A weather-windlight power storage combined prediction method and system based on an ultra-long context

Abstract

The application discloses a kind of meteorology-sight-power storage combined prediction method and system based on super-long context, belong to power system prediction technical field.The method constructs the multi-agent collaborative architecture of data cleaning, time alignment, context generation and combined prediction, realizes the automatic processing and efficient organization of wind power, photovoltaic and energy storage historical data;By introducing the super-long context input mechanism across seasons, across years, break through the context length limit of traditional model, realize the unified modeling and collaborative representation of different frequency multi-source data;On this basis, carry out zero-sample combined prediction using time series large model, and combine wind power curve, photovoltaic irradiance and energy storage state of charge and other physical constraints, build wind-sight-storage coupling consistency correction mechanism, optimize the prediction result.The method of the application effectively improves the multi-source covariant fusion capability and prediction accuracy, with good generalization performance and engineering adaptability.

Description

Technical Field

[0001] This invention relates to a method and system for joint forecasting of meteorological, wind, solar and energy storage power based on ultra-long context, belonging to the field of power system forecasting technology. Background Technology

[0002] With the rapid increase in the proportion of wind power, photovoltaic power, and energy storage in power systems, renewable energy power forecasting has become an important foundation for smart grid operation, dispatch optimization, and electricity market transactions. However, traditional meteorological and power forecasting methods have several technical limitations and are difficult to meet the needs of modern renewable energy systems.

[0003] First, existing prediction methods based on statistical models (such as ARIMA and VAR) or machine learning models (such as SVR and random forest) have limited contextual modeling capabilities when dealing with long-sequence, multi-seasonal, and multi-year meteorological and power data. They cannot effectively capture long-term periodic and interannual climate change patterns, resulting in limited prediction accuracy.

[0004] Secondly, while traditional deep learning models (such as LSTM, GRU, Transformer, etc.) perform well in short-term predictions, they are limited by the length of the input sequence and the scale of the training data, and can usually only use a limited historical window (such as hundreds to thousands of steps), making it difficult to achieve long-term dependency learning across years or seasons. At the same time, when dealing with different frequency data, wind power, photovoltaic and energy storage data often have different sampling frequencies and time deviations, and existing methods are difficult to unify and align them, leading to the accumulation of prediction errors.

[0005] Finally, existing wind, solar, and energy storage forecasting methods often employ a sequential process of "predicting weather first, then predicting power," or model wind power, solar power, and energy storage power separately. This lack of joint forecasting and coupling consistency constraints makes them prone to error propagation and unmet physical constraints. Furthermore, existing methods typically rely on simple interpolation or removal to handle outliers and missing values, failing to fully utilize the reasoning capabilities of large models on incomplete sequences and unable to effectively integrate third-party weather forecast information as auxiliary covariates.

[0006] Therefore, existing technologies cannot simultaneously meet the comprehensive requirements of zero-sample generalization, multi-source heterogeneous frequency data processing, joint prediction of wind, solar and energy storage, covariate fusion and physical consistency constraints. A new technical solution is needed to solve the above problems. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a meteorological-wind-solar-storage power joint prediction method and system based on ultra-long context. It solves the problems of limited context length of traditional models, difficulty in uniformly processing different frequency data, and insufficient covariate fusion capability. It has the advantages of strong zero-sample generalization ability, high degree of domestic controllability, and strong engineering adaptability.

[0008] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0009] On the one hand, this invention provides a method for joint prediction of meteorological, wind-solar-storage power based on ultra-long context, including:

[0010] Obtain historical data on wind, solar, and energy storage;

[0011] Based on a large model with ultra-long context, the combination method of historical data on the time scale is determined according to historical wind, solar and energy storage data.

[0012] Based on the combination of historical data on a time scale, the filtered historical data of wind, solar and energy storage is obtained by filtering from the historical data of wind, solar and energy storage.

[0013] Using the filtered historical data of wind, solar and energy storage as input, the predicted power data of wind, solar and energy storage at the target time is obtained by making predictions based on the ultra-long context large model.

[0014] The wind-solar-storage power prediction data at the target time are subjected to wind-solar-storage coupling consistency constraints, and the constraint results are used to correct the wind-solar-storage power prediction data.

[0015] Furthermore, the historical data for wind, solar, and energy storage includes power data, measured meteorological data, and equipment operation data. The power data includes the active power of the wind farm, the active power of the photovoltaic power station, the charging and discharging power of the energy storage system, and the grid-connected output power, expressed as:

[0016] ;

[0017] in, Represents power data. This indicates the active power of the wind farm. Indicates the active power of a photovoltaic power station. Indicates the charging and discharging power of the energy storage system. Indicates grid-connected output power;

[0018] The measured meteorological data includes wind speed, wind direction, horizontal irradiance, temperature, humidity, and air pressure, and its expression is:

[0019] ;

[0020] in, This represents measured meteorological data. Indicates wind speed. Indicates wind direction. Indicates horizontal irradiance. Indicates temperature. Indicates humidity. Indicates air pressure;

[0021] The equipment operation data includes unit availability, fault identification, power limitation status, and remaining energy storage.

