Photovoltaic-load-price prediction method, device, equipment, medium and product fusing large model and error limiting correction

Abstract

The application discloses a photovoltaic-load-price prediction method and device fusing a large model and error limiting correction, equipment, medium and product, and relates to the field of power system prediction. The method comprises the following steps: acquiring prediction meteorological data of a to-be-predicted time period, inputting a trained day-ahead trend prediction large model, and obtaining a day-ahead trend prediction result; acquiring photovoltaic power true values and electrical load true values collected at a second sampling interval before the to-be-predicted time period, inputting a trained real-time rolling prediction large model, and obtaining a source-load prediction result; and performing error limiting correction on the source-load prediction result based on the day-ahead trend prediction result, and obtaining a final source-load prediction result. The two-stage prediction architecture takes into account the day-ahead trend and the super-short-term high-resolution prediction demand, the error limiting correction suppresses error accumulation in the super-short-term rolling prediction, and the coupling relationship among the photovoltaic power, the load and the electricity price, thereby improving the prediction accuracy and robustness of the photovoltaic power, the load and the electricity price.

Description

Technical Field

[0001] This application relates to the field of power system forecasting technology, and in particular to a photovoltaic-load-price forecasting method, device, equipment, medium and product that integrates large model and error limiting correction. Background Technology

[0002] With the construction of new power systems, the penetration rate of renewable energy sources such as photovoltaics is constantly increasing. Their inherent intermittency and volatility pose significant challenges to the safe and stable operation of the power grid and the economic dispatch of the electricity market. Against this backdrop, accurate forecasting of photovoltaic output, power load, and market electricity prices has become a key technology supporting intelligent grid dispatch, efficient energy management, and market risk mitigation.

[0003] In related technologies, forecasting methods for photovoltaic load-electricity price mainly suffer from the following limitations: First, they typically employ single-time-scale forecasting models. Second, while day-ahead forecasting models can provide trend forecasts for the next 24 hours with a time resolution of 1 hour, they cannot meet the urgent demand of the real-time balancing market for ultra-short-term, high-time-resolution forecasting data. Third, to address real-time requirements, some studies employ rolling time-series forecasting models, but these models suffer from the problem of prediction errors accumulating and amplifying with increasing prediction step size during recursive forecasting, leading to decreased reliability over longer forecasting time spans. Furthermore, traditional forecasting methods usually treat photovoltaic, load, and electricity price as independent time series for separate forecasting, failing to effectively utilize the coupling relationship between them, resulting in insufficient information utilization in the forecasting model. Simultaneously, the performance of advanced artificial intelligence forecasting models highly depends on a large amount of high-quality historical data. In rare scenarios such as extreme weather and special events, historical data is often insufficient or lacks diversity, leading to poor model generalization ability.

[0004] Furthermore, with breakthroughs in large-scale modeling techniques in natural language processing, time-series large-scale models have gradually become a research hotspot in the field of time-series forecasting. These models are pre-trained on massive amounts of time-series data, possessing powerful sequence modeling capabilities and generalization performance, and can be adapted to specific domain tasks through fine-tuning. However, there is currently no mature solution that combines time-series large-scale models with error limiting correction mechanisms for joint forecasting of photovoltaic-load-electricity prices.

[0005] Therefore, there is an urgent need for a photovoltaic-load-price forecasting method that integrates large models and error limiting correction to solve the problems that single-timescale models cannot take into account both day-ahead trend forecasting and ultra-short-term high-resolution forecasting, and that ultra-short-term rolling forecasting suffers from error accumulation, isolated forecasting elements, and poor model generalization ability, thereby improving the forecasting accuracy and robustness of photovoltaic power, load, and electricity price. Summary of the Invention

[0006] The purpose of this application is to provide a photovoltaic-load-price forecasting method, device, equipment, medium, and product that integrates large model and error limiting correction. It can meet the dual needs of day-ahead trend forecasting and ultra-short-term high-resolution forecasting. Through the error limiting correction mechanism, it effectively suppresses the accumulation of errors in ultra-short-term rolling forecasting, and at the same time explores the coupling relationship between photovoltaic, load, and electricity price, thereby improving the forecasting accuracy and robustness of photovoltaic power, load, and electricity price.

[0007] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a photovoltaic-load-electricity price forecasting method that integrates a large model with error limiting correction, including: Obtain forecast meteorological data for the time period to be predicted; the forecast meteorological data includes meteorological parameter forecast data and satellite cloud image forecast data; The predicted meteorological data for the time period to be predicted is input into the trained daytime trend prediction model to obtain the daytime trend prediction results for the time period to be predicted. The daytime trend prediction results include photovoltaic power prediction, load prediction, and electricity price prediction. The daytime trend prediction model is obtained by fine-tuning the parameters of a pre-trained first time series model using a first sample dataset. The first sample dataset is obtained by collecting multimodal datasets of historical time periods at a first sampling interval. The multimodal dataset includes real meteorological parameter data, real satellite cloud image data, and corresponding real values ​​of photovoltaic power, electricity load, and electricity price. Obtain the true values ​​of photovoltaic power and electrical load collected at the second sampling interval within the first historical time window before the time period to be predicted; The real values ​​of photovoltaic power and electrical load, collected at the second sampling interval within the first historical time window before the time period to be predicted, are input into the trained real-time rolling prediction model to obtain the source load prediction results for the time period to be predicted. The source load prediction results include photovoltaic power prediction values ​​and load prediction values. The real-time rolling prediction model is obtained by fine-tuning the parameters of the pre-trained second time series model using the second sample dataset, which is obtained by collecting the real values ​​of photovoltaic power and electrical load from the multimodal dataset of historical time periods at the second sampling interval. Based on the day-ahead trend forecast results, the source-load forecast results are adaptively error-limited and corrected to obtain the final source-load forecast results for the forecast period. The final source-load forecast results include the electricity price forecast, the photovoltaic power forecast after error-limiting correction, and the load forecast from the day-ahead trend forecast results.

[0008] Secondly, this application provides a photovoltaic-load-price forecasting device that integrates a large model and error limiting correction, comprising: The forecast meteorological data acquisition module is used to acquire forecast meteorological data for the time period to be predicted; the forecast meteorological data includes meteorological parameter forecast data and satellite cloud image forecast data; The day-ahead trend prediction module is used to input the predicted meteorological data for the time period to be predicted into a pre-trained day-ahead trend prediction model to obtain the day-ahead trend prediction results for the time period to be predicted. The day-ahead trend prediction results include photovoltaic power prediction values, load prediction values, and electricity price prediction values. The day-ahead trend prediction model is obtained by fine-tuning the parameters of a pre-trained first time-series model using a first sample dataset. The first sample dataset is collected from multi-modal datasets of historical time periods at a first sampling interval. The multi-modal dataset includes real meteorological parameter data, real satellite cloud image data, and corresponding real values ​​of photovoltaic power, electricity load, and electricity price. The real-time data acquisition module is used to acquire the real values ​​of photovoltaic power and electrical load collected at the second sampling interval within the first historical time window before the time period to be predicted. The real-time rolling prediction module is used to input the actual photovoltaic power and electrical load values ​​collected at a second sampling interval within a first historical time window before the time period to be predicted into a trained real-time rolling prediction model to obtain the source load prediction results for the time period to be predicted. The source load prediction results include photovoltaic power prediction values ​​and load prediction values. The real-time rolling prediction model is obtained by fine-tuning the parameters of a pre-trained second time series model using a second sample dataset, which is obtained by collecting the actual photovoltaic power and electrical load values ​​from a multimodal dataset of historical time periods at a second sampling interval. The error limiting correction module is used to adaptively limit the source load prediction results based on the day-ahead trend prediction results to obtain the final source load prediction results for the time period to be predicted; the final source load prediction results include the electricity price prediction value, the photovoltaic power prediction value after error limiting correction, and the load prediction value in the day-ahead trend prediction results.

