A method and system for predicting power generation of an offshore wind farm based on big data

Abstract

The present application relates to the technical field of big data processing, and particularly relates to a method and system for offshore wind farm power generation prediction based on big data; the present application adopts an ultra-short-term power generation prediction model based on data-physical hybrid driving to predict offshore wind farm power generation, wherein the ultra-short-term power generation prediction model based on data-physical hybrid driving comprises a deep learning model, a physical model module, a reinforcement learning module and a dynamic switching mechanism; the dynamic switching mechanism is used to dynamically select the deep learning model module or the physical model module according to a wind speed fluctuation threshold and feedback of the reinforcement learning module; the model selection logic provided by the present application can significantly improve prediction accuracy; at the same time, the dynamic switching mechanism improves the robustness and generalization ability of the prediction model in combination with the wind speed fluctuation threshold and feedback of the reinforcement learning.

Description

Technical Field

[0001] This invention relates to the field of big data processing technology, and in particular to a method and system for predicting the power generation of offshore wind farms based on big data. Background Technology

[0002] In recent years, with the increasing global demand for clean energy, offshore wind power, as an efficient, clean, and sustainable energy form, has developed rapidly. Offshore wind farms are typically located in sea areas far from land, offering advantages such as abundant wind energy resources and small footprint. However, the complex operating environment of offshore wind farms makes it difficult to predict their power generation.

[0003] For forecasting the power generation of offshore wind farms, existing technologies employ a hybrid model consisting of a physical model and a data-driven model. However, these solutions lack a clear switching mechanism, resulting in low accuracy in power generation forecasting. Furthermore, the poor timestamp alignment of the large dataset of offshore wind farm data used for power generation forecasting in existing technologies also affects the accuracy of the forecasts. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method and system for predicting the power generation of offshore wind farms based on big data, which solves the problems existing in the prior art.

[0005] This invention provides a method for predicting the power generation of offshore wind farms based on big data, comprising the following steps:

[0006] S1: Obtain big data for predicting the power generation of offshore wind farms;

[0007] S2: Perform data preprocessing operations on the big data to obtain preprocessed big data;

[0008] S3: Establish a prediction model for the power generation of offshore wind farms;

[0009] The offshore wind farm power generation prediction model is an ultra-short-term power generation prediction model based on a data-physical hybrid driving system. The ultra-short-term power generation prediction model includes:

[0010] A deep learning model is used to extract time-series features from the preprocessed big data;

[0011] The physical model module is used to predict power based on the physical characteristics of the wind farm;

[0012] A reinforcement learning module is used to dynamically adjust the switching strategy of the ultra-short-term power generation prediction model;

[0013] A dynamic switching mechanism is used to dynamically select either the deep learning model module or the physical model module based on the wind speed fluctuation threshold and the feedback from the reinforcement learning module.

[0014] S4: Input the real-time wind data in the preprocessed big data into the offshore wind farm power generation prediction model to obtain the power generation prediction result.

[0015] Preferably, in step S3, the input to the deep learning model module is the preprocessed big data; the network architecture of the deep learning model module includes: a multi-layer convolutional neural network, a long short-term memory network, a Transformer encoder, and an output layer.

[0016] Preferably, the multi-layer convolutional neural network is used to extract spatial features from the preprocessed big data, and its specific structure includes:

[0017] Convolutional layers: use multiple convolutional kernels of different sizes to extract multi-scale features;

[0018] Pooling layers: Max pooling or average pooling is used after each convolutional layer to reduce feature dimensionality;

[0019] Feature fusion layer: fuses the feature maps extracted by the convolutional layer.

[0020] Preferably, the Long Short-Term Memory (LSTM) network is used to capture long-term dependencies in time series from the preprocessed big data, and its specific structure includes:

[0021] Input layer: The feature maps extracted by the multi-layer convolutional neural network are used as input;

[0022] Multi-layer LSTM: A multi-layer LSTM structure is used, where each LSTM layer performs deeper feature extraction on the input time series; each LSTM unit of the multi-layer LSTM contains an input gate, a forget gate, and an output gate to control the inflow, retention, and outflow of information;

[0023] Dropout layer: Add a Dropout layer between LSTM layers to prevent overfitting.