[0022] Furthermore, the method for determining the combination of historical data on a time scale based on the ultra-long context large model and historical wind, solar and energy storage data includes:

[0023] Based on dynamic time warping enhancement, the historical data of wind, solar and energy storage are time-aligned to obtain a preliminary time series of historical data of wind, solar and energy storage.

[0024] Based on the ultra-long context large model, outlier detection is performed on the preliminary wind, solar and energy storage historical data time series to obtain the wind, solar and energy storage historical data time series.

[0025] Using historical time series data of wind, solar and energy storage as input, the task semantic adaptive reasoning is performed based on a large model with ultra-long context to obtain the combination of historical data on the time scale.

[0026] Furthermore, the preliminary wind, solar, and energy storage historical data time series obtained by time alignment of the historical data based on dynamic time warping enhancement includes:

[0027] The optimal time mapping path is obtained by minimizing the similarity error between any two data sequences in the historical wind, solar, and energy storage data over time, using this as the objective function. The expression for the objective function is as follows:

[0028] ;

[0029] ;

[0030] ;

[0031] in, Indicates the first A set of wind, solar and energy storage data awaiting time alignment. Represents the baseline timeline sequence. Represents a local distance metric function. express and Similarity in the time dimension Represents a set of time-mapped paths. Indicates the time mapping path, Indicates the first The first of the wind, solar and energy storage data awaiting time alignment One element, Represents the first in the reference timeline sequence One element, This represents the weight hyperparameter. This indicates a global alignment path consistency regular expression. This indicates the number of wind, solar, and storage data points awaiting time alignment. Indicates the first The wind, solar, and energy storage data awaiting time alignment are in the first... The corresponding values ​​at each reference time point Indicates variance operation;

[0032] Preliminary time series of historical wind, solar and energy storage data are obtained by resampling the historical data of wind, solar and energy storage using the optimal time mapping path.

[0033] Furthermore, the step of performing outlier detection on the preliminary wind, solar, and energy storage historical data time series based on an ultra-long context large model to obtain the wind, solar, and energy storage historical data time series includes:

[0034] S1. Using the preliminary historical data time series of wind, solar and energy storage and the meteorological conditions at the current moment as input, the model with ultra-long context determines whether the wind, solar and energy storage data at the current moment is abnormal based on the relationship between meteorological conditions and wind, solar and energy storage power. If abnormal, the data is set to empty; otherwise, it is retained, and the preliminary historical data time series of wind, solar and energy storage after anomaly detection is obtained.

[0035] S2. Using the preliminary wind, solar and energy storage historical data time series after outlier detection and the meteorological conditions at the next moment as input, repeat S1 until the wind, solar and energy storage historical data time series is obtained.

[0036] Furthermore, the expression for the historical data combination method on the aforementioned time scale is:

[0037] ;

[0038] in, This indicates how historical data is combined on a given time scale. Indicates the length of the most recent consecutive historical data. Indicates the length of periodic historical data. Indicates the length of historical data spanning multiple years. Indicates the length of similar meteorological sample data.

[0039] Furthermore, the method for screening similar meteorological sample data includes:

[0040] Traverse the historical time series of wind, solar, and solar storage data, and calculate its similarity to the meteorological data at the target time. The similarity calculation expression is as follows:

[0041] ;

[0042] in, express real-time weather data and Similarity of meteorological data at different times express Real-time weather data, express Real-time weather data, Indicates the modulus;

[0043] Historical wind, solar and energy storage data with similarity greater than a set threshold are combined to form similar meteorological sample data.

[0044] Furthermore, when meteorological forecast data for the target time is available, the ultra-long context large model uses the meteorological forecast data for the target time as covariates to achieve joint prediction. The modeling form of the joint prediction is reconstructed into a conditional generation problem, the expression of which is:

[0045] ;

[0046] in, This represents a large model inference mapping with very long contexts. This indicates a very long context composed of filtered historical data on wind, solar, and energy storage. This represents the covariates composed of meteorological forecast data at the target time. This represents the wind, solar, and energy storage power data at the target time. Indicates the prediction step, This indicates the start time point of the prediction.

[0047] Furthermore, the expression for the wind-solar-storage coupling consistency constraint is as follows:

[0048] ;

[0049] ;

[0050] ;

[0051] ;

[0052]

[0053] in, This represents the active power of the wind farm at the target time. This represents the active power of the photovoltaic power station at the target time. This represents the charging and discharging power of the energy storage system at the target time. This represents the grid-connected output power at the target time. This represents the charging power at the time preceding the target time. This represents the discharge power at the time preceding the target time. Indicates the time step. Indicates the rated charging power. Indicates the rated discharge power. This represents the stored energy at the target time. This represents the stored energy at the previous time step before the target time. This represents the minimum energy storage capacity. Indicates the maximum value of stored energy. Indicates the maximum power of the power grid. This represents the active power of the wind farm at the corrected target time. This represents the active power of the photovoltaic power plant at the corrected target time. This represents the corrected charging and discharging power of the energy storage system at the target time. Indicates the prediction step, This indicates the start time point of the prediction.