[0009] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the photovoltaic-load-price forecasting method with fusion of large model and error limiting correction as described above.

[0010] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the photovoltaic-load-price forecasting method with fusion of large model and error limiting correction described above.

[0011] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the photovoltaic-load-price forecasting method with fusion of large model and error limiting correction described above.

[0012] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a photovoltaic-load-electricity price forecasting method, apparatus, equipment, medium, and product that integrates a large-scale model with error limiting correction. By acquiring forecast meteorological data for the forecast period, including meteorological parameter forecast data and satellite cloud image forecast data, and inputting the forecast meteorological data into a trained day-ahead trend forecasting large-scale model, a day-ahead trend forecasting result containing photovoltaic power forecast, load forecast, and electricity price forecast is obtained. This solves the problems of isolated forecasting elements and failure to effectively utilize the coupling relationship between photovoltaic, load, and electricity price in the prior art, and realizes multi-task joint forecasting. It enables the day-ahead trend forecasting large-scale model to learn the intrinsic relationship between the three and provide a macro trend benchmark for subsequent real-time forecasting. By acquiring the actual photovoltaic power and electrical load values ​​collected at a second sampling interval within the first historical time window before the predicted time period, and inputting these values ​​into a trained real-time rolling forecasting model, a source-load forecasting result containing both photovoltaic power and load forecasts is obtained. This solves the problem that existing single-timescale forecasting models cannot meet the real-time balancing market's demand for ultra-short-term, high-temporal-resolution forecasting data, achieving ultra-short-term, high-resolution rolling forecasting of photovoltaic power and load. By adaptively limiting the source-load forecasting result based on the day-ahead trend forecasting result, a final source-load forecasting result is obtained, containing the electricity price forecast from the day-ahead trend forecasting result, the error-limited photovoltaic power forecast, and the load forecast. This solves the error accumulation problem in the recursive process of existing ultra-short-term rolling forecasting, achieving the goal of confining the rolling forecasting result to near the day-ahead macroeconomic trend while retaining the high-resolution advantage of ultra-short-term forecasting, effectively suppressing error propagation and amplification. By training a large-scale daily trend prediction model using a first sample dataset collected at a first sampling interval, and training a large-scale real-time rolling prediction model using a second sample dataset collected at a second sampling interval, combined with the construction of a multimodal dataset, the problem of dependence on data quality and quantity in existing data-driven models is solved, providing a foundation for subsequent data augmentation. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is an application environment diagram of a photovoltaic-load-electricity price forecasting method that integrates a large model and error limiting correction in one embodiment of this application; Figure 2 A flowchart illustrating a photovoltaic-load-price forecasting method that integrates a large model and error limiting correction, provided as an embodiment of this application; Figure 3 A flowchart illustrating a photovoltaic-load-price forecasting method that integrates a large model and error limiting correction, provided for another embodiment of this application; Figure 4 This is a schematic diagram illustrating the effect of electrical load prediction provided in an embodiment of this application; Figure 5 A schematic diagram of the functional modules of a photovoltaic-load-price forecasting device that integrates a large model and error limiting correction, provided as an embodiment of this application; Figure 6 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0015] First, some technical terms involved in the embodiments of this application will be introduced.

[0016] Forecast: This refers to the prediction of the power system status for the next 24 hours, with a time resolution typically of 1 hour.

[0017] Ultra-short-term forecasting refers to rolling forecasts of the power system's state over the next few minutes to hours, with a time resolution typically between 5 and 15 minutes.

[0018] Generative Adversarial Networks (GANs) consist of a generator and a discriminator. The generator produces supplementary data that matches the distribution of the real data, while the discriminator distinguishes between the real and generated data. Through adversarial training, they compete and improve together, ultimately enabling the generator to produce high-quality supplementary data.

[0019] Time Series Generative Adversarial Network (TimeGAN): A variant of generative adversarial network specifically designed for time series data. It introduces a temporal supervision mechanism on the basis of traditional GAN, and can simultaneously learn the static distribution and dynamic temporal features of real time series data to generate supplementary time series data with the same distribution as the original data but different temporal patterns for data augmentation.

[0020] Large-scale time series models refer to deep learning models pre-trained on massive amounts of cross-domain time series data. These models typically have hundreds of millions to tens of billions of parameters and possess powerful time series representation capabilities and zero-shot prediction capabilities. They can be quickly adapted to specific domain tasks through fine-tuning. Representative models include Chronos, MOMENT, and sundials.

[0021] Parameter-Efficient Fine-Tuning (PEFT) refers to techniques that update only a small number of additional parameters while freezing the original pre-trained weights when adapting a large pre-trained model to a downstream task. Examples include LoRA (Lower-Rank Adaptation), Adapter, and Prefix Tuning. This technique can significantly reduce training costs while maintaining the generalization ability of large models.

[0022] LoRA (Low-Rank Adaptation): A parameter-efficient fine-tuning method that inserts low-rank matrices into the attention layer of a pre-trained model, trains only the parameters of these matrices, freezes the original pre-trained weights, and achieves adaptation to downstream tasks with a small number of trainable parameters.

[0023] Recursive strategy: This refers to using the predicted value output by the model as part of the input during the rolling prediction process to predict the value at the next time point, and so on recursively until the prediction of the preset number of steps is completed.

[0024] The relevant technologies have the following main limitations: a. Isolated Forecasting Elements, Ignoring Coupling Relationships: Traditional forecasting methods typically treat photovoltaic (PV) power, load, and electricity prices as independent time series for separate forecasting. However, there is a strong coupling and interaction among these three factors. Solar irradiance directly affects PV output, which in turn influences market electricity prices through supply and demand, while electricity price signals, in turn, affect demand-side load. Existing technologies fail to effectively utilize this inherent physical and economic connection, resulting in insufficient information utilization in forecasting models and limited upper limits to forecast accuracy.

[0025] b. Day-ahead forecasts fail to meet real-time requirements: Existing forecasting studies mostly focus on the day-ahead market, providing trend forecasts for the next 24 hours with a time resolution of 1 hour. While this is crucial for day-ahead bidding planning, it cannot meet the urgent need of the real-time balancing market for ultra-short-term, high-time-resolution (e.g., 5-15 minutes) forecast data. Real-time grid dispatching and high-frequency trading decisions require more refined forecast information that is closer to the current moment.

