecasting device provided below can be found in the limitations of the multi-building short-term load forecasting method described above, and will not be repeated here.

[0150] In one exemplary embodiment, such as Figure 5 As shown, a multi-building short-term load forecasting device is provided, including: a multi-source data acquisition module 502, a preprocessing and text retrieval module 504, a weight prior generation module 506, an expert fusion weight generation module 508, and a load forecasting module 510, wherein:

[0151] The multi-source data acquisition module is used to acquire unstructured natural language data of public events, as well as building metadata of multiple target buildings and structured multi-source basic data for load forecasting.

[0152] The preprocessing and text retrieval module is used to perform hourly alignment and robust preprocessing on multi-source basic data and building metadata for load forecasting to obtain aligned preprocessed data, and to perform load forecasting-related text retrieval based on public event natural language data to obtain the retrieval text.

[0153] The weight prior generation module is used to input the retrieved text, public event natural language data, and prediction time into the preset parsing model, parse out the event records, and generate expert weight priors based on the event records.

[0154] The expert fusion weight generation module is used to generate expert fusion weights based on expert weight priors, aligned preprocessed data, and event records through a context encoder and a horizontal dynamic gating mechanism.

[0155] The load forecasting module is used to acquire multiple candidate load forecasting sequences generated based on a preset scenario expert database, and generate load forecasting values ​​for multiple target buildings based on the expert fusion weights and each candidate load forecasting sequence.

[0156] The aforementioned multi-building short-term load forecasting device first acquires unstructured public event natural language data, as well as building metadata and structured multi-source load forecasting basic data for each target building. It then performs hourly alignment and robust preprocessing on the multi-source load forecasting basic data and building metadata to obtain aligned preprocessed data. Based on the public event natural language data, it performs load forecasting-related text retrieval to obtain the retrieval text. Next, it inputs the retrieval text, public event natural language data, and forecast time into a preset parsing model to parse event records and generate expert weight priors based on these records. Further, through a context encoder and a horizontal dynamic gating mechanism, it generates expert fusion weights based on the expert weight priors, aligned preprocessed data, and event records. Finally, based on these expert fusion weights and multiple candidate load forecasting sequences generated from a preset scenario expert database, it generates load forecast values ​​for multiple target buildings. This entire process effectively integrates public event natural language data, building metadata, and structured multi-source load forecasting basic data. Through the preset scenario expert database and horizontal dynamic gating, it improves the prediction accuracy, robustness, and interpretability of building loads under various scenarios, thereby effectively improving the accuracy of building load forecasting.

[0157] In one embodiment, the multi-source basic data for load forecasting includes: meteorological data, calendar data, and historical load data. The preprocessing and text retrieval module includes:

[0158] The resampling interpolation and filling unit is used to resample at least a portion of the meteorological data that is not evenly spaced after being sorted by timestamp to the hourly grid, and to perform interpolation and filling processing on at least a portion of the meteorological data with missing meteorological values ​​to obtain usable meteorological data.

[0159] The hourly alignment processing unit is used to construct hourly time stamps for calendar data and to stitch together the calendar data with constructed hourly time stamps, historical load data, and available meteorological data to form hourly aligned data.

[0160] The robust preprocessing unit is used to robustly preprocess abnormal data in historical load data and building metadata to obtain robust preprocessed data.

[0161] The data aggregation unit is used to aggregate hourly aligned data and robust preprocessed data to obtain aligned preprocessed data.

[0162] In one embodiment, the preprocessing and text retrieval module includes:

[0163] The correlation assessment unit is used to perform correlation assessments on target buildings, prediction time, and load change in natural language data of public events, and obtain correlation assessment results.

[0164] The text retrieval filtering unit is used to filter natural language data of public events based on the relevance assessment results to obtain the retrieval text.

[0165] In one embodiment, the weight prior generation module includes:

[0166] The unified event representation acquisition unit is used to encode event records into event vectors and aggregate the event vectors into a unified event representation using a preset weighted pooling method.

[0167] The expert weight prior acquisition unit is used to process the unified event representation into expert weight prior through a preset activation function.

[0168] In one embodiment, the expert fusion weight generation module includes:

[0169] The context representation extraction unit is used to extract the context representation from the aligned preprocessed data through the context encoder;

[0170] The expert fusion weight generation unit is used to generate expert fusion weights based on context representation, expert weight priors, and unified event representation generated based on event records, through a horizontal dynamic gating mechanism.

[0171] In one embodiment, multi-building short-term load forecasting includes a candidate load generation module, which comprises:

[0172] The scenario expert database construction unit is used to acquire multiple expert models and construct a preset scenario expert database based on these models.