[0054] Furthermore, it also includes an uncertainty assessment of the wind, solar, and energy storage forecast data, the calculation expression for which the uncertainty is:

[0055] ;

[0056] ;

[0057]

[0058]

[0059]

[0060] in, This represents the predicted mean. Indicates the first Predicted data for secondary wind, solar, and energy storage power. Indicates the number of predictions. Indicates the standard deviation of the forecast. This represents the quantile value of the standard normal distribution. Indicates the lower bound of confidence. Indicates the upper bound of confidence. Indicates the width of the interval.

[0061] On the other hand, the present invention also provides a meteorological-wind-solar-storage power joint prediction system based on ultra-long context, for implementing the meteorological-wind-solar-storage power joint prediction method based on ultra-long context as described in any of the above claims, comprising:

[0062] The wind, solar and energy storage historical data acquisition module is configured to acquire historical wind, solar and energy storage data.

[0063] The time-scale data combination determination module is configured to determine the combination method of historical data on the time scale based on the wind, solar and energy storage historical data, using a large model with ultra-long context.

[0064] The wind, solar and energy storage historical data filtering module is configured to filter the wind, solar and energy storage historical data from the historical data based on the combination of historical data on the time scale.

[0065] The wind, solar and energy storage power prediction module is configured to take the filtered historical wind, solar and energy storage data as input and predict the wind, solar and energy storage power data at the target time based on the ultra-long context large model.

[0066] The prediction result correction module is configured to perform wind-solar-storage coupling consistency constraints on the wind-solar-storage power data at the target time, and use the constraint results to correct the wind-solar-storage power prediction data.

[0067] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: Strong zero-shot generalization ability: This invention constructs an ultra-long contextual input structure with millions of tokens, unifying the organization of historical meteorological data across seasons and years with wind, solar, and energy storage operation data. Combined with a covariate fusion mechanism based on contextual learning (ICL), the large model can directly complete prediction tasks without fine-tuning for specific power plants. Compared to traditional models that rely on large amounts of labeled data and targeted training, this method fully utilizes the general modeling capabilities of the large model, enabling it to automatically learn data distribution characteristics and potential patterns from the context, achieving rapid adaptation to power plants with different geographical regions, climatic conditions, and equipment configurations. This capability significantly reduces model deployment and migration costs, improves the application efficiency of the system in new power plants or data-scarce scenarios, and has excellent generalization performance and practical engineering application value.

[0068] Support for multi-source, multi-frequency data processing: Addressing the common issues in wind power, photovoltaic, and energy storage systems such as diverse data sources, inconsistent sampling frequencies, and timestamp offsets, this invention designs a collaborative data cleaning and time alignment mechanism. The data cleaning module detects and removes outliers and standardizes missing values, ensuring data quality. Simultaneously, the time alignment mechanism reconstructs a unified timeline for data at different time scales (e.g., minute-level, hour-level) and automatically corrects time offsets, achieving a unified representation of multi-source, multi-frequency data. Based on this, a context construction strategy efficiently organizes the processed data into a continuous sequence input model, enabling large models to simultaneously capture multi-time-scale features and cross-variable relationships. This method effectively improves data utilization and prediction stability, solving the problem of traditional models struggling to handle complex data structures.

[0069] Joint Prediction and Physical Consistency Constraints: This invention breaks through the traditional single-variable prediction model, constructing a joint prediction framework for wind power, photovoltaic power, and energy storage power. This enables the model to simultaneously output multi-variable future states and learn their inherent coupling relationships. After the prediction results are generated, a physical consistency constraint mechanism is introduced to verify and correct the results, including constraints on wind power curves, the relationship between photovoltaic irradiance and power, and the state of charge (SOC) constraints of energy storage, thereby ensuring that the prediction results conform to actual physical laws. This mechanism effectively avoids the error propagation and inconsistency problems caused by independent multi-variable predictions, improving the overall reliability and usability of the prediction. By combining data-driven methods with physical mechanisms, this invention significantly enhances the model's adaptability and engineering reliability in complex power systems.

[0070] This invention integrates third-party weather forecasts: It introduces a flexible external information access mechanism during covariate encoding, allowing third-party weather forecast data (such as Numerical Weather Prediction (NWP) data) to be used as conditional inputs to construct a context sequence together with historical observation data. Through a unified encoding method, large models can comprehensively utilize historical data and future weather forecast information within the context, achieving more accurate inferences about future power change trends. This mechanism not only enhances the model's response to extreme weather and abrupt changes but also improves the predictive foresight and stability. Furthermore, this method exhibits good compatibility with external data sources, allowing for flexible integration with different meteorological services according to actual needs, thus improving the system's scalability and practical value.