[0026] c. The problem of error accumulation in ultra-short-term forecasting: To address real-time requirements, some studies employ rolling time series forecasting models (recursive forecasting). However, these models have an inherent flaw: error accumulation. That is, the forecasting error at each step propagates and amplifies the errors in subsequent steps, causing the predicted trajectory to deviate significantly from the true value when the forecast time span is long (e.g., predicting the next few hours), resulting in a sharp decline in reliability. How to effectively control the spread of errors in rolling forecasting is a pressing technical challenge that needs to be addressed.

[0027] d. Dependence of large data-driven models on data quality and quantity: The performance of advanced artificial intelligence prediction models (such as TCN (Temporal Convolutional Network) and Attention mechanisms) is highly dependent on a large amount of high-quality historical data. However, in the energy sector, especially in the face of rare scenarios such as extreme weather and special events, historical data is often insufficient or lacks diversity, resulting in poor model generalization ability and prediction failure in "black swan" events.

[0028] Therefore, this application provides a source-load prediction method that can: (1) combine multimodal feature inputs (meteorological parameters, satellite cloud images, etc.) to make full use of the coupling information between photovoltaic, load and electricity price for collaborative prediction; (2) take into account the dual needs of day-ahead trend prediction and ultra-short-term refined prediction; (3) effectively solve the error accumulation problem in ultra-short-term rolling prediction; and (4) improve the generalization ability of the model in complex scenarios through data augmentation.

[0029] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0030] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, this application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0031] The photovoltaic-load-electricity price forecasting method integrating a large model and error limiting correction provided in this application can be applied to, for example... Figure 1 In the application environment shown, terminal 101 communicates with server 102 via a network. A data storage system can store the data that server 102 needs to process. The data storage system can be set up independently, integrated into server 102, or placed in the cloud or on other servers. Terminal 101 can send the predicted meteorological data for the time period to be predicted to server 102. After receiving the predicted meteorological data for the time period, server 102 inputs the predicted meteorological data for the time period into a trained day-ahead trend prediction model to obtain the day-ahead trend prediction result for the time period. The day-ahead trend prediction result includes predicted photovoltaic power, predicted load, and predicted electricity price. The day-ahead trend prediction model is obtained by fine-tuning the parameters of a pre-trained first time-series model using a first sample dataset. The first sample dataset is collected from a multi-modal dataset of historical time periods at a first sampling interval. The multi-modal dataset includes real meteorological parameter data, real satellite cloud image data, and corresponding real values ​​of photovoltaic power, electricity load, and electricity price. The system also obtains photovoltaic power data collected at a second sampling interval within a first historical time window before the time period to be predicted. The system takes the actual photovoltaic power and electrical load values ​​as input. It inputs the actual photovoltaic power and electrical load values ​​collected within a first historical time window prior to the predicted time period at a second sampling interval into a trained real-time rolling prediction model to obtain the source-load prediction result for the predicted time period. The source-load prediction result includes the predicted photovoltaic power value and the predicted load value. The real-time rolling prediction model is obtained by fine-tuning the parameters of a pre-trained second time-series model using a second sample dataset, which is obtained by collecting the actual photovoltaic power and electrical load values ​​from a multi-modal dataset of historical time periods at a second sampling interval. Based on the day-ahead trend prediction result, the source-load prediction result is adaptively error-limited and corrected to obtain the final source-load prediction result for the predicted time period. The final source-load prediction result includes the predicted electricity price value from the day-ahead trend prediction result, the photovoltaic power prediction value after error-limiting correction, and the predicted load value. The server 102 can feed back the obtained final source-load prediction result to the terminal 101. In addition, in some embodiments, the photovoltaic-load-price forecasting method that integrates large models and error limiting correction can also be implemented separately by server 102 or terminal 101. For example, terminal 101 can directly perform photovoltaic-load-price forecasting processing on the forecast meteorological data of the time period to be predicted, or server 102 can obtain the forecast meteorological data of the time period to be predicted from the data storage system and perform photovoltaic-load-price forecasting processing on the forecast meteorological data of the time period to be predicted.

[0032] The terminal 101 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. The server 102 can be implemented using a standalone server or a server cluster consisting of multiple servers, or it can be a cloud server.

[0033] In one exemplary embodiment, such as Figure 2 As shown, a photovoltaic-load-price forecasting method integrating a large model and error limiting correction is provided. This method is executed by computer equipment, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 Taking server 102 as an example, the explanation includes the following steps 201 to 205. Wherein: Step 201: Obtain the predicted meteorological data for the time period to be predicted; the predicted meteorological data includes meteorological parameter prediction data and satellite cloud image prediction data.

[0034] Step 202: Input the predicted meteorological data for the time period to be predicted into the trained day-ahead trend prediction model to obtain the day-ahead trend prediction results for the time period to be predicted; the day-ahead trend prediction results include photovoltaic power prediction values, load prediction values, and electricity price prediction values; the day-ahead trend prediction model is obtained by fine-tuning the parameters of the pre-trained first time series model using the first sample dataset, and the first sample dataset is obtained by collecting multimodal datasets of historical time periods at the first sampling interval; the multimodal dataset includes real meteorological parameter data, real satellite cloud image data, and corresponding real values ​​of photovoltaic power, real values ​​of electricity load, and real values ​​of electricity price.

[0035] Step 203: Obtain the true values ​​of photovoltaic power and electrical load collected at the second sampling interval within the first historical time window before the time period to be predicted.

[0036] Step 204: Input the actual photovoltaic power and electrical load values ​​collected within the first historical time window before the time period to be predicted at the second sampling interval into the trained real-time rolling prediction model to obtain the source load prediction results for the time period to be predicted; the source load prediction results include photovoltaic power prediction values ​​and load prediction values; the real-time rolling prediction model is obtained by fine-tuning the parameters of the pre-trained second time series model using the second sample dataset, and the second sample dataset is obtained by collecting the actual photovoltaic power and electrical load values ​​in the multimodal dataset of the historical time period at the second sampling interval.

[0037] Step 205: Based on the day-ahead trend forecast results, perform adaptive error limiting correction on the source-load forecast results to obtain the final source-load forecast results for the forecast period; the final source-load forecast results include the electricity price forecast value, the photovoltaic power forecast value after error limiting correction, and the load forecast value in the day-ahead trend forecast results.

[0038] By implementing steps 201 to 205 above, this application can meet the dual needs of day-ahead trend forecasting and ultra-short-term high-resolution forecasting. It can effectively suppress error accumulation in ultra-short-term rolling forecasting through an error limiting correction mechanism, and at the same time use multi-task learning to explore the coupling relationship between photovoltaic power, load and electricity price, thereby improving the forecast accuracy and robustness of photovoltaic power, load and electricity price.

[0039] In another exemplary embodiment of this application, the parameter fine-tuning process of the large-scale model for day-ahead trend prediction specifically includes: The real meteorological parameter data and real satellite cloud image data in the first sample dataset are used as inputs to the pre-trained first time series large model. The pre-trained first time series large model is fine-tuned using a parameter-efficient fine-tuning method. The model outputs the photovoltaic power prediction value, electricity load prediction value, and electricity price prediction value corresponding to the real meteorological parameter data and real satellite cloud image data. Training stops when the loss function value is less than the preset convergence threshold or the preset maximum number of iterations is reached, and the trained day-ahead trend prediction large model is obtained. The loss function is constructed based on the photovoltaic power prediction value, electricity load prediction value, electricity price prediction value, and the corresponding real values ​​of photovoltaic power, electricity load, and electricity price.

