Power supply and demand prediction method and system based on multi-model cooperation

Abstract

The invention relates to a power supply and demand prediction method and system based on multi-model cooperation. The method comprises the steps of obtaining multi-source data corresponding to a current prediction period, wherein the multi-source data at least comprises power market structured data and unstructured text data related to power operation; processing the structured data to form basic time sequence features; calling a large language model customized in the power field to analyze the unstructured text data, and mapping the unstructured text data into structured event features; performing time sequence alignment and fusion on the basic time sequence feature and the structured event feature to obtain a multi-modal fusion feature vector; inputting the vector into prediction models corresponding to at least two different preset scenes to obtain power supply and demand prediction sub-results; and representing the current market scene by using the vector, calculating the scene similarity between the current market scene and each preset scene in the historical scene feature library, endowing each prediction sub-result with a normalized weight according to the scene similarity, carrying out weighted fusion, and outputting a final power supply and demand prediction result.

Description

Technical Field

[0001] This application relates to the field of power supply and demand forecasting technology, and in particular to a power supply and demand forecasting method and system based on multi-model collaboration. Background Technology

[0002] With the ongoing reform of the power market, the operation of the power system is gradually shifting from being primarily planned to being primarily based on market transactions and real-time dispatch. Especially in business scenarios such as the spot market and competitive bidding, the power supply and demand relationship not only determines the safe and stable operation of the power system but also directly affects market clearing results, trading strategy formulation, and risk management levels. Therefore, in the face of a short-cycle, highly volatile, and highly disruptive power market environment, building high-precision and robust power supply and demand forecasting capabilities has become an important requirement for power trading institutions, electricity sales companies, and dispatching and operation departments.

[0003] Existing power supply and demand forecasting technologies typically use structured power market data as primary input, such as load curves, generating unit output, renewable energy output, maintenance plans, meteorological data, and transaction data, and employ statistical models or machine learning / deep learning models for time-series forecasting. However, the formation mechanism of power supply and demand is complex, significantly influenced by the coupling of multiple factors, and a large amount of key disturbance information is generated and disseminated in unstructured form during market operation, such as policy notices, market announcements, maintenance notices, weather warnings, emergency reports, and public opinion information. Because unstructured information cannot be directly incorporated into traditional forecasting models, the forecasting system lacks sensitivity to typical scenarios such as extreme weather, holiday effects, overlapping maintenance, and sudden changes in renewable energy output. Forecasting biases are more pronounced during peak periods and extreme scenarios, making it difficult to meet the requirements for peak and extreme biases and cycle matching, such as bidding space. Summary of the Invention

[0004] To overcome, to some extent, the problems of poor generalization of power supply and demand forecasting models in related technologies and the limited forecasting ability for special scenarios due to insufficient integration of multi-source heterogeneous data, this application provides a power supply and demand forecasting method and system based on multi-model collaboration.

[0005] The proposed solution is as follows: According to a first aspect of the embodiments of this application, a power supply and demand forecasting method based on multi-model collaboration is provided, comprising: Acquire multi-source data corresponding to the current forecast period, wherein the multi-source data includes at least: structured data of the power market and unstructured text data related to power operation; The structured data is subjected to timestamp alignment, missing data handling, and normalization to form basic time-series features; A customized large language model for the power industry is invoked to perform semantic parsing on the unstructured text data, identify the types of events, affected objects, and intensity of impact related to power supply and demand, and map the parsing results into structured event features; The basic temporal features and the structured event features are temporally aligned and fused to obtain the multimodal fusion feature vector for the current prediction period; The multimodal fusion feature vector is input into at least two prediction models corresponding to different preset scenarios to obtain the power supply and demand prediction sub-results output by each prediction model; the prediction models include at least: a baseline scenario prediction model, an extreme weather scenario prediction model, a holiday scenario prediction model, a maintenance scenario prediction model, and a high proportion of new energy scenario prediction model. The current market scenario is represented by the multimodal fusion feature vector, and the scenario similarity between the current market scenario and each preset scenario in the historical scenario feature library is calculated; the preset scenario includes multimodal fusion feature vectors from multiple historical prediction periods; The power supply and demand prediction sub-results output by each prediction model are weighted according to the scene similarity between the current market scenario and each preset scenario in the historical scenario feature library; the weights are obtained by normalizing the scene similarity and the sum of the weights is 1. The power supply and demand forecast sub-results are weighted and fused to obtain the final power supply and demand forecast result; the final power supply and demand forecast result includes at least the total power demand forecast and the total power supply forecast for the current forecast period.

[0006] Preferably, the method further includes: After at least one forecast period, the actual observed total electricity demand and total electricity supply values ​​are obtained, and the corresponding actual supply and demand results are generated. Based on a pre-constructed comprehensive evaluation index system for adapting power supply and demand forecasting, the final power supply and demand forecast results for each forecast period are compared with the actual supply and demand results on a periodic basis, and the index values ​​are calculated. The comprehensive evaluation index system includes conventional time-series forecasting indicators and special indicators. The conventional time-series forecasting indicators include: mean absolute error index and root mean square error index. The special indicators are used to target the characteristics of bidding space and include: peak error index, extreme scenario deviation index and period matching degree index. The prediction results are judged to be up to standard based on the index values ​​and dynamic error thresholds for each prediction period. The forecast period with excessive output error and the corresponding market scenario for that forecast period.

[0007] Preferably, the method further includes: An error source analysis is performed on the prediction period where the error exceeds the standard; the error source analysis includes: Verify the feature dimension integrity, timestamp consistency, and update timeliness of the multimodal fusion feature vector to pinpoint the sources of error at the data level; The process of multi-model collaborative prediction is broken down to evaluate the accuracy of market scenario identification, the adaptability and parameter stability of each prediction model, and to pinpoint the sources of error at the model level. Obtain electricity market operating conditions, conduct external disturbance analysis based on electricity market operating conditions, and locate the sources of external disturbance errors; The source conclusions for generating and predicting error exceedance cycles are based on data-level error sources, model-level error sources, and external disturbance error sources; the source conclusions include at least the error source category and the optimization target object.