[0071] Domestically Controllable: This invention is built upon a domestically developed and controllable large-scale model system, supporting both local deployment and API call modes, allowing for flexible selection of deployment methods based on actual application scenarios. In local deployment mode, the system can run on enterprise intranets or dedicated computing platforms, ensuring data security and privacy; in API call mode, it enables efficient resource utilization and rapid expansion. By adopting a domestic software and hardware ecosystem, this method achieves independent control over key technologies, reducing dependence on external technologies and meeting the high requirements of the power industry for security, stability, and compliance. Furthermore, this architecture facilitates integration with existing power system platforms and possesses strong engineering implementation capabilities.

[0072] Uncertainty Output: To enhance the reference value of prediction results in actual scheduling, this invention introduces an uncertainty assessment mechanism, providing corresponding confidence intervals or probability distribution information along with the output prediction mean. This mechanism statistically models the prediction results by analyzing the sources of uncertainty in the model output and contextual information, thereby quantifying the prediction error range. Compared to traditional methods that only provide point predictions, this invention provides more comprehensive information support for grid scheduling, energy storage optimization control, and risk decision-making, enabling decision-makers to formulate more reasonable operating strategies under different risk preferences. Furthermore, uncertainty information can also be used for model performance evaluation and continuous optimization, improving the overall robustness of the system.

[0073] Multi-Agent Collaborative Processing: This invention employs a modular multi-agent architecture, encapsulating functions such as data cleaning, time alignment, context construction, joint prediction, and result verification into independent agents. A collaborative mechanism automates the overall process. Agents communicate and transfer data through standardized interfaces, ensuring both system structural clarity and flexibility in functional expansion. When new functional requirements or data types are added, only the corresponding agent modules need to be expanded or replaced, without requiring a large-scale reconstruction of the entire system. This architecture not only improves system maintainability and scalability but also supports distributed deployment and parallel computing, contributing to increased processing efficiency and meeting the needs of large-scale engineering applications. Attached Figure Description

[0074] Figure 1 This is a flowchart illustrating a method for joint prediction of meteorological, wind, solar, and energy storage power based on ultra-long context in one embodiment of the present invention. Detailed Implementation

[0075] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0076] Example 1:

[0077] like Figure 1 As shown, this embodiment of the invention provides a joint prediction method for meteorological-wind-solar-storage power based on ultra-long context, including the following steps:

[0078] To acquire wind, solar and energy storage data, considering that this embodiment relies on millions of context input capabilities and a zero-sample inference mechanism, data acquisition includes not only real-time data collection but also multi-year historical data.

[0079] In this embodiment, the wind, solar, and energy storage data includes power data, measured meteorological data, equipment operation data, and optional weather forecast data. Specifically, the power data includes, but is not limited to, the active power of the wind farm, the active power of the photovoltaic power station, the charging and discharging power of the energy storage system, and the grid-connected output power, and its expression is:

[0080] ;

[0081] in, Represents power data. This indicates the active power of the wind farm. Indicates the active power of a photovoltaic power station. Indicates the charging and discharging power of the energy storage system. This indicates the grid-connected output power.

[0082] Measured meteorological data include wind speed, wind direction, horizontal irradiance, temperature, humidity, and air pressure, and their expression is:

[0083] ;

[0084] in, This represents measured meteorological data. Indicates wind speed. Indicates wind direction. Indicates horizontal irradiance. Indicates temperature. Indicates humidity. Indicates air pressure;

[0085] Equipment operation data includes unit availability, fault identification, power limitation status, and remaining energy storage. The equipment operation data is mainly used to assist in subsequent anomaly detection.

[0086] Data access methods support both local deployment and proxy API calls to adapt to domestic power system environments and cloud deployment scenarios.

[0087] In terms of time scale design, the wind, solar and storage data in this embodiment supports multi-frequency data access, including different sampling frequencies such as 1 minute, 5 minutes, 15 minutes, 30 minutes and 1 hour. There is no fixed limit on the frequency. Instead, the data is collected and stored uniformly through timestamps to provide raw data input for subsequent time alignment.

[0088] Regarding the length of historical data, this embodiment retains no less than 3 years of historical data, and preferably 5 years or more of historical data when conditions permit. This is because the power of new energy sources is significantly affected by seasonal cycles, interannual climate change and extreme weather events, and multi-year data helps to achieve cross-year sample retrieval in the ultra-long context construction stage.