[0173] The candidate load generation unit is used to input aligned preprocessed data into a preset scenario expert library to obtain multiple candidate load prediction sequences.

[0174] Each module in the aforementioned multi-building short-term load forecasting device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0175] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 6As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores short-term load forecasting data for multiple buildings. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a short-term load forecasting method for multiple buildings.

[0176] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0177] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0178] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0179] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0180] It should be noted that the information (including but not limited to device information, personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0181] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, database, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0182] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0183] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for short-term load forecasting of multiple buildings, characterized in that, The method includes: Acquire unstructured natural language data of public events, as well as building metadata for multiple target buildings and structured multi-source basic data for load forecasting; The load forecast multi-source basic data and the building metadata are aligned and robustly preprocessed at the hour level to obtain aligned preprocessed data. Load forecast related text is retrieved based on the public event natural language data to obtain the retrieved text. The search text, the public event natural language data, and the predicted time are input into a preset parsing model to parse the event records, and expert weight priors are generated based on the event records. Expert fusion weights are generated based on the expert weight priors, the alignment preprocessing data, and the event records, using a context encoder and a horizontal dynamic gating mechanism. Multiple candidate load prediction sequences generated based on a preset scenario expert database are obtained, and load prediction values ​​for multiple target buildings are generated according to the expert fusion weights and each candidate load prediction sequence.

2. The method according to claim 1, characterized in that, The multi-source basic data for load forecasting includes: meteorological data, calendar data, and historical load data; the hourly alignment and robust preprocessing of the multi-source basic data for load forecasting and the building metadata to obtain aligned preprocessed data includes: At least a portion of the meteorological data that is not evenly spaced after being sorted by timestamp is resampled to an hourly grid, and at least a portion of the meteorological data with missing meteorological values ​​is interpolated and filled to obtain usable meteorological data. Hour-level time stamps are constructed on the calendar data, and the calendar data with constructed hour-level time stamps, the historical load data, and the available meteorological data are spliced ​​together to form hour-level aligned data; Robust preprocessing is performed on the historical load data and the abnormal data in the building metadata to obtain robust preprocessed data; The hourly aligned data and the robust preprocessed data are combined to obtain the aligned preprocessed data.

3. The method according to claim 1, characterized in that, The load prediction-related text retrieval based on the public event natural language data yields the retrieved text, including: The natural language data of the public events is used to perform target building correlation assessment, prediction time correlation assessment, and load change correlation assessment to obtain correlation assessment results. Based on the relevance assessment results, the natural language data of the public events is filtered to obtain the search text.

4. The method according to claim 1, characterized in that, The generation of expert weight priors based on the event records includes: The event records are encoded into event vectors, and the event vectors are aggregated into a unified event representation using a preset weighted pooling method. The unified event representation is processed into the expert weight prior by using a preset activation function.

5. The method according to claim 1, characterized in that, The process of generating expert fusion weights through a context encoder and a horizontal dynamic gating mechanism, based on the expert weight priors, the aligned preprocessed data, and the event records, includes: The context encoder extracts the context representation from the aligned preprocessed data. Based on the context representation, the expert weight prior, and the unified event representation generated based on the event records, the expert fusion weights are generated through the horizontal dynamic gating mechanism.

6. The method according to claim 1, characterized in that, The multiple candidate load prediction sequences are generated in the following manner: Acquire multiple expert models, and construct a pre-defined scenario expert database based on these expert models; The aligned preprocessed data is input into the preset scenario expert library to obtain multiple candidate load prediction sequences.

7. A multi-building short-term load forecasting device, characterized in that, The device includes: The multi-source data acquisition module is used to acquire unstructured natural language data of public events, as well as building metadata of multiple target buildings and structured multi-source basic data for load forecasting. The preprocessing and text retrieval module is used to perform hourly alignment processing and robust preprocessing on the multi-source basic data of load forecasting and the building metadata to obtain aligned preprocessed data, and to perform load forecasting-related text retrieval based on the public event natural language data to obtain the retrieval text; The weight prior generation module is used to input the search text, the public event natural language data and the prediction time into a preset parsing model, parse out the event records, and generate expert weight priors based on the event records. The expert fusion weight generation module is used to generate expert fusion weights based on the expert weight priors, the alignment preprocessing data, and the event records through a context encoder and a horizontal dynamic gating mechanism. The load forecasting module is used to acquire multiple candidate load forecasting sequences generated based on a preset scenario expert database, and generate multiple load forecasting values ​​for the target buildings according to the expert fusion weights and each candidate load forecasting sequence.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images