[0008] Preferably, the method further includes: Iterative updates are performed based on the aforementioned source tracing conclusions; including: When the error source category is a data-level error source, update the transformation rules from unstructured text to structured event features, as well as the spatiotemporal alignment threshold between basic time-series features and structured event features; When the error source category is a model-level error source, the mapping relationship between the preset scenario and the prediction model is updated, the model parameters of the prediction model are optimized, and incremental training is performed on the prediction model when the incremental training trigger condition is met; wherein, the model parameters include at least one or more of tree model depth and learning rate; When the error source category is an external disturbance error source, the historical scene feature library used for market scene identification is updated to increase scene features corresponding to the external disturbance and improve the ability to identify new market disturbances. After the update is completed, the updated transformation rules, spatiotemporal alignment thresholds, mapping relationships, model parameters, and historical scene feature libraries are called in subsequent prediction cycles to perform predictions. Differential optimization instructions, update content, and update effects are recorded to form an optimization log for long-term iteration.

[0009] Preferably, the step of invoking a large language model customized for the power industry to perform semantic parsing on the unstructured text data includes: Events are extracted from the unstructured text data to obtain a set of event elements related to power supply and demand; the set of event elements includes at least: event type, affected object, and impact intensity; wherein, the event type includes: extreme weather events, holiday events, maintenance events, and abnormal output events of new energy sources; The event element set is mapped into structured event features based on preset feature mapping rules; the structured event features include at least: the event occurrence time, duration, and quantified value of the impact intensity associated with the affected object; When there are multiple source conflicts in the unstructured text data, the structured event features are updated based on the data source credibility or consistency check structure.

[0010] Preferably, the step of aligning and fusing the basic temporal features with the structured event features in a temporal sequence includes: The structured event features are mapped to a time axis consistent with the basic time series features according to a preset time granularity, and time window slicing is performed on the structured event features whose start and end times span multiple prediction periods. The basic temporal features and the aligned structured event features are subjected to one or more of the following: feature concatenation, gating fusion, or attention fusion, to obtain the multimodal fusion feature vector.

[0011] Preferably, the method further includes: The scene similarity is calculated using one or more of cosine similarity, Euclidean distance similarity, or Mahalanobis distance similarity. Based on the similarity between the current market scenario and the historical scenario feature library, the top K preset scenarios are selected, and weights are assigned only to the power supply and demand prediction sub-results output by the prediction model corresponding to the top K preset scenarios; where K is an integer greater than or equal to 2.

[0012] Preferably, the method further includes: The supply-demand gap forecast is calculated based on the total electricity demand forecast and the total electricity supply forecast. The predicted supply-demand gap is output as part of the final electricity supply-demand forecast result.

[0013] Preferably, the construction of the dynamic error threshold includes: A rolling evaluation window is constructed based on the indicator values ​​of historical prediction periods, and the statistics of the indicator values ​​within the rolling evaluation window are calculated. The statistics include at least one or more of the following: mean, standard deviation, and quantiles. Based on the statistical measures, dynamic error thresholds are determined for the conventional time-series forecast indicators and special indicators, respectively, and the dynamic error thresholds are adaptively adjusted when the power market conditions change. The adjusted dynamic error threshold is used to determine whether the indicator values ​​in subsequent prediction periods meet the standards.

[0014] According to a second aspect of the embodiments of this application, a power supply and demand forecasting system based on multi-model collaboration is provided, comprising: Processor and memory; The processor and memory are connected via a communication bus: The processor is used to call and execute the program stored in the memory; The memory is used to store a program, which is at least used to execute a power supply and demand forecasting method based on multi-model collaboration as described in any of the above.

[0015] The technical solution provided in this application may include the following beneficial effects: This technical solution collaboratively utilizes structured data from the electricity market with unstructured textual information about electricity operation: A large language model extracts and quantifies event types, affected objects, and impact intensity to form structured event features, which are then aligned and fused with basic time-series features to obtain multimodal feature vectors. Based on similarity to historical scenario feature libraries, the outputs of different preset scenario prediction models are normalized and weighted, thereby achieving adaptive modeling and multi-model complementarity for the current market scenario. This improves the accuracy and stability of supply and demand forecasts in complex scenarios such as extreme weather, holidays, overlapping maintenance, and high proportion of renewable energy, reduces prediction bias during peak periods and extreme scenarios, and outputs more reliable total electricity demand and supply forecasts.

[0016] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] Figure 1 This is a flowchart illustrating a power supply and demand forecasting method based on multi-model collaboration, provided in one embodiment of this application. Detailed Implementation

[0019] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0020] Example 1 Traditional power supply and demand forecasting technologies primarily rely on structured data, resulting in insufficient sensitivity to typical scenarios triggered by unstructured information such as extreme weather, holiday effects, overlapping maintenance, and sudden changes in renewable energy output. This leads to significant deviations in forecasting systems during peak periods and under extreme conditions, making it difficult to meet the power market's requirements for high-precision and robust forecasting capabilities.

[0021] To address this, this embodiment provides a power supply and demand forecasting method based on multi-model collaboration, referring to... Figure 1 ,include: S1. Obtain multi-source data corresponding to the current forecast period; the multi-source data should include at least: structured data of the power market and unstructured text data related to power operation; S2. Perform timestamp alignment, missing data handling, and normalization on the structured data to form basic time series features; S3. Call the large language model customized in the power field to perform semantic parsing on unstructured text data, identify the event types, affected objects and impact intensity related to power supply and demand, and map the parsing results into structured event features; S4. Align and fuse the basic temporal features with the structured event features to obtain the multimodal fusion feature vector for the current prediction period; S5. Input the multimodal fusion feature vector into the prediction model corresponding to at least two different preset scenarios to obtain the power supply and demand prediction sub-results output by each prediction model; the prediction model includes at least: a baseline scenario prediction model, an extreme weather scenario prediction model, a holiday scenario prediction model, a maintenance scenario prediction model, and a high proportion of new energy scenario prediction model. S6. Represent the current market scenario with a multimodal fusion feature vector, and calculate the scenario similarity between the current market scenario and each preset scenario in the historical scenario feature library; the preset scenarios include multimodal fusion feature vectors from multiple historical prediction periods; S7. Assign weights to the power supply and demand prediction sub-results output by each prediction model based on the scene similarity between the current market scene and each preset scene in the feature library of historical scenes; the weights are obtained by normalizing the scene similarity and the sum of the weights is 1. S8. Weighted fusion of the sub-results of power supply and demand forecasts to obtain the final power supply and demand forecast result; the final power supply and demand forecast result shall include at least the forecast value of total power demand and the forecast value of total power supply in the current forecast period.