[0089] Because different data sources often have different sampling frequencies, sampling time offsets, or nonlinear time differences in wind, solar, and energy storage forecasting, simply aligning them based on timestamps can lead to information distortion. To improve the fusion quality, this embodiment designs a joint alignment mechanism based on Dynamic Time Warping (DTW) enhancement, combined with statistical and learning-based matching strategies, to make time alignment more refined and robust. Specifically, it includes:

[0090] To measure the time synchronization error between different sequences, an alignment distance function is constructed to measure the similarity error of any two data sequences in the historical wind, solar, and energy storage data along the time dimension. The objective function is to minimize the overall similarity error while ensuring that each data sequence is mutually aligned. The expression of the objective function is as follows:

[0091] ;

[0092] ;

[0093] ;

[0094] in, Indicates the first A set of wind, solar and energy storage data awaiting time alignment. Represents the baseline timeline sequence. Represents a local distance metric function. express and Similarity in the time dimension Represents a set of time-mapped paths. Indicates the time mapping path, Indicates the first The first of the wind, solar and energy storage data awaiting time alignment One element, Represents the first in the reference timeline sequence One element, This represents the weight hyperparameter. This represents a global alignment path consistency regularization term, used to balance overall alignment quality with the temporal relationships within each sequence. This indicates the number of wind, solar, and storage data points awaiting time alignment. Indicates the first The wind, solar, and energy storage data awaiting time alignment are in the first... The corresponding values ​​at each reference time point This indicates variance calculation.

[0095] After obtaining the optimal path mapping based on the objective function, each data sequence is resampled according to the optimal path mapping to form a preliminary wind, solar and energy storage historical data time series with a unified time step.

[0096] Using the initial historical time series of wind, solar and energy storage data, a structured context is constructed by combining historical power change trends, multi-dimensional meteorological conditions, equipment operation logic, and temporal semantic information (day and night, season, etc.). A large model with an ultra-long context performs causal consistency inference to determine whether the data at each moment conforms to physical and operational laws. If the data at the current moment is determined to be an outlier, the value is set to empty; otherwise, it is retained, thus obtaining the historical time series of wind, solar and energy storage data.

[0097] We construct an ultra-long context using historical time series data of wind, solar and energy storage, and use it as input. Based on the ultra-long context large model, we perform task semantic adaptive reasoning to obtain the combination of historical data on the time scale.

[0098] In this embodiment, the construction of the ultra-long context does not employ a fixed-length sliding window approach, but rather dynamically generates it through a multi-round dialogic adaptive planning mechanism involving large models in decision-making. Since current mainstream large models support input capabilities of millions of tokens, this embodiment fully utilizes this capability, organizing multi-year historical power data, historical meteorological data, and optional third-party meteorological forecast information into a structured context input. However, it does not simply stack historical data, but rather constructs it semantically around a specific prediction task.

[0099] Specifically, before the forecast begins, a structured task description is first input into the large model, including the forecast target (wind power, solar power, energy storage, or a combination thereof), the forecast time range, the site area information, the current time, and the maximum range of available historical data. The large model performs semantic reasoning based on the renewable energy power formation mechanism, meteorological driving characteristics, and periodic patterns to determine the required historical reference depth to complete the forecast task and provides suggestions on the length of historical data, such as whether recent consecutive hours of trend data are needed, whether cross-day periodic samples are needed, whether cross-year comparison samples are needed, and whether samples of historically similar meteorological conditions are needed.

[0100] Building upon this, a second round of interaction with the large model guides it to analyze potential special meteorological scenarios during the forecast period, such as cold waves, strong winds, low radiation, or severe convective weather. This further determines whether it's necessary to expand the historical retrieval scope to include historical power response data under extreme weather conditions or similar weather conditions. Through these multiple rounds of semantic interaction, the large model outputs a structured historical data selection scheme, including combinations of historical fragments at different time scales, expressed as follows:

[0101] ;

[0102] in, This indicates how historical data is combined on a given time scale. Indicates the length of the most recent consecutive historical data. Indicates the length of periodic historical data. Indicates the length of historical data spanning multiple years. Indicates the length of similar meteorological sample data.

[0103] Based on the combination of historical data on a time scale, the corresponding time period data is automatically retrieved from the historical wind, solar and energy storage data to obtain the filtered historical wind, solar and energy storage data.

[0104] Methods for screening similar meteorological samples include:

[0105] Traverse the historical time series of wind, solar, and solar storage data, and calculate its similarity to the meteorological data at the target time. The similarity calculation expression is as follows:

[0106] ;

[0107] in, express real-time weather data and Similarity of meteorological data at different times express Real-time weather data, express Real-time weather data, Indicates the modulus;

[0108] Historical wind, solar and energy storage data with similarity greater than a set threshold are combined into a similar meteorological sample set.

[0109] To avoid information confusion caused by simple time splicing, this embodiment adopts a structured tag organization method, which separates historical data from different sources with semantic tags, such as recent trend segments, periodic reference segments, cross-year comparison segments, and similar meteorological sample segments. This enables the large model to distinguish historical information of different functions when allocating attention, thereby improving inference stability.