[0040] The first time-series large model comprises a multimodal coding layer, a Transformer backbone layer, and a multi-task output layer connected in sequence. The multimodal coding layer encodes meteorological parameter data and satellite cloud image data into a unified representation to obtain a multimodal fusion feature vector. The Transformer backbone layer extracts time-series features based on the multimodal fusion feature vector through a self-attention mechanism to obtain a time-series feature representation. The multi-task output layer simultaneously outputs photovoltaic power prediction, load prediction, and electricity price prediction based on the time-series feature representation through a multi-task learning architecture.

[0041] In another exemplary embodiment of this application, the method for constructing the first sample dataset and the second sample dataset specifically includes: Obtain a first raw dataset collected from the multimodal dataset within a historical time period at a first sampling interval, and a second raw dataset collected from the actual photovoltaic power and actual electrical load values ​​in the multimodal dataset within the historical time period at a second sampling interval.

[0042] The first original multimodal dataset and the second original multimodal dataset are preprocessed to obtain the preprocessed first original dataset and the second original dataset; the preprocessing operations include data cleaning and normalization operations.

[0043] Based on the preprocessed first and second original datasets, time-series generative adversarial networks are used to generate a first supplementary dataset and a second supplementary dataset, respectively. The first supplementary dataset has the same data volume as the preprocessed first original dataset but different data values. The second supplementary dataset has the same data volume as the preprocessed second original dataset but different data values.

[0044] The first supplementary dataset is merged with the preprocessed first original dataset to obtain the first sample dataset.

[0045] The second supplementary dataset is merged with the preprocessed second original dataset to obtain the second sample dataset.

[0046] In another exemplary embodiment of this application, step 204 specifically includes: Step 204-1: Divide the time period to be predicted into multiple consecutive rolling periods; each rolling period contains multiple consecutive rolling time periods divided according to the second sampling interval.

[0047] Step 204-2: Predict the photovoltaic power forecast and load forecast for all rolling time periods within the k-th rolling cycle through the following steps; where k={1, 2, ..., K}, s={1, 2, ..., S}, K represents the total number of rolling cycles, and S represents the total number of rolling time periods within a rolling cycle.

[0048] Step 204-3: When s=1, the real values ​​of photovoltaic power and electrical load collected at the second sampling interval within the first historical time window before the start time of the kth rolling cycle are used as the input sequence and input into the trained real-time rolling prediction model to predict the source and load prediction results for the first rolling time period within the kth rolling cycle.

[0049] Step 204-4: When s≥2, add the source load prediction results from the first rolling time period to the (s-1)th rolling time period generated in the kth rolling cycle to the input sequence, and remove the earliest s-1th rolling time period length of the actual photovoltaic power and actual electrical load values ​​in the input sequence to form an updated input sequence. Input the updated input sequence into the trained real-time rolling prediction model to obtain the source load prediction results for the sth rolling time period in the kth rolling cycle, until s=S, thus completing the prediction for the kth rolling cycle.

[0050] Step 204-5: Let k=k+1, repeat the above prediction steps for the rolling cycle until k=K, and obtain the source load prediction results for the time period to be predicted.

[0051] In another exemplary embodiment of this application, adaptive error limiting correction is performed on the source load prediction results based on the day-ahead trend prediction results, specifically including: When s=S within each rolling cycle, the source load prediction result for the S-th rolling time period is adaptively error-limited and corrected using the following formula: .

[0052] in, This represents the predicted object after adaptive error limiting correction in the S-th rolling time period. i The predicted value; ,when When, it indicates photovoltaic power; when The time indicates the electrical load; This represents the predicted object output by the large model in the real-time rolling prediction during the S-th rolling time period. i The predicted value; This indicates the forecast object corresponding to the S-th rolling time period in the current trend forecast results. i The trend value; Indicates the corrected strength coefficient; This indicates an adaptively adjusted threshold.

[0053] In another exemplary embodiment of this application, the loss function of the large-scale day-ahead trend prediction model is: .

[0054] in, This represents the loss function value of the large-scale model for predicting current trends. The weighting coefficients represent the photovoltaic power prediction task. This represents the mean square error value of photovoltaic power prediction; Indicates the weighting coefficients for load forecasting tasks; This represents the mean square error value of the load forecast; This represents the weighting coefficient for the electricity price forecasting task; This represents the average absolute error of the electricity price forecast.

[0055] This application acquires multimodal historical data at two time scales—a first sampling interval (1 hour) and a second sampling interval (5 minutes)—through a data acquisition phase. In the data augmentation phase, a time-series generative adversarial network is used to learn the joint distribution and temporal characteristics of the historical data, generating high-quality supplementary data to enhance the training set. In the multi-time-scale prediction phase, firstly, in day-ahead prediction, a pre-trained first time-series large model is used, and its parameters are efficiently fine-tuned using the first sample dataset to train a multi-task learning day-ahead prediction large model, achieving joint trend prediction of photovoltaic power, load, and electricity price. Subsequently, based on a pre-trained lightweight second time-series large model, a parameter-efficient fine-tuning technique is used to train a real-time prediction large model. Combined with a recursive prediction strategy, ultra-short-term time-series rolling predictions of photovoltaic power and load are performed. The large model dynamically identifies error patterns and generates an adaptive limiting ratio of 5%. Therefore, the limiting function constrains the prediction results within ±5% of the day-ahead prediction trend to suppress error accumulation, effectively improving the accuracy and robustness of new energy power plant source-load prediction and solving the problem of error accumulation in ultra-short-term rolling predictions. This approach is highly practical.

[0056] The following example illustrates this application using a specific fusion of large-scale models and error limiting correction for photovoltaic-load-electricity price forecasting.

[0057] like Figure 3 The diagram illustrates a flowchart of a photovoltaic-load-price forecasting method that integrates a large model with error limiting correction, specifically including the following steps: S1. Acquire historical data of multimodal input features such as meteorology and cloud images, as well as historical data of multi-task outputs of photovoltaic power, electricity load, and electricity price, including two time scales: 1 hour and 5 minutes. The 1-hour time scale data will be used for day-ahead forecasting to understand the future trends of photovoltaic power and load and time-of-use electricity prices, while also providing a basis for real-time forecast correction. The 5-minute time scale is used for real-time forecast correction, to accurately grasp the ultra-short-term fluctuations of photovoltaic power and load by shortening the forecast time resolution and time span.

[0058] As an optional implementation, step S1 includes the following steps: S1.1 Obtain historical and forecast meteorological data from the meteorological department's API (Application Programming Interface), including global horizontal irradiance (GHI), ambient temperature, etc., and read satellite cloud images.

[0059] S1.2 Obtain historical and real-time total output data and regional electrical load data of photovoltaic power plants from the Energy Management System (EMS) or Supervisory Control and Data Acquisition (SCADA).

[0060] S1.3 Obtain historical and real-time Locational Marginal Price (LMP) data from the power trading center.