[0022] For ease of understanding, the following explains some key terms in this embodiment: Multi-source data refers to a collection of various types of data used for electricity supply and demand forecasting. It covers all aspects of electricity market operation, including but not limited to information on market transactions, operating status, and external environment.

[0023] Structured data refers to data with a well-defined data model, stored in tabular form, and easy to query and analyze. In the power sector, it typically includes load curves, unit output, renewable energy output, maintenance plans, meteorological data, and transaction data.

[0024] Unstructured text data refers to text information that does not have a predefined data model or fixed format. In the power sector, this may include policy notices, market announcements, maintenance notices, weather warnings, emergency reports, and public opinion information, which are usually in natural language form.

[0025] A customized large language model for the power sector refers to a natural language processing model trained and optimized using specific power industry corpora. This model is capable of understanding power industry terminology, identifying power events, and analyzing the semantics of power-related texts. It can also extract and structure key information from unstructured texts.

[0026] Basic time-series features refer to the features extracted from structured data and arranged in a time sequence. These features, after timestamp alignment, missing data handling, and normalization, reflect the operating patterns and trends of the power system in the time dimension.

[0027] Structured event features refer to event information related to electricity supply and demand that is parsed and extracted from unstructured text data and transformed into a structured form that can be processed by models. These features typically include quantitative information such as event type, affected objects, impact intensity, and event occurrence time and duration.

[0028] A multimodal fusion feature vector is a comprehensive feature representation formed by aligning and fusing basic time-series features with structured event features. This vector can comprehensively characterize various influencing factors of power supply and demand within the current forecast period, including time-series trends and event disturbances.

[0029] A prediction model refers to an algorithm or model used to predict electricity supply and demand. In this embodiment, the prediction model is designed to be trained and optimized for different preset scenarios to improve prediction accuracy under specific market conditions.

[0030] Preset scenarios refer to representative market operation scenarios defined based on historical electricity market operation data and experience. These scenarios can include baseline scenarios, extreme weather scenarios, holiday scenarios, maintenance scenarios, and scenarios with a high proportion of renewable energy, etc. Each scenario corresponds to one or a set of specific market characteristics.

[0031] Scene similarity refers to the degree of similarity between the current market scene and various preset scenes in the historical scene feature library. This similarity is quantified by calculating the distance or correlation between multimodal fusion feature vectors, and is used to evaluate the matching degree between the current market environment and typical historical scenes.

[0032] A historical scene feature library is a database that stores multimodal fusion feature vectors from multiple historical prediction periods. This library is used to construct and identify different preset scenes and serves as a reference benchmark for calculating scene similarity.

[0033] The electricity supply and demand forecast sub-results refer to the electricity supply and demand forecast values ​​independently output by various forecasting models targeting different preset scenarios. These sub-results are weighted and fused in subsequent steps to form the final forecast result.

[0034] Weight refers to the relative importance coefficient assigned to each power supply and demand forecast sub-result during the weighted fusion process. This weight is dynamically adjusted based on the similarity between the current market scenario and the preset scenario to reflect the applicability of different forecasting models under current market conditions.

[0035] The final electricity supply and demand forecast refers to the comprehensive forecast of total electricity demand and total electricity supply for the current forecast period, obtained through weighted fusion. This result is the final output of this method and is used to guide electricity market trading and dispatch decisions.

[0036] This embodiment provides a power supply and demand forecasting method based on multi-model collaboration.

[0037] First, acquire multi-source data corresponding to the current forecast period. This multi-source data includes at least structured electricity market data and unstructured text data related to electricity operation. Data can be acquired manually, such as operators manually inputting electricity market transaction data and text information extracted from news reports. Alternatively, data can be acquired through simple file reading methods, such as reading structured data in CSV format and unstructured text data in TXT format from a preset local file path. As a preferred implementation, data can be acquired by connecting to external databases or data interfaces, for example, periodically pulling structured data from the electricity trading center's database and crawling unstructured text data from publicly available web platforms.

[0038] Furthermore, the structured data undergoes timestamp alignment, missing data handling, and normalization to form basic temporal characteristics. Timestamp alignment can be achieved through simple linear interpolation or by removing mismatched time points. Missing data handling can be done by filling missing data points with zero or average values. Normalization can employ methods such as max-min normalization or Z-score normalization. Specifically, during data preprocessing, a fixed time granularity can be set, such as every hour or every 15 minutes, and the timestamps of all structured data can be adjusted to that granularity. For missing data, forward or backward padding can be used. Normalization can be performed by selecting an appropriate scaling method based on the overall distribution characteristics of the data, such as scaling the data to the [0, 1] interval.

[0039] Subsequently, a customized large language model for the power sector is invoked to perform semantic parsing on the unstructured text data. This identifies event types, affected objects, and impact intensities related to power supply and demand, and maps the parsing results into structured event features. The large language model can be configured to receive unstructured text input and output raw text fragments containing event types, affected objects, and impact intensities. For typical textual information such as extreme weather warnings issued by meteorological departments and equipment maintenance notices issued by dispatch centers, the large model is invoked to perform deep semantic parsing. For example, the natural language description "Affected by strong cold air, temperatures have plummeted for several days, and photovoltaic output is expected to decrease significantly" can be automatically converted into structured event features such as "cold weather event," "significant temperature drop," and "expected reduction in photovoltaic output." The parsed information is then transformed into a structured data table containing these fields.

[0040] Based on this, the basic temporal features and the structured event features are temporally aligned and fused to obtain the multimodal fused feature vector for the current prediction period. Temporal alignment can be achieved by matching the timestamps of the structured event features with the timestamps of the basic temporal features. For example, if an event occurs at a certain time point, the event feature is appended to the basic temporal feature at that time point. Fusion can be achieved through simple feature concatenation, that is, the event feature is directly added as a new dimension to the end of the temporal feature. As an alternative, during temporal alignment, a feature vector can be created for each time step. If an event occurs within that time step, the corresponding structured event feature is filled into the vector for that time step; otherwise, zero values ​​are filled in. During fusion, element-wise addition or multiplication can be used to combine the event features and temporal features.