[0024] Preferably, the Transformer encoder is used to further improve the deep learning model module's ability to model time series data, and its specific structure includes:

[0025] Input embedding layer: Embeds the output feature map of each LSTM layer into the input space of the Transformer encoder;

[0026] Multi-head self-attention mechanism: enables the Transformer encoder to capture features from different subspaces;

[0027] Feedforward Neural Network: Each Transformer encoder contains a feedforward neural network for further processing of features;

[0028] Residual connectivity and normalization: used to improve model stability and training efficiency.

[0029] Preferably, in step S3, the physical model module is used to predict power generation using the physical characteristics of the wind farm; the physical model module includes a wind speed-power curve fitting layer and a dynamic system model; wherein, the wind speed-power curve fitting layer is a physical model describing the relationship between wind speed and power generation; the dynamic system model is used to describe the dynamic relationship between wind speed changes and power generation output.

[0030] Preferably, the implementation steps of the dynamic switching mechanism are as follows:

[0031] To calculate wind speed fluctuations, at each time step t, calculate the difference between the wind speed at the current time step and the wind speed at the previous time step.

[0032] Set the wind speed fluctuation threshold;

[0033] If the wind speed fluctuation exceeds the wind speed fluctuation threshold, the physical model module is selected for power generation prediction; if the wind speed fluctuation is less than or equal to the wind speed fluctuation threshold, the deep learning model module or the physical model module is selected based on the feedback from the reinforcement learning module.

[0034] Preferably, in step S2, data preprocessing includes timestamp alignment; the timestamp alignment process is as follows:

[0035] Sa: Select the reference time resolution based on the acquisition frequency of different types of data;

[0036] Sb: Multi-level resampling is used to adjust large datasets from different acquisition frequencies to the reference time resolution.

[0037] Preferably, step Sb specifically comprises:

[0038] The large datasets at different frequencies are resampled at the first stage to achieve initial alignment.

[0039] A second-level resampling is performed on the large datasets of different frequencies that have undergone initial alignment;

[0040] The second-level resampling includes: performing multidimensional interpolation on sensor data that has been initially aligned and whose acquisition frequency is lower than the reference time resolution; and performing dynamic weight aggregation on sensor data that has been initially aligned and whose acquisition frequency is higher than the reference time resolution.

[0041] On the other hand, the present invention also provides a big data-based offshore wind farm power generation prediction system, which adopts the above-mentioned big data-based offshore wind farm power generation prediction method. The offshore wind farm power generation prediction system includes:

[0042] The data acquisition module is used to acquire big data for predicting the power generation of offshore wind farms;

[0043] The preprocessing module is used to perform data preprocessing operations on the big data;

[0044] The model building module is used to build a prediction model for the power generation of offshore wind farms and obtain preprocessed big data.

[0045] The power prediction module is used to input the real-time wind data in the preprocessed big data into the offshore wind farm power prediction model to obtain the power prediction result.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] This invention employs a data-physics hybrid-driven ultra-short-term power generation prediction model for offshore wind farm power generation prediction. The data-physics hybrid-driven ultra-short-term power generation prediction model includes a deep learning model module, a physical model module, a reinforcement learning module, and a dynamic switching mechanism. The dynamic switching mechanism dynamically selects either the deep learning model module or the physical model module based on wind speed fluctuation thresholds and feedback from the reinforcement learning module. The model selection logic provided by this invention significantly improves the prediction accuracy of offshore wind farm power generation and reduces the average normalized root mean square error (RMSE), verifying the effectiveness and practicality of the reinforcement learning module. The model selection logic provided by this invention can dynamically adjust the prediction model according to real-time wind speed fluctuations and prediction errors, adapting to different operating conditions. Simultaneously, by combining wind speed fluctuation thresholds, reinforcement learning feedback, and the dynamic switching mechanism, the robustness and generalization ability of the prediction model are improved.

