A self-supervised learning prediction method and system for spatiotemporal meteorological-power series
By employing a self-supervised learning method, a virtual wind farm pre-training framework and a self-supervised calibration gating fusion module were constructed. This solved the problems of scarce labeled data and data heterogeneity in wind power prediction, achieving high-precision and generalizable wind power prediction to meet the needs of different scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI JIAOTONG UNIV
- Filing Date
- 2026-03-12
- Publication Date
- 2026-06-02
AI Technical Summary
Existing wind power prediction technologies suffer from problems such as scarce labeled data, data heterogeneity, fixed feature fusion methods, and insufficient model adaptability when processing spatiotemporal series data of wind farm meteorology and power, making it difficult to achieve a balance between high accuracy and generalization ability.
A self-supervised learning method is adopted, which constructs a virtual power station self-supervised pre-training framework and a self-supervised calibration gating fusion module, and combines spatiotemporal interpolation, spatial grid alignment, adaptive feature fusion and dynamic weight calculation to achieve wind power prediction.
It achieves high-precision prediction under different wind farm and meteorological conditions, reduces dependence on labeled data, improves the model's adaptability and generalization ability, and meets the real-time dispatching needs of the power grid.
Smart Images
Figure CN122133872A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of new energy power generation prediction technology, self-supervised learning technology, and spatiotemporal sequence data analysis. It is a self-supervised learning prediction method and system for wind farm meteorological-power spatiotemporal sequence data, which is applicable to ultra-short-term and medium-short-term wind power prediction scenarios. Background Technology
[0002] As a core component of clean and renewable energy, the large-scale development and utilization of wind energy has become crucial for energy structure transformation. Accurate wind power forecasting is a core technological support for ensuring the safe and stable operation of the power system, optimizing grid dispatch, and improving wind power absorption efficiency. However, wind power output is highly random, volatile, and intermittent, and its variations are highly correlated with meteorological factors such as wind speed, wind direction, and air pressure. Furthermore, meteorological data and power data exhibit significant spatiotemporal heterogeneity and time-frequency inconsistencies, posing a significant challenge to wind power forecasting.
[0003] Existing wind power prediction technologies are mainly divided into three categories:
[0004] Physical methods: Power output is derived based on fluid dynamics, wind turbine aerodynamic characteristics and meteorological numerical forecasting models, but they have stringent requirements for terrain data, high computational complexity, and the prediction accuracy is heavily dependent on the accuracy of meteorological forecasts, making them difficult to meet the needs of refined prediction under complex meteorological conditions. Traditional statistical methods model by mining statistical patterns in historical data, but they can only capture linear relationships and cannot adapt to the complex nonlinear spatiotemporal dependence between meteorology and power. Machine learning methods, including support vector machines and traditional neural networks, can fit nonlinear relationships, but they have two major drawbacks: First, they rely on large-scale, high-quality labeled data, while wind farm labeled data collection is costly and samples are scarce; second, they lack customized designs for the spatiotemporal characteristics of meteorological-power sequences, making it difficult to handle cross-frequency and data feature transfer, and resulting in insufficient generalization ability.
[0005] This invention provides a new approach to solving the problem of scarce labeled data by learning effective features from massive amounts of unlabeled data through self-supervised learning and designing unsupervised pre-training tasks. However, existing applications of self-supervised learning in wind power forecasting have significant shortcomings: First, the pre-training task design is too simplistic and fails to fully explore key information such as the spatiotemporal correlation, frequency alignment, and migration of meteorological-power data; second, the feature fusion method is rigid and cannot dynamically balance the weights of spatiotemporal and temporal features, resulting in insufficient model adaptability to different scenarios; third, it does not specifically address data heterogeneity issues such as inconsistent data frequencies and spatial locations. Therefore, there is an urgent need for a self-supervised learning forecasting method that can deeply adapt to the spatiotemporal characteristics of meteorological-power sequences, fully explore the inherent patterns in the data, and balance prediction accuracy and generalization ability. Summary of the Invention
[0006] The purpose of this invention is to provide a self-supervised learning prediction method and system for spatiotemporal meteorological-power sequences. By constructing a virtual station self-supervised pre-training framework and a self-supervised calibration gating fusion module, wind power prediction can be achieved.