[0041] Next, the multimodal fusion feature vector is input into prediction models corresponding to at least two different preset scenarios to obtain the power supply and demand prediction sub-results output by each prediction model. These prediction models include at least a baseline scenario prediction model, an extreme weather scenario prediction model, a holiday scenario prediction model, a maintenance scenario prediction model, and a high-renewable-capacity scenario prediction model. Multiple independent prediction models can be pre-trained, each optimized specifically for a preset scenario. For example, one model is specifically used to predict power supply and demand under the baseline scenario, and another model is specifically used to predict power supply and demand under extreme weather scenarios. When the multimodal fusion feature vector is input, all these models corresponding to preset scenarios will perform predictions in parallel and output their respective sub-results. The prediction models can employ different machine learning algorithms, such as decision tree-based models, support vector machine models, or simple linear regression models. Each model is trained independently to learn the power supply and demand patterns under a specific scenario. When the fusion feature vector is received, each model independently generates its prediction sub-result.

[0042] Simultaneously, the current market scenario is represented by this multimodal fusion feature vector, and the scenario similarity between the current market scenario and each preset scenario in the historical scenario feature library is calculated. This preset scenario includes multimodal fusion feature vectors from multiple historical prediction periods. The representation of the current market scenario directly uses the multimodal fusion feature vector from the current prediction period. Scenario similarity can be determined by calculating the Euclidean distance between the current vector and the average vector of each preset scenario in the historical scenario feature library. The smaller the distance, the higher the similarity. The historical scenario feature library can be constructed as multiple representative multimodal fusion feature vectors containing each preset scenario. Scenario similarity can be evaluated by calculating the Manhattan distance between the current market scenario vector and all preset scenario vectors in the historical scenario feature library.

[0043] Then, weights are assigned to the power supply and demand prediction sub-results output by each prediction model based on the scene similarity between the current market scenario and each preset scenario in the historical scenario feature library. These weights are obtained by normalizing the scene similarity, and the sum of the weights is 1. The calculated scene similarity value can be directly used as the initial weight. For example, if the similarity between a preset scenario and the current market scenario is 0.8, the initial weight of the corresponding prediction model sub-result is 0.8. Subsequently, all initial weights are normalized to ensure their sum is 1. The weights can be set to be proportional to the scene similarity. For example, the prediction model corresponding to the scenario with the highest similarity receives the highest weight, and the model corresponding to the scenario with lower similarity receives a lower weight. Normalization can be achieved by dividing each weight by the sum of all weights.

[0044] Finally, the power supply and demand forecast sub-results are weighted and fused to obtain the final power supply and demand forecast result. This final power supply and demand forecast result includes at least the predicted total power demand and the predicted total power supply for the current forecast period. Weighted fusion can be achieved by multiplying each forecast sub-result by its corresponding weight and then summing all the weighted sub-results. For example, if the weight of forecast sub-result A is W_A and the weight of forecast sub-result B is W_B, then the final result is (sub-result A * W_A) + (sub-result B * W_B). After obtaining the sub-results of each forecast model and their corresponding weights, a linear weighted average can be used for fusion. For example, for the total power demand forecast, the total power demand forecast sub-results of all models are multiplied by their weights and then summed to obtain the final total power demand forecast. The total power supply forecast is obtained in the same way.

[0045] This method effectively improves the sensitivity of the power supply and demand forecasting system to complex market environments and key disturbance events by fusing structured time-series data with unstructured event information and utilizing a multi-model collaborative forecasting mechanism. As a result, forecasting bias is effectively controlled when facing typical scenarios such as extreme weather, holiday effects, overlapping maintenance, and sudden changes in renewable energy output, thus meeting the high-precision requirements of the power market, especially the bidding space, for peak value, extreme bias, and cycle matching.

[0046] Example 2 It should be noted that the method also includes: After at least one forecast period, the actual observed total electricity demand and total electricity supply values ​​are obtained, and the corresponding actual supply and demand results are generated. Based on the pre-constructed comprehensive evaluation index system for adapting power supply and demand forecasting, the final power supply and demand forecast results for each forecast period are compared with the actual supply and demand results on a periodic basis, and the index values ​​are calculated. The comprehensive evaluation index system includes conventional time-series forecasting indicators and special indicators. Conventional time-series forecasting indicators include: mean absolute error index and root mean square error index. Special indicators are used to target the characteristics of bidding space, including: peak error index, extreme scenario deviation index and period matching degree index. The prediction results are judged to be up to standard based on the index values ​​and dynamic error thresholds for each prediction period. The forecast period with excessive output error and the corresponding market scenario for that forecast period.

[0047] Specifically, at the end of each forecast period, the system obtains the actual total electricity demand and total electricity supply values ​​for that period from the power dispatch center, market trading platform, or relevant monitoring equipment. These actual observation data serve as the benchmark for evaluating the performance of the forecast model. By performing necessary cleaning, verification, and formatting on these raw data, actual supply and demand results corresponding to the forecast results can be generated, providing reliable real values ​​for subsequent comparative analysis.

[0048] To comprehensively and accurately assess the quality of electricity supply and demand forecasts, a comprehensive evaluation index system needs to be designed and constructed in advance. This system includes not only conventional time-series forecasting indicators, such as the mean absolute error (MAE) and root mean square error (RMSE), used to measure the overall deviation between predicted and actual values; but also specialized indicators tailored to the characteristics of electricity market bidding, such as peak error indicators to assess the accuracy of forecasts during peak load periods; extreme scenario deviation indicators to measure forecast performance under abnormal or sudden events; and cycle matching indicators to assess the degree of agreement between the forecast curve and the actual curve over time. These indicators together constitute a multi-dimensional evaluation framework, capable of reflecting the performance of forecast results in more detail across different dimensions.

[0049] After obtaining the actual supply and demand results, the system precisely compares the final power supply and demand forecast generated within each forecast period with the corresponding actual supply and demand results. This comparison process is performed periodically to ensure consistency over time. By substituting the predicted and actual values ​​into the formulas of various indicators in a pre-built comprehensive evaluation indicator system, the specific indicator values ​​for each forecast period are calculated. These values ​​quantify the accuracy, degree of deviation, and performance under specific scenarios of the forecast, providing a data foundation for subsequent compliance determination.

[0050] To objectively assess the quality of forecast results, the system calculates the indicator values ​​for each forecast period and combines them with a preset dynamic error threshold to determine compliance. The dynamic error threshold is dynamically adjusted based on historical data, market volatility, or specific business needs, rather than being fixed, making the evaluation criteria more adaptable. When any indicator value for a forecast period exceeds its corresponding dynamic error threshold, the forecast result is deemed unsatisfactory, indicating a significant forecast error.