[0110] Compared with the traditional fixed window input method, this embodiment achieves adaptive determination of context length through multi-turn dialogue, making the selection of historical information relevant to the problem and specific to the scenario, avoiding invalid token occupation, improving the utilization efficiency of millions of context resources, and significantly enhancing the cross-year generalization ability and adaptability to extreme weather scenarios under zero-sample conditions, thereby constructing a truly problem-driven ultra-long context prediction framework.

[0111] Using filtered historical wind, solar, and energy storage data as input, when meteorological forecast data for the target time is available, the ultra-long context large model uses the meteorological forecast data for the target time as a covariate to achieve joint prediction; otherwise, only meteorological-power joint prediction is performed. The modeling form of the joint prediction is restructured into a conditional generation problem, the expression of which is:

[0112] ;

[0113] in, This represents a large model inference mapping with very long contexts. This indicates a very long context composed of filtered historical data on wind, solar, and energy storage. This represents the covariates composed of meteorological forecast data at the target time. This represents the wind, solar, and energy storage data at the target time. Indicates the prediction step, This indicates the start time point of the prediction.

[0114] The core of this joint generation mechanism lies in the fact that meteorological and power are no longer related as upstream and downstream in a causal chain, but rather as joint random variables in the same probability space, generated conditionally. This allows for the dynamic adjustment of power predictions while generating meteorological forecasts, avoiding the amplification of errors caused by the propagation of meteorological errors.

[0115] Regarding covariate embedding, this embodiment adopts the ICL (In-Context Learning) mechanism, which directly embeds the historical-power mapping relationship as an example into the context, rather than completing the modeling through parameter training. Specifically, multiple "historical weather-power corresponding segments" are constructed in the ultra-long context, for example: "when the wind speed is X1, the irradiance is Y1, and the temperature is Z1, the wind power is A1, the photovoltaic power is B1, and the energy storage output is C1."

[0116] By providing a sufficient number of historical mapping examples, the large model abstracts the mapping relationships within the context, thereby generating the corresponding power output under future weather conditions. If third-party weather forecast data exists, it is directly embedded as a future covariate into the prediction interval input; otherwise, only joint weather-power prediction is performed.

[0117] In the integrated wind-solar-storage scenario, wind power Photovoltaic power With energy storage power There is a strong coupling relationship between the physical and operational levels. Traditional prediction methods typically model these three factors separately, neglecting power conservation constraints, energy storage power balance constraints, and grid-connected power limitations, resulting in physically infeasible predictions. This embodiment proposes a coupling consistency constraint reconstruction mechanism based on the output of a large model, which performs consistency constraint correction on the prediction results after joint prediction generation.

[0118] First, establish the power balance relationship:

[0119] ;

[0120] in, This represents the active power of the wind farm at the target time. This represents the active power of the photovoltaic power station at the target time. This represents the charging and discharging power of the energy storage system at the target time. This represents the grid-connected output power at the target time.

[0121] Secondly, the energy recursion constraints for energy storage are established as follows:

[0122] ;

[0123] in, This represents the charging power at the time preceding the target time. This represents the discharge power at the time preceding the target time. Indicates the time step. Indicates the rated charging power. Indicates the rated discharge power. This represents the stored energy at the target time. This represents the stored energy at the previous time step before the target time.

[0124] The consistency constraint is:

[0125]

[0126] in, This represents the minimum energy storage capacity. Indicates the maximum value of stored energy. This indicates the maximum power of the power grid.

[0127] If the wind, solar, and energy storage data at the target time do not meet the above constraints, then a minimum correction optimization problem is constructed to correct the wind, solar, and energy storage data at the target time. Its expression is:

[0128] ;

[0129] in, This represents the active power of the wind farm at the corrected target time. This represents the active power of the photovoltaic power plant at the corrected target time. This represents the corrected charging and discharging power of the energy storage system at the target time. Indicates the prediction step, This indicates the start time point of the prediction.

[0130] Power dispatching systems require probabilistic predictions rather than single-value predictions; therefore, this embodiment introduces a structured uncertainty estimation mechanism.

[0131] ;

[0132] ;

[0133] ;

[0134] ;

[0135] ;

[0136] in, This represents the predicted mean. Indicates the first Predicted data for secondary wind, solar, and energy storage power. Indicates the number of predictions. Indicates the standard deviation of the forecast. This represents the quantile value of the standard normal distribution. Indicates the lower bound of confidence. Indicates the upper bound of confidence. Indicates the width of the interval.

[0137] Interval width can effectively represent the uncertainty of prediction. The larger the value, the higher the uncertainty of the prediction result; the smaller the value, the more reliable the prediction result.