[0048] Meanwhile, this invention employs a two-stage resampling approach for timestamp alignment of big data. In the second-stage resampling, two key steps—multidimensional interpolation and dynamic weight aggregation—further optimize timestamp alignment. The multidimensional interpolation method considers both time and data feature dimensions simultaneously, generating more accurate interpolation results. The dynamic weight aggregation method dynamically adjusts the aggregation weights based on the actual data collection frequency and timestamp deviation, generating more accurate aggregation results. The combination of these two steps effectively handles dynamic changes in collection frequency and timestamp deviation, further improving data temporal consistency. Attached Figure Description

[0049] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0050] Figure 1 This is a flowchart of a method for predicting the power generation of an offshore wind farm based on big data, provided in an embodiment of the present invention.

[0051] Figure 2 This is a flowchart of timestamp alignment provided in an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0053] Example 1

[0054] like Figure 1 As shown, a method for predicting the power generation of offshore wind farms based on big data includes the following steps:

[0055] S1: Obtain big data for predicting the power generation of offshore wind farms;

[0056] The big data includes current wind speed data, historical wind speed data, and historical power generation data.

[0057] S2: Perform data preprocessing operations on the big data to obtain preprocessed big data;

[0058] Data preprocessing directly impacts the accuracy and reliability of subsequent data analysis and early warning models. In this step, data preprocessing includes data cleaning, timestamp alignment, data noise reduction, data normalization, and feature extraction.

[0059] Data cleaning is used to remove outlier and duplicate data, ensuring data quality and reliability. Specifically, the large dataset in step S1 is checked by comparing timestamps and sensor measurements to identify identical duplicate data records. By setting reasonable threshold ranges, data points outside the normal range are removed, thus preprocessing outlier data. For example, the normal range for wind speed sensors is 0-30 m / s, and the normal range for vibration frequency is 0-10 Hz. Data points outside these ranges are considered outlier data.

[0060] Timestamp alignment adjusts inconsistent data of different timestamp types to a unified time resolution, while ensuring data integrity and accuracy.

[0061] In this embodiment, as Figure 2 As shown, the timestamp alignment process includes:

[0062] Sa: Select the reference time resolution based on the acquisition frequency of different types of data;

[0063] The least common multiple (LCM) of the acquisition frequencies of all data types is selected as the baseline time resolution. For data types with lower acquisition frequencies, the time resolution is appropriately reduced, while for data types with higher acquisition frequencies, a higher time resolution is maintained in order to retain more details.

[0064] Sb: Multi-level resampling is used to adjust large datasets from different acquisition frequencies to a reference time resolution. The specific process includes:

[0065] First-level resampling is performed on big data of different frequencies to achieve initial alignment;

[0066] The first-level resampling mentioned above includes first upsampling and first downsampling; the first upsampling is to upsample sensor data with a collection frequency lower than the reference time resolution and use linear interpolation to generate data for missing time points; the first downsampling is to downsample sensor data with a collection frequency higher than the reference time resolution and use an average aggregation method to merge multiple data points into one data point.

[0067] In reality, the varying hardware precision of different sensors can lead to discrepancies between the actual and nominal values ​​of the sampling frequency. For example, some sensors may experience unstable sampling frequencies due to hardware aging, malfunctions, or environmental interference during operation; or the internal clock mechanism of the sensor may drift, causing the sampling frequency to gradually deviate from the nominal value. All of these situations make it difficult for the initial alignment step to completely unify all the large datasets to the reference time resolution. Based on these findings, this embodiment also includes a second-level resampling of the initially aligned large datasets at different frequencies to further improve the accuracy of the time resolution alignment.