[0051] Once the system determines that the result for a prediction period fails to meet the target, it will immediately output information about the prediction period in which the error exceeds the limit. Simultaneously, to facilitate subsequent error tracing and analysis, the system will also output the corresponding market scenario for that prediction period. Market scenario information typically includes market characteristics represented by the multimodal fusion feature vector within that period, such as the presence of key events like extreme weather, holidays, or equipment maintenance, and the intensity of these events' impact. This helps to quickly pinpoint the specific context in which the problem occurred, providing direction for subsequent optimization.

[0052] The method also includes: Error source analysis is performed on the prediction period where the error exceeds the standard; the error source analysis includes: Verify the feature dimension integrity, timestamp consistency, and update timeliness of the multimodal fusion feature vector to pinpoint the sources of error at the data level; The process of multi-model collaborative prediction is broken down to evaluate the accuracy of market scenario identification, the adaptability and parameter stability of each prediction model, and to pinpoint the sources of error at the model level. Obtain electricity market operating conditions, conduct external disturbance analysis based on electricity market operating conditions, and locate the sources of external disturbance errors; The source conclusions for generating and predicting error exceedance cycles are based on the sources of error at the data level, the sources of error at the model level, and the sources of error from external disturbances. The source conclusions include at least the types of error sources and the target objects for optimization.

[0053] Error source analysis aims to systematically diagnose the root causes of substandard power supply and demand forecasts. By meticulously examining each stage of the forecasting process, problems at different levels, such as data input, model processing, or the external environment, can be identified, providing a clear direction for subsequent system optimization. When locating the source of errors at the data level, the first step is to verify the integrity of the feature dimensions of the multimodal fusion feature vector. Feature dimension integrity refers to checking whether all expected features used to construct the multimodal fusion feature vector have been correctly acquired and included. For example, if a key structured event feature (such as the impact intensity of extreme weather events) is not correctly extracted or mapped, it will lead to missing information in the feature vector, thus affecting forecast accuracy. Secondly, timestamp consistency is verified. Timestamp consistency ensures that all input data, including basic time-series features and structured event features, are precisely aligned on the timeline. Any deviation in timestamps can lead to data misalignment, causing the model to associate irrelevant events or time-series information at incorrect points in time, resulting in forecast errors. Thirdly, update timeliness is verified. Update timeliness refers to checking whether the data used for forecasting is the latest available data. For example, if structured data from the electricity market or unstructured text data related to electricity operation is not updated in a timely manner, the model will make predictions based on outdated information, failing to reflect current market dynamics and thus introducing errors. By validating these data-level attributes, prediction errors caused by data quality, processing, or timeliness issues can be effectively identified.

[0054] When locating the sources of error at the model level, the first step is to break down the multi-model collaborative prediction process. This involves decomposing the entire prediction workflow into multiple sub-modules, such as a market scenario identification module, prediction model modules corresponding to each preset scenario, and a weighted fusion module, to examine their operational status one by one. Secondly, the accuracy of market scenario identification is evaluated. This refers to assessing whether the multimodal fusion feature vector for the current prediction period is accurately matched to the corresponding preset scenario in the historical scenario feature library. Inaccurate scenario identification, such as incorrectly identifying an extreme weather scenario as the baseline scenario, will lead to the subsequent selection and weighting of inappropriate prediction models, resulting in significant errors. Thirdly, the adaptability and parameter stability of each prediction model are evaluated. This involves checking whether the prediction model invoked in a specific scenario is still applicable to the current market conditions, and whether its internal parameters (such as learning rate, regularization coefficient, tree model depth, etc.) remain in an optimal or stable state. Long-term model operation may lead to a decrease in its adaptability to new data or parameter drift, thus affecting prediction performance. By evaluating these model-level elements, errors caused by improper model selection, parameter configuration problems, or model performance degradation can be identified.

[0055] When locating the sources of external disturbance errors, the first step is to obtain the operating conditions of the electricity market. Obtaining these conditions involves collecting actual electricity market operations relevant to the forecast period, including but not limited to grid operation status, unforeseen events (such as equipment failures and line trips), policy and regulatory adjustments, and macroeconomic fluctuations. This information may not be fully reflected in structured or unstructured data. Secondly, external disturbance analysis is conducted based on the electricity market operating conditions. This involves comparing the actual electricity market conditions with the forecast results to identify errors that cannot be explained by data-level or model-level issues. For example, a sudden power plant outage not reported in the news or recorded in the data, or a temporary government-imposed power rationing policy, can lead to significant deviations between the forecast and the actual situation. Through this analysis, errors caused by external, unpredictable, or difficult-to-quantify factors can be identified.

[0056] Finally, based on the sources of error at the data level, model level, and external disturbance levels, a source tracing conclusion is generated for the error exceeding the prediction cycle. This source tracing conclusion is a structured report that synthesizes the above analysis results and clearly identifies the main causes of prediction errors. The conclusion includes at least the error source category and the optimization target. The error source category classifies errors into data-level, model-level, or external disturbance level, facilitating subsequent classification processing. The optimization target specifically indicates the system components or data sources that need improvement, such as "updating the conversion rules from unstructured text to structured event features," "optimizing the parameters of the extreme weather scenario prediction model," or "expanding the historical scenario feature library to include new external disturbances." This conclusion provides clear and actionable guidance for subsequent iterative updates.

[0057] The method also includes: Iterative updates are performed based on the source tracing conclusions; including: When the error source category is data-level error source, update the transformation rules from unstructured text to structured event features, as well as the spatiotemporal alignment threshold between basic time-series features and structured event features; When the error source category is a model-level error source, update the mapping relationship between the preset scenario and the prediction model, optimize the model parameters of the prediction model, and perform incremental training on the prediction model when the incremental training trigger condition is met; wherein, the model parameters include at least one or more of the tree model depth and learning rate; When the error source category is external disturbance error source, update the historical scene feature library used for market scene identification to increase scene features corresponding to external disturbances and improve the ability to identify new market disturbances. After the update is completed, the updated transformation rules, spatiotemporal alignment thresholds, mapping relationships, model parameters, and historical scene feature libraries are called in subsequent prediction cycles to perform predictions. Differential optimization instructions, update content, and update effects are recorded to form an optimization log for long-term iteration.