[0007] To achieve the above-mentioned objectives, the specific steps of the technical solution adopted by this invention include: S1. Data preprocessing: Missing values are filled using spatiotemporal interpolation. This method calculates a weighted average by weighting the data around the missing data based on distance weights. The distance weight formula is as follows:
[0008] in, The distance between the kth valid data point and the missing data point; Outliers are removed using the 3σ criterion when the data meets the following conditions: When a value is identified as an outlier, μ is the data mean and σ is the standard deviation. Normalization uses Min-Max standardization, and the formula is:
[0009] in The minimum value of the data. The maximum value of the data; Spatial grid alignment unifies wind turbine coordinates and meteorological station coordinates into the same spatial grid system. The grid index calculation formula is as follows:
[0010] in , For the target coordinates, , The starting coordinates of the grid. , This refers to the grid resolution.
[0011] S2, the dual-branch feature extraction network includes a main branch and an auxiliary branch: The main branch is the adaptive multi-scale spatiotemporal fusion module, and its attention calculation follows the formula:
[0012] Where Q, K, and V are the query, key, and value matrices, respectively. As the key dimension, Encoding of spatiotemporal relative positions; The auxiliary branch is a bidirectional GRU network, and its hidden state update formula is:
[0013] in To update the door, To reset the door, , This is the weight matrix. , For bias.
[0014] S3. The gating weights of the self-supervised calibration gating fusion module are calculated as follows:
[0015] in, The spatiotemporal fusion features output by the main branch, To aid the timing dependency features of branch outputs, , For gating parameters, For the Sigmoid function, For self-supervised calibration weights.
[0016] The fusion features are calculated as follows:
[0017] in, For gating weights, As a feature of spatiotemporal fusion, It is a time-dependent feature.
[0018] S4. The loss function for the fine-tuning stage is the MSE loss, and the formula is:
[0019] To predict power, This represents the actual power. The calculation of spatiotemporal relative position coding includes temporal coding and spatial coding, and the calculation formulas are as follows: Time encoding:
[0020] in , The start time, For time step Spatial encoding:
[0021] in , Relative coordinates For batch indexing, For coordinate index.
[0022] S5. The wind power forecast results include short-to-medium term (1-5 days) and ultra-short term (15 minutes-1 hour) wind power forecasts, with a time frequency of 15 minutes. The inverse normalization formula for the forecast results is as follows:
[0023] in For normalized predicted values, , These represent the maximum and minimum power values, respectively.
[0024] This application also provides a computer system, including: A processor is used to execute instructions stored in a storage medium to implement a self-supervised learning prediction method for spatiotemporal sequence meteorological-power.
[0025] The memory is used to store the data and intermediate results generated during the calculation process of a self-supervised learning prediction method for spatiotemporal series meteorological-power. The input / output interface is used to receive specified relevant information and output the transformed prediction results.
[0026] A self-supervised learning prediction method and system for spatiotemporal series meteorological-power, characterized in that the device is configured as follows: A self-supervised learning prediction method and system for spatiotemporal meteorological-power sequences has the following advantages: The virtual power station pre-training self-supervised pre-training framework realizes in-depth feature mining: through the virtual power station pre-training task, time series frequency alignment + future power prediction mines general features, respectively capturing the local correlation features, cross-frequency adaptation features and related general features of meteorological-power data, solving the problem that a single task is difficult to fully characterize the data pattern. Adaptive feature fusion enhances model adaptability: The self-supervised calibration gated fusion module achieves intelligent balance between spatiotemporal features and temporal features by dynamically calculating weights, adapting to the differences in feature distribution under different wind farms and meteorological conditions; The data preprocessing scheme addresses the heterogeneity problem: methods such as spatiotemporal interpolation and spatial grid alignment effectively handle data missingness and spatial location dispersion, while the 3σ criterion and Min-Max standardization ensure data quality and provide reliable input for model training. Reduced dependence on labeled data and strong generalization ability: By pre-training on unlabeled data to learn general features, high-precision prediction can be achieved with only a small amount of labeled data for fine-tuning, which significantly reduces data collection costs. At the same time, the transfer task improves the model's adaptability to different wind farms. Predictions cover multiple scenarios: Supports ultra-short-term and medium-short-term dual-scenario predictions, with 15-minute high-frequency output to meet the real-time dispatching needs of the power grid, providing full-cycle technical support for large-scale wind power grid connection. Attached Figure Description
[0027] Figure 1 This is the overall flowchart of the method; Figure 2 This is a schematic diagram of the core model structure. Detailed Implementation
[0028] A wind farm covers a total area of 55 km² and is equipped with 45 2.0MW direct-drive wind turbines, each with a rated power of 2.0MW. The turbines are evenly distributed in a grid with a spacing of about 1.3 km. Three meteorological stations are evenly distributed in the east, central and west areas of the wind farm to monitor four core meteorological indicators: wind speed, wind direction, air pressure and temperature. These stations are used to support wind power forecasting and power system dispatch decisions.