[0068] The second-level resampling includes: multidimensional interpolation of sensor data that has been initially aligned and whose acquisition frequency is lower than the reference time resolution; and dynamic weight aggregation of sensor data that has been initially aligned and whose acquisition frequency is higher than the reference time resolution.

[0069] Among them, multidimensional interpolation considers both time and data feature dimensions, which improves the accuracy of interpolation. Specifically, in the big data of different frequencies after preliminary alignment, the time points to be interpolated are selected, and the semivariance function between the time points to be interpolated is calculated. This semivariance function describes the spatial correlation between data points. Based on the semivariance function, a kriging model is constructed. This kriging model generates the value of missing data points by minimizing the variance of the prediction error.

[0070] The dynamic weight aggregation specifically involves: in the initially aligned big data of different frequencies, calculating the time deviation of each data point, where the time deviation refers to the difference between the actual timestamp and the ideal timestamp of the data point. Then, based on the time deviation, dynamically determining the aggregation weight; where the ideal timestamp is a standard time series generated according to the reference time resolution.

[0071] Furthermore, the size of the aggregation window is dynamically adjusted based on the statistical characteristics of the time deviation; specifically:

[0072] Statistical analysis is performed on the time deviation of all data points to calculate the mean and standard deviation. The aggregate weight wi for each data point is dynamically determined based on the mean and standard deviation. The formula for calculating the aggregate weight wi is:

[0073] ;

[0074] In the formula, α is an adjustment parameter used to control the dynamic range of the weight wi. When α is large, the change in weight is more significant; when α is small, the change in weight is more gradual. μΔt is the mean of the time deviation, σΔt is the standard deviation of the time deviation, and Δti is the time deviation of the i-th data point.

[0075] The data points are weighted and aggregated according to the aggregation weight of each data point, thereby achieving dynamic weight aggregation.

[0076] In the second-level resampling, two key steps—multidimensional interpolation and dynamic weight aggregation—further optimize the timestamp alignment of the data. The multidimensional interpolation method considers both time and data feature dimensions, generating more accurate interpolation results. Meanwhile, the dynamic aggregation method dynamically adjusts the aggregation weights based on the actual data collection frequency and timestamp deviation, producing more accurate aggregation results. The combination of these two steps effectively handles dynamic changes in collection frequency and timestamp deviation, further improving the temporal consistency of the data.

[0077] S3: Establish a prediction model for the power generation of offshore wind farms;

[0078] In this step, the offshore wind farm power generation prediction model is an ultra-short-term power generation prediction model based on data-physics hybrid driving. The ultra-short-term power generation prediction model architecture combines deep learning, reinforcement learning and dynamic switching mechanisms, which can more accurately capture the spatiotemporal characteristics of power generation and dynamically adjust the early warning strategy according to real-time wind speed fluctuations.

[0079] Specifically, the aforementioned ultra-short-term power generation prediction model includes:

[0080] The deep learning model module is used to extract time-series features from preprocessed big data.

[0081] The physical model module is used for power prediction based on the physical characteristics of wind farms;

[0082] The reinforcement learning module is used to dynamically adjust the switching strategy of the ultra-short-term power generation prediction model.

[0083] A dynamic switching mechanism is used to dynamically select either the deep learning model module or the physical model module based on the wind speed fluctuation threshold and the feedback from the reinforcement learning module.

[0084] Furthermore, the input to the aforementioned deep learning model module is preprocessed big data (preprocessed big data); the network architecture of the deep learning model module includes: multi-layer convolutional neural network, long short-term memory network, Transformer encoder, and output layer.

[0085] Among them, multi-layer convolutional neural networks are used to extract spatial features from preprocessed big data, and their specific structure includes:

[0086] Convolutional layers: Multiple convolutional kernels of different sizes (such as 3×3, 5×5, 7×7) are used to extract multi-scale features. Each convolutional layer is followed by a ReLU activation function.

[0087] Pooling layers: Max pooling or average pooling is used after each convolutional layer to reduce feature dimensionality, reduce computation, and preserve important features.