[0058] Specifically, when the error source category is a data-level error source, the system will update the conversion rules from unstructured text to structured event features, as well as the spatiotemporal alignment threshold between basic time-series features and structured event features. Updating the conversion rules may involve adjusting the event extraction templates, keyword libraries, or mapping logic used in the semantic parsing process of the customized large language model for the power sector, to improve the accuracy of identifying event types, affected objects, and impact intensity from unstructured text data. Updating the spatiotemporal alignment threshold is used to optimize the matching accuracy between basic time-series features and structured event features in the time dimension, ensuring that event information can be accurately fused with the corresponding time-series data, avoiding information bias caused by time misalignment.

[0059] When the error source category is a model-level error source, the system will update the mapping relationship between the preset scenarios and the prediction model, optimize the model parameters of the prediction model, and perform incremental training on the prediction model when the incremental training trigger condition is met. The model parameters include at least one or more of tree model depth and learning rate. Updating the mapping relationship means adjusting the association between different preset scenarios and specific prediction models based on actual prediction results; for example, reallocating certain scenarios to be predicted by a more suitable model. Optimizing model parameters involves fine-tuning the existing prediction model; for example, adjusting the tree model depth or learning rate through methods such as grid search and Bayesian optimization to improve the model's prediction accuracy in specific scenarios. Incremental training, on the basis of the existing model, uses new, validated data to train the model locally or continuously, enabling it to quickly adapt to new market dynamics and data patterns, maintaining the model's timeliness and accuracy.

[0060] When the error source category is an external disturbance error source, the system will update the historical scene feature library used for market scene identification to increase scene features corresponding to the external disturbance and improve the ability to identify new market disturbances. This means that the external disturbance events that cause prediction errors and their corresponding multimodal fusion feature vectors will be included in the historical scene feature library, thereby expanding the coverage of the scene library. In this way, when similar new disturbances occur in the future, the system can more accurately identify the scene through scene similarity calculation and adjust the prediction strategy or weight allocation accordingly, thereby effectively improving the ability to respond to unknown or rare market disturbances.

[0061] After completing the aforementioned updates, in subsequent prediction cycles, the system will invoke the updated transformation rules, spatiotemporal alignment thresholds, mapping relationships, model parameters, and historical scene feature libraries to perform predictions. Simultaneously, the system will record differentiated optimization instructions, update content, and update effects, forming an optimization log. This optimization log serves as a crucial basis for the system's self-learning and long-term iteration, not only tracking the specific details and effects of each optimization but also providing valuable data support for subsequent system maintenance, performance analysis, and further improvements.

[0062] Example 3 The system invokes a large language model customized for the power industry to perform semantic parsing on unstructured text data, including: Events are extracted from unstructured text data to obtain a set of event elements related to power supply and demand. The set of event elements includes at least: event type, affected objects, and impact intensity. Among them, event types include: extreme weather events, holiday events, maintenance events, and abnormal output events of new energy sources. Based on preset feature mapping rules, the set of event elements is mapped into structured event features; the structured event features include at least: the event occurrence time, duration, and quantified value of the impact intensity associated with the affected object; When there are multiple conflicting sources in unstructured text data, the features of structured events are updated based on the credibility or consistency of the data sources.

[0063] Specifically, event extraction from the unstructured text data aims to accurately extract key information that directly or indirectly affects power supply and demand from raw, unformatted unstructured text data related to power operation. Event extraction is a core task in natural language processing, with the goal of identifying events mentioned in the text and extracting elements such as participants, time, and location. Here, the event element set is specifically limited to event type, affected objects, and impact intensity. Event types are further refined into extreme weather events (such as typhoons, cold waves, and high temperatures), holiday events (such as Spring Festival and National Day), maintenance events (such as planned maintenance of generator units and transmission line maintenance), and abnormal renewable energy output events (such as sudden drops or increases in wind and solar power output). Among these, a large language model customized for the power sector, with its powerful semantic understanding capabilities and domain knowledge, can more accurately identify these specific types of events and their elements.

[0064] Building upon this, the set of event elements is mapped into structured event features based on preset feature mapping rules. This step transforms the unstructured event elements extracted in the previous step into standardized, numerical, structured features that can be processed by subsequent models. The preset feature mapping rules are crucial for this transformation; they define how textual information such as event type, affected objects, and impact intensity are converted into specific numerical or coded representations. For example, the event occurrence time can be extracted from the text and converted into a standard timestamp format; the duration can be parsed from the description and quantified into hours or days; the impact intensity may need to be quantified and scored based on the event description (such as "severe impact" or "minor fluctuation") or a preset domain knowledge base. For instance, "typhoon warning" is mapped to high-intensity impact, and "small-scale maintenance" is mapped to low-intensity impact. The generation of structured event features allows unstructured information to be effectively integrated with basic time-series features, providing more comprehensive input for subsequent prediction models.

[0065] Example 4 The basic temporal features and structured event features are temporally aligned and fused, including: The structured event features are mapped to a time axis consistent with the basic time series features according to a preset time granularity, and time window slicing is performed on the structured event features whose start and end times span multiple prediction periods. Perform one or more of the following on the basic temporal features and the aligned structured event features: feature concatenation, gating fusion, or attention fusion, to obtain a multimodal fusion feature vector.

[0066] Mapping the structured event features to a time axis consistent with the base time-series features at a preset time granularity aims to resolve inconsistencies in the time dimension of different modalities. For example, if the base time-series features are at the hourly granularity, while structured event features (such as extreme weather events) may be at the daily granularity or have specific start and end times, then the event features need to be refined or aggregated to the hourly granularity. This can be achieved by evenly distributing the event's impact across each time unit within the event's duration, or by marking the time unit where the event occurs or begins. Simultaneously, performing time window slicing on structured event features whose start and end times span multiple prediction periods ensures that the feature vector within each prediction period only contains information about events that actually occurred or continued within that period. For example, for a three-day holiday event, when making a single-day prediction, only the impact of the event on the current prediction date is captured, avoiding the erroneous introduction of future or past event impacts into the current prediction. This approach ensures the timeliness and relevance of the features, avoiding information redundancy or misalignment.

[0067] Example 5 The method also includes: Scene similarity is calculated using one or more of the following methods: cosine similarity, Euclidean distance similarity, or Mahalanobis distance similarity. Based on the similarity between the current market scenario and the historical scenario feature library, the top K preset scenarios are selected, and weights are assigned only to the power supply and demand prediction sub-results output by the prediction model corresponding to the top K preset scenarios; where K is an integer greater than or equal to 2.