[0061] S2. Data preprocessing and augmentation: Outliers are removed from historical data, and missing values ​​are supplemented using interpolation. Data is normalized, and data augmentation and expansion are achieved based on a time-series generative adversarial network. This step improves data quality while addressing issues such as insufficient initial data volume.

[0062] As an optional implementation, step S2 includes the following steps: S2.1 Data Cleaning: Process the raw data obtained in step S1, fill in missing values ​​using linear interpolation, and identify and remove outliers that are significantly beyond physical meaning (such as positive irradiance at night).

[0063] S2.2 Normalization: Max-min normalization is applied to the cleaned meteorological, photovoltaic, load, and electricity price data respectively, scaling them to the [0, 1] interval. The normalization method is as follows: .

[0064] in, The value is the normalized value; These are the initial values ​​for historical data; It is the minimum value in historical data; This represents the maximum value from historical data.

[0065] S2.3. Input the normalized historical data into the Time Series Generative Adversarial Network (TimeGAN) for training. The TimeGAN's generator and discriminator learn the joint probability distribution and dynamic temporal patterns of real historical data through adversarial training. After training, the generator is used to generate supplementary time series data that is similar in size to the original historical data but different in content.

[0066] S2.4 Construct an enhanced dataset by merging the original normalized data obtained in step S2.1 with the supplementary data generated in step S2.3 to form an enhanced training dataset for subsequent model training, thereby improving the model's generalization ability under conditions of scarce or abnormal data.

[0067] S3. Execute Day-ahead Forecasts: Considering the coupling relationship between photovoltaic (PV), electricity load, and electricity price, based on historical data on a 1-hour timescale, multimodal data such as meteorological parameters and satellite cloud images are used as inputs, with PV, electricity load, and electricity price as outputs. A pre-trained first-time-series large-scale model is combined with multi-task learning techniques to achieve joint forecasting of the three. Efficient parameter fine-tuning techniques are used to adapt the time-series large-scale model to the domain, resulting in a joint forecasting large-scale model. Subsequently, based on this large-scale model, the meteorological parameters and cloud image trends for the next 24 hours are input to obtain the forecast results for PV power, electricity load, and electricity price for the next 24 hours. The day-ahead forecast results for PV power and load are used as an important reference for real-time correction.

[0068] As an optional implementation, step S3 includes the following steps: S3.1 Input and Output Definitions: Meteorological parameters and satellite cloud images are used as multimodal inputs during the large model training process, while photovoltaic power, electrical load, and electricity price at corresponding times are used as the outputs of the large model.

[0069] S3.2 Model Construction: A pre-trained time series large-scale model is used as the backbone network to construct a multi-task joint prediction architecture. In this embodiment, the Chronos-2 time series large-scale model released by Amazon or the MOMENT time series basic model released by Carnegie Mellon University are selected. These models are pre-trained on hundreds of millions of time series data and have powerful time series representation capabilities.

[0070] S3.2.1 Model Input: Multimodal data such as meteorological parameters and satellite cloud images are encoded into the model input format. The meteorological parameters are numerical time series, and the satellite cloud images are image sequences. Features are extracted separately using a multimodal encoder and then fused.

[0071] S3.2.2 Backbone Network: Employing the encoder portion of a large-scale time-series model, it performs deep modeling of the input multimodal fusion features, capturing the long-range dependencies and coupling relationships between photovoltaics, load, and electricity prices. The self-attention mechanism of the large-scale time-series model can effectively model the correlations between different time steps, exhibiting stronger expressive power and a larger receptive field compared to traditional artificial intelligence prediction methods.

[0072] S3.2.3 Multi-task output layer: Based on the hidden layer representation output by the time series large model, it simultaneously outputs photovoltaic power prediction, load prediction and electricity price prediction through a multi-task learning architecture.

[0073] S3.3 Model Fine-tuning Training: The augmented training dataset obtained in step S2.3 is used for fine-tuning training of the large vertical model. The LoRA parameter efficient fine-tuning technique is employed, updating only a small number of adaptation parameters while freezing the original weights of the large model. The loss function is defined as the weighted sum of the mean squared errors (MSEs) of the three tasks.

[0074] .

[0075] in, This represents the loss function value of the large-scale model for predicting current trends. The weighting coefficients represent the photovoltaic power prediction task. This represents the mean square error value of photovoltaic power prediction; Indicates the weighting coefficients for load forecasting tasks; This represents the mean square error value of the load forecast; This represents the weighting coefficient for the electricity price forecasting task; This represents the average absolute error value of the electricity price forecast. In this embodiment, = =0.4, =0.2.

[0076] S3.4. Based on the large model trained in step S3.3, using the meteorological parameters and satellite cloud image prediction results for the next 24 hours as input, the prediction results for photovoltaic power, load, and electricity price for the next 24 hours are obtained. The prediction results for photovoltaic power and electricity load are then used to obtain the 5-minute timescale prediction trend using the Hermitian difference method, denoted as... ( , ).

[0077] S4. Real-time Rolling Correction: Initialize the prediction time t=0, and the rolling time period s=0. Since the time-of-use pricing is on a 1-hour time scale, the real-time stage only performs rolling prediction corrections for photovoltaic power and load. Based on a single column of historical photovoltaic / load data on a 5-minute time scale, a lightweight second time-series large model is used to train the time series prediction large model; in the future data prediction process, each prediction is made for the next 5 minutes ( Photovoltaic / electrical load data (5 minutes) is collected and combined with a recursive strategy to obtain the forecast results for the next 15 minutes. A large-model-driven adaptive error limiting correction mechanism is used to correct the rolling forecast results. Error patterns are dynamically identified based on the real-time scenario, and an adaptive correction threshold is generated to constrain the 15-minute forecast results to near the day-ahead forecast trend. Rolling is performed every 15 minutes until a 24-hour forecast is completed. Real-time forecasting shortens the time resolution (from 1 hour to 5 minutes) and time span (from 24 hours to 15 minutes) of the forecast task. Combined with corrections based on day-ahead trend forecast results, this facilitates accurate forecasting of ultra-short-term fluctuations in photovoltaic power and load.

[0078] As an optional implementation, step S4 includes the following steps: S4.1 Data Preparation: Extract 5-minute granular photovoltaic / load data and satellite cloud images from historical databases to form a training dataset.

[0079] S4.2 Input and Output Definitions: Every 15 minutes, the latest real-time historical data within the first historical time window (1 hour, i.e., 12 5-minute points) is obtained as input, and the data for the next 5 minutes is used as output.

[0080] S4.3 Construction and Fine-tuning of the Real-Time Prediction Model: A lightweight time-series large-scale model is used as the backbone network for real-time prediction. LoRA technology is employed to fine-tune the lightweight time-series large-scale model, with the fine-tuning based on historical photovoltaic / load data at a 5-minute granularity. After training, two independent real-time prediction large-scale models for photovoltaics and load are obtained.

[0081] S4.4. Based on the large-scale prediction model obtained in step S4.2, perform a real-time prediction task. Input the data of 12 time steps obtained by sampling at 5-minute intervals within the latest first historical time window prepared in step S4.1 into the prediction model. The model outputs the predicted value of photovoltaic or load for the next 5 minutes. ( , The first 5-minute input data does not include predicted values, and therefore is not affected by error accumulation.