[0088] Feature fusion layer: The feature maps extracted by the convolutional layer are fused using methods such as weighted summation, concatenation, or attention mechanisms, so that the multi-layer convolutional neural network can utilize feature information at different scales at the same time.

[0089] Long Short-Term Memory (LSTM) networks are used to capture long-term dependencies in time series data during preprocessing. Their specific structure includes:

[0090] Input layer: The feature maps extracted by the multi-layer convolutional neural network are used as input;

[0091] Multi-layer LSTM: Using a multi-layer LSTM structure, each LSTM layer performs deeper feature extraction on the input time series. Each LSTM unit of this multi-layer LSTM contains an input gate, a forget gate, and an output gate, which are used to control the inflow, retention, and outflow of information.

[0092] Dropout layer: Add a Dropout layer between LSTM layers to prevent overfitting.

[0093] The Transformer encoder further enhances the ability of deep learning models to model time series data, especially by capturing global dependencies through a self-attention mechanism. Its specific structure includes:

[0094] Input embedding layer: Embeds the output feature map of each LSTM layer into the input space of the Transformer encoder;

[0095] Multi-head self-attention mechanism: The Transformer encoder is able to capture features from different subspaces through the multi-head self-attention mechanism;

[0096] Feedforward Neural Network: Each Transformer encoder contains a feedforward neural network for further processing of features;

[0097] Residual connections and normalization: Residual connections and layer normalization are used after each sub-layer to improve model stability and training efficiency.

[0098] The output layer mentioned above includes a fully connected layer and an output module. The fully connected layer is used to integrate the output feature vectors of the Transformer encoder, while the output module is used to generate the final power generation prediction value through a linear layer or a Softmax layer.

[0099] Specifically, the physical model module mentioned above is used to predict power generation by utilizing the physical characteristics of wind farms, especially when wind speed fluctuates greatly, which can effectively make up for the shortcomings of data-driven models.

[0100] Furthermore, the physical model module includes a wind speed-power curve fitting layer and a dynamic system model; wherein, the wind speed-power curve fitting layer is a physical model describing the relationship between wind speed and power generation; the steps for establishing the wind speed-power curve of the wind speed-power curve fitting layer are as follows: collect historical data of the wind farm, including wind speed and corresponding power generation; and use polynomial regression to fit the relationship between wind speed and power generation.

[0101] A dynamic system model is used to describe the dynamic relationship between wind speed changes and power generation output. This model considers the physical characteristics and dynamic response of the wind farm, enabling more accurate prediction of power generation. The specific steps for predicting power generation include:

[0102] Establish a state-space model between wind speed and power to describe the dynamic behavior of offshore wind farm systems;

[0103] Common methods for estimating model parameters using historical data include least squares estimation and maximum likelihood estimation; the model parameters are polynomial coefficients.

[0104] Dynamic prediction: Based on the current state and model parameters, predict the power generation at the next moment; where the current state includes wind speed, current output power, etc.

[0105] Reinforcement learning is a machine learning method that learns optimal behavioral policies through the interaction between an agent and its environment. The agent selects actions based on its current state, and the environment provides rewards based on those actions. The agent learns the optimal policy by maximizing the accumulated rewards. In a data-physics hybrid-driven ultra-short-term power generation prediction model, the reinforcement learning module is used to dynamically adjust the model's switching strategy to select the optimal prediction model (data-driven model or physical model) based on real-time wind speed fluctuations and prediction errors.

[0106] The reinforcement learning module includes a state space, an action space, a reward function, a policy network, an experience replay, and a dynamic switching mechanism.

[0107] Furthermore, the state space is used to describe the agent's observation information at each time step; in power generation prediction, the state space includes:

[0108] Current wind speed fluctuation: the difference between the wind speed at the current moment and the wind speed at the previous moment; Historical prediction error: the prediction error over the past N time steps (the difference between the actual power and the predicted power).

[0109] Current power output: The power output at the current moment;

[0110] Historical prediction error: Prediction error over the past few time steps (the difference between actual power and predicted power); Current power output: Power output at the current moment.