[0068] When calculating scene similarity, a suitable similarity measurement method can be flexibly selected based on the characteristics of the multimodal fusion feature vectors and business requirements. For example, when focusing on the directionality of feature vectors rather than their absolute values, cosine similarity can be used. Cosine similarity measures the similarity between two vectors by calculating the cosine of the angle between them; the closer the value is to 1, the more similar they are. When the dimensions of the feature vectors have similar dimensions and have been normalized, Euclidean distance similarity can be used. Euclidean distance measures the straight-line distance between two points in a multidimensional space; the smaller the distance, the higher the similarity. Furthermore, considering the complex statistical correlations that may exist between different feature dimensions in power supply and demand forecasting, and the potential inconsistencies in the dimensions, Mahalanobis distance similarity can be used. Mahalanobis distance effectively handles the correlations between dimensions and corrects for scale differences between different dimensions, thus more accurately reflecting the true distance between scenes. In some cases, multiple similarity calculation methods can be combined; for example, one method can be used for initial screening, followed by another method for fine-tuning.

[0069] After calculating the scenario similarity between the current market scenario and each preset scenario in the historical scenario feature library, this application no longer performs weighted fusion of the power supply and demand prediction sub-results output by the prediction models corresponding to all preset scenarios. Instead, it introduces a selective fusion mechanism. Specifically, the system sorts all preset scenario similarity values ​​in descending order and selects the top K preset scenarios with the highest similarity. Here, K is a preset integer, and the value of K is greater than or equal to 2, to ensure that at least two most relevant models participate in the fusion, thereby reflecting the advantages of multi-model collaboration.

[0070] Example 6 The method also includes: The supply-demand gap is calculated based on the forecasts of total electricity demand and total electricity supply. The predicted supply-demand gap will be used as part of the final electricity supply-demand forecast.

[0071] Based on the forecasts of total electricity demand and total electricity supply, the project further calculates and outputs a supply-demand gap forecast, greatly enhancing the practicality and decision support capabilities of the electricity supply-demand forecast results. This supply-demand gap forecast directly reveals the potential supply-demand imbalance in the power system during the forecast period, enabling decision-makers to clearly identify potential risks.

[0072] Example 7 The construction of the dynamic error threshold includes: A rolling evaluation window is constructed based on the indicator values ​​of historical forecast periods, and the statistics of the indicator values ​​within the rolling evaluation window are calculated. The statistics include at least one or more of the following: mean, standard deviation, and quantile. Dynamic error thresholds are determined based on statistical measures for both conventional time-series forecasting indicators and specific indicators, and these thresholds are adaptively adjusted when power market conditions change. The adjusted dynamic error threshold is used to determine whether the indicator values ​​in subsequent prediction periods meet the standards.

[0073] The purpose of constructing a rolling evaluation window is to dynamically capture recent trends in historical forecast performance, rather than relying on all historical data. This window can be set to a fixed length, such as the most recent N forecast periods. Whenever a new forecast period ends and actual supply and demand results are generated, the corresponding indicator values ​​for that period are included in the rolling evaluation window, while the oldest period data is removed, thus ensuring the timeliness and relevance of the data within the window.

[0074] Within the rolling evaluation window, calculating the statistical measures of the indicator values ​​is a crucial step in determining the dynamic error threshold. These statistics may include the mean, reflecting the average level of prediction error within the window; the standard deviation, measuring the volatility and dispersion of the prediction error; and quantiles, such as the 90th or 95th percentile, used to set an upper limit on the error, ensuring that the vast majority of predictions meet the requirements. These statistics can be used individually or in combination; for example, the threshold can be set as the mean plus or minus a certain number of standard deviations, or a high quantile can be used directly as the threshold.

[0075] Based on the calculated statistics, dynamic error thresholds are determined separately for conventional time-series forecasting indicators and specific indicators. This method of determining thresholds separately takes into account the different forecasting performance focuses of different types of indicators. For example, for conventional time-series forecasting indicators such as mean absolute error and root mean square error, thresholds can be set based on the mean and standard deviation to reflect the overall forecasting accuracy and stability; while for specific indicators such as peak error, extreme scenario deviation, and period matching degree, quantiles may be used to set thresholds to ensure forecasting quality under specific key scenarios.

[0076] Furthermore, this application proposes an adaptive adjustment of the dynamic error threshold when electricity market conditions change. Changes in electricity market conditions may include, but are not limited to, increased load fluctuations, significant fluctuations in renewable energy output, major policy adjustments, or extreme weather events. When the system detects these changes, it can trigger a recalculation of the threshold or adjust it based on preset rules. For example, the threshold can be appropriately relaxed when market uncertainty increases, or tightened when the market tends to stabilize, to ensure that the threshold always matches the current market environment.

[0077] Finally, the indicator values ​​for subsequent forecast periods are judged to meet the target based on the adjusted dynamic error threshold. This means that the forecast result for each forecast period will be compared with a threshold that is dynamically adjusted according to the current market conditions to determine whether it meets the target.

[0078] Example 8 A power supply and demand forecasting system based on multi-model collaboration includes: Processor and memory; The processor and memory are connected via a communication bus: The processor is used to call and execute programs stored in memory. A memory is used to store a program, which is at least used to execute a power supply and demand forecasting method based on multi-model collaboration as described in the above embodiments.

[0079] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0080] It should be noted that in the description of this application, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of this application, unless otherwise stated, "a plurality of" means at least two.

[0081] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the function involved, as will be understood by those skilled in the art to which embodiments of this application pertain.

[0082] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0083] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0084] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0085] The storage media mentioned above can be read-only memory, disk, or optical disk, etc.