[0029] S1: Data Preprocessing 1. Missing value imputation (spatiotemporal interpolation), calculated using a given distance weight formula:
[0030] Where dk is the spatial Euclidean distance between the missing point and the k-th valid data point, in km.
[0031] For a wind turbine in the central part of the wind farm, with coordinates N38.5°, E112.3°, power data for October 12, 2020 at 09:45 is missing. The coordinates, distances, and power values of the three surrounding valid wind turbines are as follows:
[0032] Calculation process: Distance weight: , ,
[0033] Filler value: 1.7×0.31+1.8×0.36+1.6×0.23≈1.72MW 2. Outlier removal, The criteria calculate the mean of each indicator data. and standard deviation Eliminate those that meet the requirements Data: Taking the historical power data of wind turbine 08 as an example: Statistics of 2 years of data are obtained =1.2MW, =0.4MW, 3 =1.2MW, therefore excluded or The data shows a single unit rated power of 2.0MW; any abnormal data exceeding the rated value must be removed. A certain record with a power data of 2.5MW satisfies |2.5 1.2 |=1.3>1.2 is considered an outlier and is replaced with the spatiotemporal interpolation result of 1.9MW from the two adjacent valid data points.
[0034] 3. Normalization: Min-Max standardization reduces all data to the [0,1] interval according to the patented formula.
[0035] Power data: =0MW, =2.0MW, the normalized result of a certain power value of 1.5MW is =0.75 Wind speed data: =0m / s, =14m / s, the normalized result of a certain wind speed value of 7.5m / s is (7.5 0) / (14 0)≈0.54 All meteorological indicators and power data are normalized according to this formula to ensure consistent data scale.
[0036] 4. Spatial grid alignment: Unify the coordinates of wind turbines and meteorological stations to the same spatial grid system. Parameter settings: Grid starting coordinates: =38.2°N, =112.0°E Grid resolution: =0.1°, =0.1°, corresponding to a grid of approximately 11km × 11km on the ground. Grid index formula:
[0037] Calculate the coordinates of a weather station (N38.4°, E112.2°). , The grid index is (2,2), and all wind turbines and meteorological stations are mapped to this grid system to complete spatial alignment.
[0038] S2: Construct a self-supervised pre-training framework for virtual wind turbines: The virtual wind turbine data includes historical power data generated by physical simulation at a time frequency of 15 minutes, historical meteorological data at a time frequency of 1 hour, and future meteorological forecast data at a time frequency of 1 hour. The physical simulation power data is generated based on the wind turbine aerodynamic model and meteorological numerical simulation results.
[0039] 1. Core model parameter settings: Adaptive multi-scale spatiotemporal fusion module: basic window size Fan density Units / km², wind speed standard deviation Learnable parameters , Window size calculation:
[0040] Dual-branch feature extraction network: The main branch uses adaptive multi-scale spatiotemporal fusion: key dimension Spatiotemporal relative position encoding Calculate the temporal encoding and spatial encoding using the formula. Auxiliary branch, bidirectional GRU: hidden layer dimension 64, number of layers 2, update gate. and reset door The weights are initialized as a random normal distribution.
[0041] 2. Virtual station pre-training implementation: Training parameters: 100 pre-training rounds, batch size=32, optimizer Adam, initial learning rate 1e-4, learning rate decays by 50% every 20 rounds. Time series frequency alignment prediction: meteorological data time step. =17520, power data time step =70080, the MAE loss function minimizes the historical data recovery error, at the 100th round. =0.018.
[0042] S3: Constructing a self-supervised calibration gating fusion module 1. Gating weight calculation
[0043] in, Main branch spatiotemporal fusion features, dimensions , To assist in the temporal dependency features of the branch, the dimension , Dimension and Dimension For learnable parameters, , For the Sigmoid function. At a certain moment in a certain batch. eigenvectors and The calculated value is 0.32. Add... The gating weight is 0.42. .