[0111] Furthermore, the action space describes the actions an agent can take. In power generation prediction, the action space includes the following two options:

[0112] Select the deep learning model module: Use a deep learning model for prediction;

[0113] Select the physical model module: Use a physical model for prediction.

[0114] Furthermore, the reward function is used to evaluate the quality of the actions taken by the agent. In power generation prediction, the reward function is designed based on the prediction error. In this embodiment, the reward function is a negative error reward function, meaning the reward value is the negative of the prediction error; that is, the more accurate the prediction, the higher the reward.

[0115] Furthermore, the policy network is the core component of reinforcement learning, used to select the optimal action based on the current state. In this embodiment, the policy network is a deep Q-network, which uses a deep neural network to approximate the Q-value function and learns the optimal policy by maximizing the cumulative reward.

[0116] Furthermore, experience replay is a technique to improve the training efficiency of reinforcement learning. The agent stores the experience (state, action, reward, next state) of each interaction in an experience replay cache, and then randomly samples small batches of data from the cache for training, so as to break the correlation between data and improve the stability and convergence speed of training.

[0117] Furthermore, the dynamic switching module dynamically selects either a data-driven model or a physical model for prediction based on the output of the policy network.

[0118] This dynamic switching mechanism dynamically selects the appropriate prediction module (data-driven model or physical model) based on wind speed fluctuation thresholds and feedback from the reinforcement learning module. The implementation steps of the dynamic switching mechanism are as follows:

[0119] Calculate wind speed fluctuations. At each time step t, calculate the difference between the current wind speed and the wind speed at the previous time step.

[0120] Set the wind speed fluctuation threshold;

[0121] In this step, a wind speed fluctuation threshold is set based on historical data and experience. Specifically, the statistical characteristics of all wind speed fluctuation values ​​are calculated, such as the mean, standard deviation, maximum value, and minimum value. Then, based on the required prediction accuracy for the wind farm, a suitable percentile is selected as the threshold. Common choices include the 90th, 95th, or 99th percentile. For example, choosing the 95th percentile means that in most cases (95% of the time), wind speed fluctuations will not exceed this value.

[0122] In this embodiment, the threshold is set to the 95th percentile of historical wind speed fluctuations.

[0123] Based on the wind speed fluctuation threshold and feedback from the reinforcement learning module, the data-driven model or the physical model is dynamically selected. The specific logic is as follows: if the wind speed fluctuation exceeds the wind speed fluctuation threshold, the physical model is selected for power generation prediction; if the wind speed fluctuation is less than or equal to the wind speed fluctuation threshold, the deep learning model module or the physical model module is selected based on the feedback from the reinforcement learning module. The feedback from the reinforcement learning module is the output result of the reinforcement learning module, which includes two results: selecting the deep learning model and selecting the physical model module.

[0124] Experiments were conducted on actual wind farm data to verify and compare the prediction performance of using only a data-driven model or a physical model. Experimental results show that, with the introduction of the reinforcement learning module, the model can dynamically adjust its prediction strategy based on real-time wind speed fluctuations.

[0125] Meanwhile, compared to relying solely on wind speed fluctuations to determine whether a data-driven or physical model is used, the model selection logic provided in this embodiment can significantly improve prediction accuracy and reduce the average normalized root mean square error (RMSE), verifying the effectiveness and practicality of the reinforcement learning module. Furthermore, the model selection logic provided in this embodiment can dynamically adjust the prediction model based on real-time wind speed fluctuations and prediction errors, adapting to different operating conditions. Additionally, by combining wind speed fluctuation thresholds and reinforcement learning feedback, the dynamic switching mechanism improves the robustness and generalization ability of the prediction model.

[0126] S4: Input the real-time wind data from the preprocessed big data into the offshore wind farm power generation prediction model to obtain the power generation prediction results.