[0086] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0087] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A power supply and demand forecasting method based on multi-model collaboration, characterized in that, include: Obtain multi-source data corresponding to the current forecast period; The multi-source data includes at least: structured data of the electricity market and unstructured text data related to electricity operation; The structured data is subjected to timestamp alignment, missing data handling, and normalization to form basic time-series features; A customized large language model for the power industry is invoked to perform semantic parsing on the unstructured text data, identify the types of events, affected objects, and intensity of impact related to power supply and demand, and map the parsing results into structured event features; The basic temporal features and the structured event features are temporally aligned and fused to obtain the multimodal fusion feature vector for the current prediction period; The multimodal fusion feature vector is input into at least two prediction models corresponding to different preset scenarios to obtain the power supply and demand prediction sub-results output by each prediction model; the prediction models include at least: a baseline scenario prediction model, an extreme weather scenario prediction model, a holiday scenario prediction model, a maintenance scenario prediction model, and a high proportion of new energy scenario prediction model. The current market scenario is represented by the multimodal fusion feature vector, and the scenario similarity between the current market scenario and each preset scenario in the historical scenario feature library is calculated; the preset scenario includes multimodal fusion feature vectors from multiple historical prediction periods; The power supply and demand prediction sub-results output by each prediction model are weighted according to the scene similarity between the current market scenario and each preset scenario in the historical scenario feature library; the weights are obtained by normalizing the scene similarity and the sum of the weights is 1. The power supply and demand forecast sub-results are weighted and fused to obtain the final power supply and demand forecast result; the final power supply and demand forecast result includes at least the total power demand forecast and the total power supply forecast for the current forecast period.

2. The method according to claim 1, characterized in that, The method further includes: After at least one forecast period, the actual observed total electricity demand and total electricity supply values ​​are obtained, and the corresponding actual supply and demand results are generated. Based on a pre-constructed comprehensive evaluation index system for adapting power supply and demand forecasting, the final power supply and demand forecast results for each forecast period are compared with the actual supply and demand results on a periodic basis, and the index values ​​are calculated. The comprehensive evaluation index system includes conventional time-series forecasting indicators and special indicators. The conventional time-series forecasting indicators include: mean absolute error index and root mean square error index. The special indicators are used to target the characteristics of bidding space and include: peak error index, extreme scenario deviation index and period matching degree index. The prediction results are judged to be up to standard based on the index values ​​and dynamic error thresholds for each prediction period. The forecast period with excessive output error and the corresponding market scenario for that forecast period.

3. The method according to claim 2, characterized in that, The method further includes: An error source analysis is performed on the prediction period where the error exceeds the standard; the error source analysis includes: Verify the feature dimension integrity, timestamp consistency, and update timeliness of the multimodal fusion feature vector to pinpoint the sources of error at the data level; The process of multi-model collaborative prediction is broken down to evaluate the accuracy of market scenario identification, the adaptability and parameter stability of each prediction model, and to pinpoint the sources of error at the model level. Obtain electricity market operating conditions, conduct external disturbance analysis based on electricity market operating conditions, and locate the sources of external disturbance errors; The source conclusions for generating and predicting error exceedance cycles are based on data-level error sources, model-level error sources, and external disturbance error sources; the source conclusions include at least the error source category and the optimization target object.

4. The method according to claim 3, characterized in that, The method further includes: Iterative updates are performed based on the aforementioned source tracing conclusions; including: When the error source category is a data-level error source, update the transformation rules from unstructured text to structured event features, as well as the spatiotemporal alignment threshold between basic time-series features and structured event features; When the error source category is a model-level error source, the mapping relationship between the preset scenario and the prediction model is updated, the model parameters of the prediction model are optimized, and incremental training is performed on the prediction model when the incremental training trigger condition is met; wherein, the model parameters include at least one or more of tree model depth and learning rate; When the error source category is an external disturbance error source, the historical scene feature library used for market scene identification is updated to increase scene features corresponding to the external disturbance and improve the ability to identify new market disturbances. After the update is completed, the updated transformation rules, spatiotemporal alignment thresholds, mapping relationships, model parameters, and historical scene feature libraries are called in subsequent prediction cycles to perform predictions. Differential optimization instructions, update content, and update effects are recorded to form an optimization log for long-term iteration.

5. The method according to claim 1, characterized in that, The process of invoking a large language model customized for the power industry to perform semantic parsing on the unstructured text data includes: Events are extracted from the unstructured text data to obtain a set of event elements related to power supply and demand; the set of event elements includes at least: event type, affected object, and impact intensity; wherein, the event type includes: extreme weather events, holiday events, maintenance events, and abnormal output events of new energy sources; The event element set is mapped into structured event features based on preset feature mapping rules; the structured event features include at least: the event occurrence time, duration, and quantified value of the impact intensity associated with the affected object; When there are multiple source conflicts in the unstructured text data, the structured event features are updated based on the data source credibility or consistency check structure.

6. The method according to claim 1, characterized in that, The step of aligning and fusing the basic temporal features with the structured event features in a temporal sequence includes: The structured event features are mapped to a time axis consistent with the basic time series features according to a preset time granularity, and time window slicing is performed on the structured event features whose start and end times span multiple prediction periods. The basic temporal features and the aligned structured event features are subjected to one or more of the following: feature concatenation, gating fusion, or attention fusion, to obtain the multimodal fusion feature vector.

7. The method according to claim 1, characterized in that, The method further includes: The scene similarity is calculated using one or more of cosine similarity, Euclidean distance similarity, or Mahalanobis distance similarity. Based on the similarity between the current market scenario and the historical scenario feature library, the top K preset scenarios are selected, and weights are assigned only to the power supply and demand prediction sub-results output by the prediction model corresponding to the top K preset scenarios; where K is an integer greater than or equal to 2.

8. The method according to claim 1, characterized in that, The method further includes: The supply-demand gap forecast is calculated based on the total electricity demand forecast and the total electricity supply forecast. The predicted supply-demand gap is output as part of the final electricity supply-demand forecast result.

9. The method according to claim 2, characterized in that, The construction of the dynamic error threshold includes: A rolling evaluation window is constructed based on the indicator values ​​of historical prediction periods, and statistics of the indicator values ​​within the rolling evaluation window are calculated. The statistics include at least one or more of the following: mean, standard deviation, and quantiles. Based on the statistical measures, dynamic error thresholds are determined for the conventional time-series forecast indicators and special indicators, respectively, and the dynamic error thresholds are adaptively adjusted when the power market conditions change. The adjusted dynamic error threshold is used to determine whether the indicator values ​​in subsequent prediction periods meet the standards.

10. A power supply and demand forecasting system based on multi-model collaboration, characterized in that, include: Processor and memory; The processor and memory are connected via a communication bus: The processor is used to call and execute the program stored in the memory; The memory is used to store a program, which is at least used to execute the power supply and demand forecasting method based on multi-model collaboration as described in any one of claims 1-9.

Images