[0044] 2. Calculation of fusion features, taking g=0.60 as an example above, The contributions of the main branch of dynamic balance spatiotemporal features and the auxiliary branch of temporal features.
[0045] S4: Model fine-tuning. Fine-tuning parameters were set using 35,040 power data points from labeled data collected between January 1, 2021 and December 31, 2021. The fine-tuning process lasted 30 rounds, with the learning rate reduced to 1e-5. 80% of the underlying parameters of the pre-trained model were frozen, and only the parameters of the top-level network and gating modules were fine-tuned. The loss function was the MSE validation set. =0.015, the model converges and is stable.
[0046] S5: Wind Power Prediction and Inverse Normalization 1. Input Data Preparation Historical data for the forecast period: historical power and meteorological data from December 27, 2021 to December 31, 2021, preprocessed by S1.
[0047] Future weather forecast data: 1-hour meteorological data from January 1, 2022 to January 2022, upsampled to 15-minute level, and then normalized by S1.
[0048] 2. Prediction Results Output The normalized predicted values are output by the final prediction model. Then, after inverse normalization, the actual power value is obtained, where, , Normalized predicted value at a certain moment Then the actual power prediction value .
[0049] Evaluation index definition Standard evaluation indicators from the field of wind power forecasting are adopted: Mean Absolute Error (MAE):
[0050] Root Mean Square Error (RMSE):
[0051] Mean Absolute Percentage Error (MAPE): (removed) (Data) Ultra-short-term forecast results
[0052] Short- and medium-term forecast results
[0053] Method Comparison and Validation The method of this invention is compared with traditional prediction methods, and the results are as follows:
[0054] in conclusion Excellent prediction accuracy: The method of this invention has a MAPE of only 1.43% in ultra-short-term prediction and 3.16% in medium-short-term prediction, both of which are far lower than traditional methods, meeting the high accuracy requirements of power system dispatch for wind power prediction (the industry standard MAPE < 8%).
[0055] Outstanding generalization ability: Through the feature transfer task, the model achieved a MAPE of <3.8% in unlabeled adjacent wind farm tests, significantly reducing the dependence on labeled data and adapting to different wind farm scenarios.
[0056] This embodiment also provides a computer system, including: A processor is used to execute instructions stored in a storage medium to implement a self-supervised learning prediction method for spatiotemporal sequence meteorological-power as described in the above embodiments; The memory is used to store the text of a self-supervised learning prediction method for spatiotemporal series meteorological-power and the intermediate results generated during the calculation process; The input / output interface is used to receive specified information and output wind power prediction results.
[0057] The computer system provided in this embodiment corresponds to the method embodiment described above, and has corresponding functional modules and beneficial effects, which will not be repeated here.
[0058] This embodiment also provides a self-supervised learning prediction device for spatiotemporal sequence meteorological-power. The device is configured to: preprocess data, construct a large-scale self-supervised pre-training framework for meteorological-power mapping based on physical simulation, construct a self-supervised calibration gating fusion module, input the preprocessed labeled data into the pre-trained core model and the complete model composed of the self-supervised calibration gating fusion module for fine-tuning, input the future meteorological forecast data of the period to be predicted, the corresponding historical power data and historical meteorological data into the final prediction model, and output the wind power prediction result.
[0059] The self-supervised learning prediction device for spatiotemporal sequence meteorological-power provided in this embodiment corresponds to the above method embodiment and has corresponding functional modules and beneficial effects, which will not be elaborated here.
Claims
1. A self-supervised learning prediction method for spatiotemporal series meteorological-power, characterized in that, Includes the following steps: S1: Data Preprocessing: Acquire historical power data, historical meteorological data, and future meteorological forecast data of the wind farm, and perform preprocessing, normalization, and spatial grid alignment on the data; S2: Virtual wind power station pre-training: Design a time-series frequency alignment prediction task, input past meteorological data, future meteorological forecast data and past physical simulation power data into the core model for pre-training. The core model includes a cross-frequency time-series alignment network, an adaptive multi-scale spatiotemporal fusion module, and a dual-branch feature extraction network. The loss function for pre-training is the mean square error (MSE) of the future wind power prediction results. S3: Construct a self-supervised calibration gated fusion module: dynamically fuse the spatiotemporal features and temporal features output by the dual-branch feature extraction network to obtain fused features; S4: Site fine-tuning: The pre-processed labeled data is input into the pre-trained core model and the complete model composed of the self-supervised calibration gating fusion module for fine-tuning, optimizing the model parameters, and obtaining the final prediction model; S5: Wind Power Prediction: Input the future weather forecast data for the period to be predicted, the corresponding historical power data, and historical weather data into the final prediction model, and output the wind power prediction results.