[0127] The aforementioned power prediction was conducted at a wind farm in Shandong Province. One year of historical data was collected, and power prediction was performed based on the historical data. The prediction was then compared with the actual power in the historical data, demonstrating that the proposed scheme has a good predictive effect.

[0128] Meanwhile, the average normalized root mean square error (RMSE) of different prediction methods (this embodiment, the data-driven model only, and the physical model only) was tested. The RMSE for this embodiment was 0.1589, for the data-driven model only it was 0.2056, and for the physical model only it was 0.2234. By combining the advantages of the data-driven model and the physical model, the prediction error was significantly reduced and the prediction accuracy was improved. Furthermore, the performance of this embodiment's scheme was compared under different (seasonal) meteorological conditions. In spring: this embodiment's RMSE was 0.1654, and the data-driven model only RMSE was... =0.2102, Physical Model Only: RMSE=0.2256; Summer: This Scheme: RMSE=0.1523, Data-Driven Model Only: RMSE=0.2034, Physical Model Only: RMSE=0.2189; Autumn: This Scheme: RMSE=0.1487, Data-Driven Model Only: RMSE=0.1956, Physical Model Only: RMSE=0.2145; Winter: This Scheme: RMSE=0.1712, Data-Driven Model Only: RMSE=0.2201, Physical Model Only: RMSE=0.2312; It can be seen that the scheme in this embodiment shows high prediction accuracy under different seasons and meteorological conditions. Especially in winter, when wind speed fluctuates significantly, the advantages of physical models become more apparent. The solution in this embodiment effectively utilizes the advantages of physical models through a dynamic switching mechanism, thereby improving prediction accuracy. Simultaneously, a one-year long-term operational test was conducted on the model, recording monthly RMSE and the percentage of favorable switching. The results show that the annual average RMSE was 0.1623, with monthly RMSE fluctuations ranging from 0.1456 to 0.1897; the annual average percentage of favorable switching was 89.54%, with monthly favorable switching fluctuations ranging from 85.23% to 92.15%. This demonstrates that the solution in this embodiment exhibits high stability and robustness during long-term operation, continuously providing high-precision wind power predictions, and is suitable for the long-term operational needs of actual wind farms.

[0129] Example 2

[0130] This invention also provides a big data-based offshore wind farm power generation prediction system, which can employ a big data-based offshore wind farm power generation prediction method based on Embodiment 1, comprising:

[0131] The data acquisition module is used to acquire big data for predicting the power generation of offshore wind farms;

[0132] The preprocessing module is used to perform data preprocessing operations on the big data;

[0133] The model building module is used to build a prediction model for the power generation of offshore wind farms.

[0134] The power prediction module is used to input real-time wind data from the big data into the offshore wind farm power prediction model to obtain power prediction results.

[0135] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the power generation of offshore wind farms based on big data, characterized in that, Includes the following steps: S1: Obtain big data for predicting the power generation of offshore wind farms; S2: Perform data preprocessing operations on the big data to obtain preprocessed big data; in step S2, data preprocessing includes timestamp alignment; the timestamp alignment process is as follows: Sa: Select the reference time resolution based on the acquisition frequency of different types of data; Sb: Multi-level resampling is used to adjust large datasets from different acquisition frequencies to the reference time resolution; step Sb specifically involves: First-level resampling is performed on big data of different frequencies to achieve initial alignment; A second-level resampling is performed on the large datasets of different frequencies that have undergone initial alignment; The second-level resampling includes: performing multidimensional interpolation on sensor data that has been initially aligned and whose acquisition frequency is lower than the reference time resolution; and performing dynamic weight aggregation on sensor data that has been initially aligned and whose acquisition frequency is higher than the reference time resolution. S3: Establish a prediction model for the power generation of offshore wind farms; The offshore wind farm power generation prediction model is an ultra-short-term power generation prediction model based on a data-physical hybrid driving system. The ultra-short-term power generation prediction model includes: The deep learning model module is used to extract time series features from the preprocessed big data; The physical model module is used to predict power based on the physical characteristics of the wind farm; A reinforcement learning module is used to dynamically adjust the switching strategy of the ultra-short-term power generation prediction model; A dynamic switching mechanism is used to dynamically select either the deep learning model module or the physical model module based on the wind speed fluctuation threshold and the feedback from the reinforcement learning module; the implementation steps of the dynamic switching mechanism are as follows: To calculate wind speed fluctuations, at each time step t, calculate the difference between the wind speed at the current time step and the wind speed at the previous time step. Set the wind speed fluctuation threshold; If the wind speed fluctuation exceeds the wind speed fluctuation threshold, the physical model module is selected for power generation prediction; if the wind speed fluctuation is less than or equal to the wind speed fluctuation threshold, the deep learning model module or the physical model module is selected based on the feedback from the reinforcement learning module. S4: Input the real-time wind data in the preprocessed big data into the offshore wind farm power generation prediction model to obtain the power generation prediction result.