[0082] S4.5. Put the first 5-minute prediction result back into the historical data as input for the next 5-minute (10-minute) prediction process to obtain the second 5-minute prediction result, which accumulates the error of the first 5-minute prediction.

[0083] S4.6 Put the prediction results of the first two 5 minutes back into the historical data and use them as the input for the third 5-minute (15-minute) prediction process to obtain the prediction result of the 15th minute. At this time, the input data of the 15th minute prediction includes 2 prediction values, accumulating the prediction errors of the first two steps.

[0084] S4.7 Intelligent Error Limiting Correction Based on Large Model: Upgrading the traditional fixed threshold correction to an adaptive correction mechanism driven by a large model, specifically including: S4.7.1 Constructing an Error Pattern Knowledge Base: In advance, a large-scale time-series model is used to conduct in-depth analysis of historical prediction errors, clustering error characteristics under different scenarios into N typical patterns. Each pattern includes: Scenario characteristics: meteorological conditions (irradiance change rate, temperature), load characteristics (weekdays / holidays), rolling prediction step size, etc.; Error distribution: the statistical distribution of prediction errors under this scenario (mean, variance, skewness); Optimal correction strategy: the best correction parameters (threshold, correction strength) obtained through backtracking from historical data.

[0085] S4.7.2 Real-time error pattern recognition: During the current rolling forecast process, the meteorological change characteristics of the past hour, the rolling forecast error sequence, and the current day-ahead trend characteristics are input into the encoder part of the MOMENT time series basic model. After extracting the deep features, they are connected to the classification head, and the matching probability distribution of the current scene and N types of error patterns is output.

[0086] S4.7.3 Adaptive Correction Parameter Generation: Based on the matching probability distribution, the correction strategy parameters in the error pattern knowledge base are weighted and fused to obtain the adaptive correction threshold for the current scenario. and corrected strength coefficient .

[0087] S4.7.4 Execute intelligent correction: Apply an adaptive amplitude limiting function to correct the 15-minute forecast value.

[0088] .

[0089] in, This represents the predicted object after adaptive error limiting correction in the S-th rolling time period. i The predicted value; ,when When, it indicates photovoltaic power; when The time indicates the electrical load; This represents the predicted object output by the large model in the real-time rolling prediction during the S-th rolling time period. i The predicted value; This indicates the forecast object corresponding to the S-th rolling time period in the current trend forecast results. i The trend value; Indicates the corrected strength coefficient; This indicates an adaptively adjusted threshold.

[0090] S4.7.5 Correction Effect Feedback: Feedback the effect of this correction back to the error mode knowledge base for subsequent optimization of correction strategy parameters, forming a closed-loop learning mechanism.

[0091] In another exemplary embodiment of this application, the day-ahead and real-time forecasts of a regional power grid are taken as an example.

[0092] Step 1: Acquisition of multi-source data.

[0093] (1) Retrieve hourly historical data for the past three years and forecast data for the next two days for a certain region. Key meteorological parameters include: global horizontal irradiance (W / m²). 2 The data includes information such as ambient temperature (°C), cloud trajectories and cloud thickness captured by satellite images, and other data.

[0094] (2) Through the power grid dispatch data network, obtain historical and real-time data streams of “total output of photovoltaic power station (kW)” and “total active load of regional power grid (kW)” with time resolution of 1 hour and 5 minutes within the same period from the energy management system of a certain region.

[0095] (3) Obtain historical data of “node marginal electricity price (yuan / MWh)” at the hourly level within the same period.

[0096] Step 2: Data preprocessing and augmentation.

[0097] (1) Data cleaning: The quality of the data obtained above is checked. For missing values ​​caused by sensor failure or communication interruption, linear interpolation is used to fill them in. For outliers that are significantly beyond the physical meaning (e.g., GHI recorded at nighttime is greater than 10 W / m), 2 (or the load data is negative), identify and remove them.

[0098] (2) Data normalization: The cleaned meteorological (GHI, temperature), photovoltaic, load and electricity price data are subjected to maximum and minimum normalization respectively, and scaled to the [0, 1] interval.

[0099] (3) TimeGAN Data Augmentation: Normalized historical data of 1 hour granularity over three years (approximately 26,280 time points) was used as the training set and input into the TimeGAN model for training. After training, a generator was used to generate supplementary 1 hour granularity data (approximately 17,520 time points) equivalent to two years in length. These supplementary data are highly consistent with the real data in terms of statistical distribution and temporal dynamics, but contain new fluctuation patterns.

[0100] (4) Constructing an augmented dataset: The original three-year normalized data is merged with the generated two-year supplementary data to form an augmented training set containing five years of data, which is used for the next step of model training.

[0101] Step 3: Current multi-task joint trend forecast.

[0102] (1) Using hourly meteorological parameter data and satellite cloud images as inputs for multimodal variables, and photovoltaic, load and electricity price at the corresponding time as outputs, a multi-task joint prediction model is constructed using a pre-trained time series large model.

[0103] This embodiment uses the Amazon-released Chronos-2 temporal large-scale model as the backbone network and employs the LoRA parameter efficient fine-tuning technique to adapt the Chronos-2 model to the domain. A low-rank adaptation matrix is ​​inserted into the attention layer, and only these adaptation parameters are updated, while the original pre-trained weights of the large-scale model are frozen.

[0104] (2) Perform day-ahead forecasting and interpolation processing. Based on the meteorological forecast data for the next 24 hours, forecast the photovoltaic load and electricity price for the next 24 hours. On this basis, use the piecewise cubic Hermit interpolation method to obtain the trend values ​​of photovoltaic load and electricity load on a 5-minute time scale. Store the trend values ​​in the database as a benchmark for real-time correction.

[0105] Step 4: Real-time rolling correction of forecasts.

[0106] (1) Real-time prediction model training: Three years of 5-minute granular photovoltaic and load data were extracted from the historical database to construct a real-time prediction training set. A lightweight time-series large model was used as the backbone network for real-time prediction. In this embodiment, the Chronos-Bolt model was selected, which is an accelerated version of the Chronos series and is suitable for real-time prediction scenarios. The data of the past 1 hour (5-minute time resolution, 12 data points) was used as the model input, and the data of the next 5 minutes (1 point) was used as the output. LoRA technology was used to fine-tune the Chronos-Bolt, and the fine-tuning data was the 5-minute granular historical photovoltaic / load data. After training, two independent real-time prediction large models for photovoltaic and load were obtained, which were used to learn the ultra-short-term fluctuation patterns.

[0107] (2) Real-time forecasting. Taking electricity load forecasting as an example, assuming the current time is 10:50, the electricity load data of the previous hour (09:50 to 10:50) is obtained and input into the corresponding Chronos-Bolt model to obtain the forecast result of the first 5 minutes in the future (10:55).

[0108] (3) Put the prediction result at 10:55 into the input data for the second 5-minute prediction (11:00) (from 09:55 to 10:55) to get the prediction result at 11:00.

[0109] (4) Put the prediction data of 10:55 and 11:00 into the input data of the third 5-minute prediction (11:05) (from 10:00 to 11:00) to get the prediction result of 11:05.