[0138] Example 2:

[0139] Building upon Example 1, this example also provides a meteorological-wind-solar-storage power joint prediction system based on ultra-long context. The modules in the system operate collaboratively as intelligent agents, specifically including:

[0140] The wind, solar and energy storage historical data acquisition module is configured to acquire historical wind, solar and energy storage data.

[0141] The timescale data combination determination module is configured to determine the combination method of historical data on the timescale based on the historical data of wind, solar and energy storage, using a large model with ultra-long context.

[0142] The wind, solar and energy storage historical data filtering module is configured to filter the wind, solar and energy storage historical data from the historical data based on the combination of historical data on the time scale.

[0143] The wind, solar and energy storage power prediction module is configured to take the filtered historical wind, solar and energy storage data as input and predict the wind, solar and energy storage power data at the target time based on the ultra-long context large model. In this embodiment, the ultra-long context large model adopts a domestic large model, and the model can be called in the form of API call or local deployment.

[0144] The prediction result correction module is configured to perform wind-solar-storage coupling consistency constraints on the wind-solar-storage power data at the target time, and use the constraint results to correct the wind-solar-storage power prediction data.

[0145] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A joint prediction method for meteorological, wind-solar-storage power based on ultra-long context, characterized in that, include: Obtain historical data on wind, solar, and energy storage; Based on a large model with ultra-long context, the combination method of historical data on the time scale is determined according to historical wind, solar and energy storage data. Based on the combination of historical data on a time scale, the filtered historical data of wind, solar and energy storage is obtained by filtering from the historical data of wind, solar and energy storage. Using the filtered historical data of wind, solar and energy storage as input, the predicted power data of wind, solar and energy storage at the target time is obtained by making predictions based on the ultra-long context large model. The wind-solar-storage power prediction data at the target time are subjected to wind-solar-storage coupling consistency constraints, and the constraint results are used to correct the wind-solar-storage power prediction data.

2. The meteorological-wind-solar-storage power joint prediction method based on ultra-long context as described in claim 1, characterized in that, The historical data for wind, solar, and energy storage includes power data, measured meteorological data, and equipment operation data. The power data includes the active power of the wind farm, the active power of the photovoltaic power station, the charging and discharging power of the energy storage system, and the grid-connected output power, expressed as follows: ； in, Represents power data. This indicates the active power of the wind farm. Indicates the active power of a photovoltaic power station. Indicates the charging and discharging power of the energy storage system. Indicates grid-connected output power; The measured meteorological data includes wind speed, wind direction, horizontal irradiance, temperature, humidity, and air pressure, and its expression is: ； in, This represents measured meteorological data. Indicates wind speed. Indicates wind direction. Indicates horizontal irradiance. Indicates temperature. Indicates humidity. Indicates air pressure; The equipment operation data includes unit availability, fault identification, power limitation status, and remaining energy storage.

3. The meteorological-wind-solar-storage power joint prediction method based on ultra-long context as described in claim 1, characterized in that, The method for determining the combination of historical data on a time scale based on the ultra-long context large model and historical wind, solar and energy storage data includes: Based on dynamic time warping enhancement, the historical data of wind, solar and energy storage are time-aligned to obtain a preliminary time series of historical data of wind, solar and energy storage. Based on the ultra-long context large model, outlier detection is performed on the preliminary wind, solar and energy storage historical data time series to obtain the wind, solar and energy storage historical data time series. Using historical time series data of wind, solar and energy storage as input, the task semantic adaptive reasoning is performed based on a large model with ultra-long context to obtain the combination of historical data on the time scale.

4. The meteorological-wind-solar-storage power joint prediction method based on ultra-long context as described in claim 3, characterized in that, The preliminary time series of historical wind, solar, and energy storage data obtained by time alignment of historical wind, solar, and energy storage data based on dynamic time warping enhancement includes: The optimal time mapping path is obtained by minimizing the similarity error between any two data sequences in the historical wind, solar, and energy storage data over time, using this as the objective function. The expression for the objective function is as follows: ； ； ； in, Indicates the first A set of wind, solar and energy storage data awaiting time alignment. Represents the baseline timeline sequence. Represents a local distance metric function. express and Similarity in the time dimension Represents a set of time-mapped paths. Indicates the time mapping path, Indicates the first The first of the wind, solar and energy storage data awaiting time alignment One element, Represents the first in the reference timeline sequence One element, This represents the weight hyperparameter. This indicates a global alignment path consistency regular expression. This indicates the number of wind, solar, and storage data points awaiting time alignment. Indicates the first The wind, solar, and energy storage data awaiting time alignment are in the first... The corresponding values ​​at each reference time point Indicates variance operation; Preliminary time series of historical wind, solar and energy storage data are obtained by resampling the historical data of wind, solar and energy storage using the optimal time mapping path.