2. The method for predicting the power generation of offshore wind farms based on big data according to claim 1, characterized in that, In step S3, the input to the deep learning model module is the preprocessed big data; The network architecture of the deep learning model module includes: a multi-layer convolutional neural network, a long short-term memory network, a Transformer encoder, and an output layer.

3. The method for predicting the power generation of offshore wind farms based on big data according to claim 2, characterized in that, The multi-layer convolutional neural network is used to extract spatial features from the preprocessed big data, and its specific structure includes: Convolutional layer: Uses multiple convolutional kernels of different sizes to extract multi-scale features and obtain feature maps; Pooling layers: Max pooling or average pooling is used after each convolutional layer to reduce feature dimensionality; Feature fusion layer: fuses the feature maps extracted by the convolutional layer.

4. The method for predicting the power generation of offshore wind farms based on big data according to claim 2, characterized in that, The Long Short-Term Memory (LSTM) network is used to capture long-term dependencies in time series data from the preprocessed big data. Its specific structure includes: Input layer: The feature maps extracted by the multi-layer convolutional neural network are used as input; Multi-layer LSTM: A multi-layer LSTM structure is used, where each LSTM layer performs deeper feature extraction on the input time series; each LSTM unit of the multi-layer LSTM contains an input gate, a forget gate, and an output gate to control the inflow, retention, and outflow of information; Dropout layer: Add a Dropout layer between LSTM layers to prevent overfitting.

5. The method for predicting the power generation of offshore wind farms based on big data according to claim 4, characterized in that, The Transformer encoder is used to further enhance the deep learning model module's ability to model the time series, and its specific structure includes: Input embedding layer: Embeds the output feature map of each LSTM layer into the input space of the Transformer encoder; Multi-head self-attention mechanism: enables the Transformer encoder to capture features from different subspaces; Feedforward Neural Network: Each Transformer encoder contains a feedforward neural network for further processing of features; Residual connectivity and normalization: used to improve model stability and training efficiency.

6. The method for predicting the power generation of offshore wind farms based on big data according to claim 1, characterized in that, In S3, the physical model module is used to predict power generation by utilizing the physical characteristics of the wind farm. The physical model module includes a wind speed-power curve fitting layer and a dynamic system model. The wind speed-power curve fitting layer is a physical model that describes the relationship between wind speed and power generation. The dynamic system model is used to describe the dynamic relationship between wind speed changes and power generation output.

7. A big data-based offshore wind farm power generation prediction system, characterized in that, The offshore wind farm power generation prediction system adopts the big data-based offshore wind farm power generation prediction method described in any one of claims 1-6, which includes: The data acquisition module is used to acquire big data for predicting the power generation of offshore wind farms; The preprocessing module is used to perform data preprocessing operations on the big data to obtain preprocessed big data; The model building module is used to build a predictive model for the power generation of offshore wind farms; The power prediction module is used to input the real-time wind data in the preprocessed big data into the offshore wind farm power prediction model to obtain the power prediction result.

Images