[0110] (5) Rolling forecast constraint correction: only the forecast value of the last forecast point (11:05) is corrected. Read the day-ahead trend value corresponding to the hour (11:00-12:00) where 11:05 is located, obtained in step (2), and apply the limiting function to correct the forecast result for 11:05, such as... Figure 4 As shown, a schematic diagram of the electrical load prediction effect is provided.

[0111] (6) The system enters a waiting state. At 11:05, it repeats steps (2)-(5) to perform the next round of rolling prediction. This cycle continues until the prediction task for the day ends.

[0112] In summary, this application has the following beneficial effects: First, it has high prediction accuracy: TimeGAN data augmentation improves the data foundation for training large models; multi-task learning joint modeling utilizes the correlation information between variables; and two-stage prediction takes into account both trends and details.

[0113] Second, it is highly robust: the data augmentation process reduces the model's sensitivity to data noise and missing data; the constraint correction mechanism in the real-time stage avoids abnormal deviations in the prediction results.

[0114] Third, it is highly practical: the two-stage design meets the actual dispatching needs of the power system. The day-ahead stage provides a basis for market transactions and planning, while the real-time stage provides precise input for controls such as AGC (Automatic Generation Control).

[0115] Fourth, it is highly innovative: Based on the "day-ahead trend prediction - real-time constraint correction" strategy, and combined with artificial intelligence technology to train a large vertical model, it effectively alleviates error accumulation and achieves highly accurate prediction of photovoltaic and electrical loads.

[0116] Based on the same inventive concept, this application also provides a photovoltaic-load-price forecasting device for implementing the photovoltaic-load-price forecasting method involving the fusion of large models and error limiting correction as described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more photovoltaic-load-price forecasting device embodiments involving the fusion of large models and error limiting correction provided below can be found in the limitations of the photovoltaic-load-price forecasting method involving the fusion of large models and error limiting correction described above, and will not be repeated here.

[0117] In one exemplary embodiment, such as Figure 5 As shown, a photovoltaic-load-price forecasting device integrating a large model and error limiting correction is provided, comprising: The forecast meteorological data acquisition module 301 is used to acquire forecast meteorological data for the time period to be predicted; the forecast meteorological data includes meteorological parameter forecast data and satellite cloud image forecast data.

[0118] The day-ahead trend prediction module 302 is used to input the predicted meteorological data for the time period to be predicted into the trained day-ahead trend prediction model to obtain the day-ahead trend prediction results for the time period to be predicted. The day-ahead trend prediction results include photovoltaic power prediction values, load prediction values, and electricity price prediction values. The day-ahead trend prediction model is obtained by fine-tuning the parameters of a pre-trained first time-series model using a first sample dataset. The first sample dataset is collected from a multi-modal dataset of historical time periods at a first sampling interval. The multi-modal dataset includes real meteorological parameter data, real satellite cloud image data, and corresponding real values ​​of photovoltaic power, electricity load, and electricity price.

[0119] The real-time data acquisition module 303 is used to acquire the real values ​​of photovoltaic power and electrical load collected at the second sampling interval within the first historical time window before the time period to be predicted.

[0120] The real-time rolling prediction module 304 is used to input the real values ​​of photovoltaic power and electrical load collected at a second sampling interval within a first historical time window before the time period to be predicted into a trained real-time rolling prediction model to obtain the source load prediction results for the time period to be predicted; the source load prediction results include photovoltaic power prediction values ​​and load prediction values; the real-time rolling prediction model is obtained by fine-tuning the parameters of a pre-trained second time series model using a second sample dataset, the second sample dataset being collected at a second sampling interval from the real values ​​of photovoltaic power and electrical load in the multimodal dataset of historical time periods.

[0121] The error limiting correction module 305 is used to adaptively limit the error of the source load prediction result based on the day-ahead trend prediction result to obtain the final source load prediction result for the time period to be predicted; the final source load prediction result includes the electricity price prediction value, the photovoltaic power prediction value after error limiting correction, and the load prediction value in the day-ahead trend prediction result.

[0122] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 6 As 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 operating system and computer programs stored in the non-volatile storage media. The database stores photovoltaic-load-price forecasting processing data that integrates a large model and error limiting correction. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a photovoltaic-load-price forecasting method that integrates a large model and error limiting correction.

[0123] Those skilled in the art will understand that Figure 6 The structures shown are merely block diagrams of some structures related to the present application and do 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 shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0124] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0125] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0126] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, 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.

[0127] Those skilled in the art will understand that all or part of the processes in 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, and when executed, it can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile 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).

[0128] 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, etc., and are not limited to these.

[0129] 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 specification.

[0130] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A photovoltaic-load-electricity price forecasting method integrating a large model and error limiting correction, characterized in that, include: Obtain forecast meteorological data for the time period to be predicted; the forecast meteorological data includes meteorological parameter forecast data and satellite cloud image forecast data; The predicted meteorological data for the time period to be predicted is input into the trained daytime trend prediction model to obtain the daytime trend prediction results for the time period to be predicted. The daytime trend prediction results include photovoltaic power prediction, load prediction, and electricity price prediction. The daytime trend prediction model is obtained by fine-tuning the parameters of a pre-trained first time series model using a first sample dataset. The first sample dataset is obtained by collecting multimodal datasets of historical time periods at a first sampling interval. The multimodal dataset includes real meteorological parameter data, real satellite cloud image data, and corresponding real values ​​of photovoltaic power, electricity load, and electricity price. Obtain the true values ​​of photovoltaic power and electrical load collected at the second sampling interval within the first historical time window before the time period to be predicted; The real values ​​of photovoltaic power and electrical load, collected at the second sampling interval within the first historical time window before the time period to be predicted, are input into the trained real-time rolling prediction model to obtain the source load prediction results for the time period to be predicted. The source load prediction results include photovoltaic power prediction values ​​and load prediction values. The real-time rolling prediction model is obtained by fine-tuning the parameters of the pre-trained second time series model using the second sample dataset, which is obtained by collecting the real values ​​of photovoltaic power and electrical load from the multimodal dataset of historical time periods at the second sampling interval. Based on the current day trend forecast results, the source load forecast results are adaptively error-limited and corrected to obtain the final source load forecast results for the time period to be predicted. The final source-load forecast results include the electricity price forecast, the photovoltaic power forecast after error limit correction, and the load forecast from the day-ahead trend forecast results.

2. The photovoltaic-load-price forecasting method integrating a large model and error limiting correction as described in claim 1, characterized in that, The parameter fine-tuning process of the current trend forecasting model includes: The real meteorological parameter data and real satellite cloud image data in the first sample dataset are used as inputs to the pre-trained first time series large model. The pre-trained first time series large model is fine-tuned using a parameter-efficient fine-tuning method. The model outputs the photovoltaic power prediction value, electricity load prediction value, and electricity price prediction value corresponding to the real meteorological parameter data and real satellite cloud image data. Training stops when the loss function value is less than the preset convergence threshold or the preset maximum number of iterations is reached, and the trained day-ahead trend prediction large model is obtained. The loss function is constructed based on the photovoltaic power prediction value, electricity load prediction value, electricity price prediction value, and the corresponding real photovoltaic power value, electricity load value, and electricity price value. The first time-series large model comprises a multimodal coding layer, a Transformer backbone layer, and a multi-task output layer connected in sequence. The multimodal coding layer encodes meteorological parameter data and satellite cloud image data into a unified representation to obtain a multimodal fusion feature vector. The Transformer backbone layer extracts time-series features based on the multimodal fusion feature vector through a self-attention mechanism to obtain a time-series feature representation. The multi-task output layer simultaneously outputs photovoltaic power prediction, load prediction, and electricity price prediction based on the time-series feature representation through a multi-task learning architecture.