5. The meteorological-wind-solar-storage power joint prediction method based on ultra-long context as described in claim 3, characterized in that, The process of obtaining historical wind, solar, and energy storage data time series by performing outlier detection on the preliminary wind, solar, and energy storage historical data time series based on a large, ultra-long context model includes: S1. Using the preliminary historical data time series of wind, solar and energy storage and the meteorological conditions at the current moment as input, the model with ultra-long context determines whether the wind, solar and energy storage data at the current moment is abnormal based on the relationship between meteorological conditions and wind, solar and energy storage power. If abnormal, the data is set to empty; otherwise, it is retained, and the preliminary historical data time series of wind, solar and energy storage after anomaly detection is obtained. S2. Using the preliminary wind, solar and energy storage historical data time series after outlier detection and the meteorological conditions at the next moment as input, repeat S1 until the wind, solar and energy storage historical data time series is obtained.

6. The meteorological-wind-solar-storage power joint prediction method based on ultra-long context as described in claim 3, characterized in that, The expression for the historical data combination method at the aforementioned time scale is: ； in, This indicates how historical data is combined on a given time scale. Indicates the length of the most recent consecutive historical data. Indicates the length of periodic historical data. Indicates the length of historical data spanning multiple years. Indicates the length of similar meteorological sample data.

7. The meteorological-wind-solar-storage power joint prediction method based on ultra-long context as described in claim 6, characterized in that, The screening method for similar meteorological sample data includes: Traverse the historical time series of wind, solar, and solar storage data, and calculate its similarity to the meteorological data at the target time. The similarity calculation expression is as follows: ； in, express real-time weather data and Similarity of meteorological data at different times express Real-time weather data, express Real-time weather data, Indicates the modulus; Historical wind, solar and energy storage data with similarity greater than a set threshold are combined to form similar meteorological sample data.

8. The meteorological-wind-solar-storage power joint prediction method based on ultra-long context as described in claim 1, characterized in that, When meteorological forecast data for the target time is available, the ultra-long context large model uses the meteorological forecast data for the target time as covariates to achieve joint prediction. The modeling form of the joint prediction is reconstructed into a conditional generation problem, and the expression of the conditional generation problem is: ； in, This represents a large model inference mapping with very long contexts. This indicates a very long context composed of filtered historical data on wind, solar, and energy storage. This represents the covariates composed of meteorological forecast data at the target time. This represents the wind, solar, and energy storage power data at the target time. Indicates the prediction step, This indicates the start time point of the prediction.

9. The meteorological-wind-solar-storage power joint prediction method based on ultra-long context as described in claim 1, characterized in that, The expression for the wind-solar-storage coupling consistency constraint is: ； ； ； ； ； in, This represents the active power of the wind farm at the target time. This represents the active power of the photovoltaic power station at the target time. This represents the charging and discharging power of the energy storage system at the target time. This represents the grid-connected output power at the target time. This represents the charging power at the time preceding the target time. This represents the discharge power at the time preceding the target time. Indicates the time step. Indicates the rated charging power. Indicates the rated discharge power. This represents the stored energy at the target time. This represents the stored energy at the previous time step before the target time. This represents the minimum energy storage capacity. Indicates the maximum value of stored energy. Indicates the maximum power of the power grid. This represents the active power of the wind farm at the corrected target time. This represents the active power of the photovoltaic power plant at the corrected target time. This represents the corrected charging and discharging power of the energy storage system at the target time. Indicates the prediction step, This indicates the start time point of the prediction.

10. The meteorological-wind-solar-storage power joint prediction method based on ultra-long context as described in claim 1, characterized in that, It also includes uncertainty assessment of wind, solar, and energy storage forecast data, the calculation expression for which the uncertainty is: ； ； ； ； ； in, This represents the predicted mean. Indicates the first Predicted data for secondary wind, solar, and energy storage power. Indicates the number of predictions. Indicates the standard deviation of the forecast. This represents the quantile value of the standard normal distribution. Indicates the lower bound of confidence. Indicates the upper bound of confidence. Indicates the width of the interval.

11. A meteorological-wind-solar-storage power joint prediction system based on ultra-long context, characterized in that, The method for implementing the joint prediction method of meteorological, wind, solar and storage power based on ultra-long context as described in any one of claims 1 to 10 includes: The wind, solar and energy storage historical data acquisition module is configured to acquire historical wind, solar and energy storage data. The time-scale data combination determination module is configured to determine the combination method of historical data on the time scale based on the wind, solar and energy storage historical data, using a large model with ultra-long context. The wind, solar and energy storage historical data filtering module is configured to filter the wind, solar and energy storage historical data from the historical data based on the combination of historical data on the time scale. The wind, solar and energy storage power prediction module is configured to take the filtered historical wind, solar and energy storage data as input and predict the wind, solar and energy storage power data at the target time based on the ultra-long context large model. The prediction result correction module is configured to perform wind-solar-storage coupling consistency constraints on the wind-solar-storage power data at the target time, and use the constraint results to correct the wind-solar-storage power prediction data.

Images