2. The method according to claim 1, characterized in that, Step S1 includes: Data preprocessing includes filling missing values using spatiotemporal interpolation. This spatiotemporal interpolation method calculates a weighted average of the missing data by calculating the distance weights of the valid data surrounding the missing data. The distance weight formula is as follows: in, The distance between the k-th valid data point and the missing data point. The weights represent the results of the calculation; Outliers are removed using the 3σ criterion when the data meets the following conditions: When a value is identified as an outlier, μ is the data mean and σ is the standard deviation. Normalization uses Min-Max standardization, and the formula is: in, The minimum value of the data. The maximum value of the data; Spatial grid alignment unifies wind turbine coordinates and meteorological station coordinates into the same spatial grid system. The grid index calculation formula is as follows: in, , For the target coordinates, , The starting coordinates of the grid. , This refers to the grid resolution.
3. The method according to claim 1, characterized in that, In step S2, the dual-branch feature extraction network includes a main branch and an auxiliary branch: The main branch is the adaptive multi-scale spatiotemporal fusion module, and its attention calculation follows the formula: Where Q, K, and V are query, key, and value matrices, As the key dimension, This is a spatiotemporal relative position encoding, where T is the matrix transpose symbol; The auxiliary branch is a bidirectional GRU network, and its hidden state update formula is: in, To update the door, To reset the door, , This is the weight matrix. , For bias, This indicates the hidden state at the previous moment.
4. The method according to claim 1, characterized in that, In step S3, the gating weights of the self-supervised calibration gating fusion module are calculated as follows: in, The spatiotemporal fusion features output by the main branch, To aid the timing dependency features of branch outputs, , For gating parameters, For the Sigmoid function, For self-supervised calibration weights.
5. The method according to claim 1, characterized in that, In step S3, the fusion features are calculated as follows: in, For gating weights, As a feature of spatiotemporal fusion, It is a time-dependent feature.
6. The method according to claim 1, characterized in that, In step S4, the loss function for the fine-tuning stage is the MSE loss, and the formula is: The MSE loss formula in the pre-training phase is the same as that in the fine-tuning phase. To predict power, This represents the actual power.
7. The method according to claim 1, characterized in that, Spatiotemporal relative position coding includes temporal coding and spatial coding, and the calculation formulas are as follows: Time encoding: in , The start time, For time step Spatial encoding: in , Relative coordinates For batch indexing, For coordinate index.
8. The method according to claim 1, characterized in that, In step S5, meteorological data is used only as input and does not require reconstruction or restoration. A two-stage training mode of pre-training + site fine-tuning is adopted, using the future wind power prediction MSE as the sole training loss function. The wind power prediction results include short-to-medium term (1-5 days) and ultra-short term (15 minutes-1 hour) wind power prediction values, with a time frequency of 15 minutes. The inverse normalization formula for the prediction results is: in For normalized predicted values, , These represent the maximum and minimum power values, respectively.
9. A computer system, characterized in that, include: A processor for executing instructions stored in a storage medium to implement a self-supervised learning prediction method for spatiotemporal sequence meteorological-power according to any one of claims 1-8; The memory is used to store the text of a self-supervised learning prediction method for spatiotemporal series meteorological-power and the intermediate results generated during the calculation process; The input / output interface is used to receive specified information and output wind power prediction results.
10. A self-supervised learning prediction device for spatiotemporal series meteorological-power, characterized in that, The device is configured to: preprocess data, construct a large-scale self-supervised pre-training framework for meteorological-power mapping based on physical simulation, construct a self-supervised calibration gating fusion module, input the preprocessed labeled data into the pre-trained core model and the complete model composed of the self-supervised calibration gating fusion module for fine-tuning, input the future meteorological forecast data of the period to be predicted, the corresponding historical power data and historical meteorological data into the final prediction model, and output the wind power prediction results.