3. The photovoltaic-load-electricity price forecasting method integrating a large model and error limiting correction as described in claim 1, characterized in that, The methods for constructing the first and second sample datasets specifically include: Obtain a first raw dataset collected from a multimodal dataset within a historical time period at a first sampling interval, and a second raw dataset collected from the actual photovoltaic power and actual electrical load values ​​in the multimodal dataset within a historical time period at a second sampling interval; The first original multimodal dataset and the second original multimodal dataset are preprocessed to obtain the preprocessed first original dataset and the second original dataset; the preprocessing operations include data cleaning and normalization. Based on the preprocessed first and second original datasets, time-series generative adversarial networks are used to generate a first supplementary dataset and a second supplementary dataset, respectively. The first supplementary dataset has the same data volume as the preprocessed first original dataset but different data values. The second supplementary dataset has the same data volume as the preprocessed second original dataset but different data values. The first supplementary dataset is merged with the preprocessed first original dataset to obtain the first sample dataset. The second supplementary dataset is merged with the preprocessed second original dataset to obtain the second sample dataset.

4. The photovoltaic-load-electricity price forecasting method integrating a large model and error limiting correction as described in claim 1, characterized in that, The actual photovoltaic power and electrical load values ​​collected at the second sampling interval within the first historical time window before the time period to be predicted are input into the trained real-time rolling prediction model to obtain the source-load prediction results for the time period to be predicted, specifically including: The time period to be predicted is divided into multiple consecutive rolling periods; each rolling period contains multiple consecutive rolling time periods divided according to the second sampling interval. The following steps are used to predict the photovoltaic power and load forecast values ​​for all rolling time periods within the k-th rolling cycle; where k = {1, 2, ..., K}, s = {1, 2, ..., S}, K represents the total number of rolling cycles, and S represents the total number of rolling time periods within a rolling cycle; When s=1, the real values ​​of photovoltaic power and electrical load collected at the second sampling interval within the first historical time window before the start time of the kth rolling cycle are used as the input sequence and input into the trained real-time rolling prediction model to predict the source load prediction result for the first rolling time period within the kth rolling cycle. When s≥2, the source load prediction results from the first rolling time period to the (s-1)th rolling time period generated in the kth rolling cycle are added to the input sequence, and the real values ​​of photovoltaic power and electrical load with the earliest rolling time period of (s-1)th rolling time period are removed from the input sequence to form the updated input sequence. The updated input sequence is then input into the trained real-time rolling prediction model to obtain the source load prediction results for the sth rolling time period in the kth rolling cycle, until s=S, thus completing the prediction for the kth rolling cycle. Let k = k + 1, and repeat the above prediction steps for the rolling cycle until k = K, to obtain the source load prediction results for the time period to be predicted.

5. The photovoltaic-load-electricity price forecasting method integrating a large model and error limiting correction as described in claim 4, characterized in that, Based on the day-ahead trend forecast results, an adaptive error limiting correction is performed on the source load forecast results, specifically including: When s=S within each rolling cycle, the source load prediction result for the S-th rolling time period is adaptively error-limited and corrected using the following formula: ； in, This represents the predicted object after adaptive error limiting correction in the S-th rolling time period. i The predicted value; ,when When, it indicates photovoltaic power; when When, it indicates the electrical load; This represents the predicted object output by the large model in the real-time rolling prediction during the S-th rolling time period. i The predicted value; This indicates the forecast object corresponding to the S-th rolling time period in the current trend forecast results. i The trend value; Indicates the corrected strength coefficient; This indicates an adaptive threshold adjustment.

6. The photovoltaic-load-electricity price forecasting method integrating a large model and error limiting correction as described in claim 2, characterized in that, The loss function for the current trend prediction model is: ； in, This represents the loss function value of the large-scale model for predicting current trends. The weighting coefficients represent the photovoltaic power prediction task. This represents the mean square error value of photovoltaic power prediction; Indicates the weighting coefficients for load forecasting tasks; This represents the mean square error value of the load forecast; This represents the weighting coefficient for the electricity price forecasting task; This represents the average absolute error of the electricity price forecast.

7. A photovoltaic-load-price forecasting device integrating a large model and error limiting correction, characterized in that, The photovoltaic-load-price forecasting device integrating a large model and error limiting correction uses the photovoltaic-load-price forecasting method integrating a large model and error limiting correction as described in any one of claims 1-6. The photovoltaic-load-price forecasting device integrating a large model and error limiting correction comprises: The forecast meteorological data acquisition module is used to acquire forecast meteorological data for the time period to be predicted; the forecast meteorological data includes meteorological parameter forecast data and satellite cloud image forecast data; The day-ahead trend prediction module is used to input the predicted meteorological data for the time period to be predicted into a pre-trained day-ahead trend prediction model to obtain the day-ahead trend prediction results for the time period to be predicted. The day-ahead trend prediction results include photovoltaic power prediction values, load prediction values, and electricity price prediction values. The day-ahead trend prediction model is obtained by fine-tuning the parameters of a pre-trained first time-series model using a first sample dataset. The first sample dataset is collected from multi-modal datasets of historical time periods at a first sampling interval. The multi-modal dataset includes real meteorological parameter data, real satellite cloud image data, and corresponding real values ​​of photovoltaic power, electricity load, and electricity price. The real-time data acquisition module is used to acquire the real values ​​of photovoltaic power and electrical load collected at the second sampling interval within the first historical time window before the time period to be predicted. The real-time rolling prediction module is used to input the actual photovoltaic power and electrical load values ​​collected at a second sampling interval within a first historical time window before the time period to be predicted into a trained real-time rolling prediction model to obtain the source load prediction results for the time period to be predicted. The source load prediction results include photovoltaic power prediction values ​​and load prediction values. The real-time rolling prediction model is obtained by fine-tuning the parameters of a pre-trained second time series model using a second sample dataset, which is obtained by collecting the actual photovoltaic power and electrical load values ​​from a multimodal dataset of historical time periods at a second sampling interval. The error limiting correction module is used to adaptively limit the source load prediction results based on the day-ahead trend prediction results to obtain the final source load prediction results for the time period to be predicted; the final source load prediction results include the electricity price prediction value, the photovoltaic power prediction value after error limiting correction, and the load prediction value in the day-ahead trend prediction results.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the photovoltaic-load-price forecasting method with fusion of large model and error limiting correction as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the photovoltaic-load-price forecasting method according to any one of claims 1-6, which integrates a large model with error limiting correction.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the photovoltaic-load-price forecasting method according to any one of claims 1-6, which integrates a large model with error limiting correction.

Images