Wind power prediction method based on long short-term memory network

Abstract

The invention discloses a wind power prediction method based on a long short-term memory network, and the method comprises the following steps: 1, obtaining a multi-dimensional data set of a target wind power plant in real time through a data collection system, and the multi-dimensional data set at least comprises historical power data, numerical weather forecast data, fan operation state parameters and geographic information data; according to the invention, through a multi-dimensional feature fusion mechanism and a dynamic weight optimization strategy, the prediction precision is significantly improved; the method comprises the following steps: firstly, fusing weather forecast, a fan operation state and topographic feature data, constructing feature data with physical significance, and enhancing the characterization capability of a model to a complex power grid engineering environment; secondly, a self-adaptive feature selection module is introduced to dynamically optimize the input weight, and redundant data interference is reduced; and finally, an attention mechanism is adopted to enhance feature extraction of key time nodes, so that the fluctuation ratio of a prediction curve is reduced.

Description

Technical Field

[0001] This application relates to the field of power grid engineering technology, specifically to a wind power prediction method based on long short-term memory networks. Background Technology

[0002] As wind power becomes a clean energy source, its share in power grid projects continues to increase, but the intermittency and volatility of wind energy pose a significant challenge to the safe operation of the power grid.

[0003] Among the existing technologies, CN202110404300.6 discloses a technology for improving the wind power ultra-short-term prediction method using a bidirectional long short-term memory network. When implemented, this method utilizes the bidirectional memory performance of the bidirectional long short-term memory network for historical power data and digital weather forecasts, incorporates an attention mechanism, and weights the output of the hidden layer of the bidirectional long short-term memory network, thereby effectively improving the prediction accuracy.

[0004] However, when implementing the above scheme, the following steps are taken: Data preprocessing is performed on the training samples; historical wind power data vectors from wind farms are collected, and the data with the highest correlation to wind power data in numerical weather prediction are selected and normalized; an attention mechanism model is integrated between the bidirectional long short-term memory network and the fully connected layer; the improved bidirectional long short-term memory network algorithm model is trained and its parameters optimized using the training set; the test set is input into the improved bidirectional long short-term memory network model to establish an ultra-short-term wind power prediction model, and the output value P of wind power is calculated one hour later. pre The method of predicting wind power output is used, but when the raw data is applied directly during implementation, the prediction fluctuates greatly during periods of sudden wind speed changes due to the single feature dimension (only wind speed + power) and lack of physical constraints. In particular, when there is a wake effect in the wind turbine cluster, the prediction deviation will be further amplified, making it difficult to meet the needs of grid dispatch. Summary of the Invention

[0005] To address this issue, this application provides a wind power prediction method based on long short-term memory networks, which solves the problem that the prediction deviation in existing technologies is further amplified due to the single feature dimension (only wind speed + power) and lack of physical constraints.

[0006] To achieve the above objectives, this application provides the following technical solution:

[0007] A wind power prediction method based on long short-term memory networks includes the following steps:

[0008] Step 1: Acquire the multidimensional dataset of the target wind farm in real time through the data acquisition system. The multidimensional dataset includes at least historical power data, numerical weather forecast data, wind turbine operating status parameters, and geographic information data.

[0009] Step 2: Perform spatiotemporal alignment processing on the multidimensional dataset. The spatiotemporal alignment processing includes data cleaning, missing value imputation, and timestamp standardization to generate a standardized input matrix.

[0010] Step 3: Construct a feature fusion module to nonlinearly combine meteorological elements, equipment status, and spatial location information in the standardized input matrix to generate a high-dimensional feature vector;

[0011] Step 4: Design the main structure of the Long Short-Term Memory Network. The main structure of the Long Short-Term Memory Network includes an input layer, a hidden layer, and an output layer connected in sequence. The hidden layer is composed of multiple stacked LSTM units.

[0012] Step 5: Cascade an attention mechanism module after the hidden layer to dynamically calculate the weight distribution of feature vectors at different time steps and generate context-aware feature maps.

[0013] Step 6: Use a sliding time window strategy to divide the training set and the test set, and optimize the network parameters based on the adaptive moment estimation algorithm to obtain the network model;

[0014] Step 7: Utilize the trained network model to receive the high-dimensional feature vector from Step 3 and output the wind power prediction sequence for the next 12-72 hours.

[0015] Step 8: The predicted sequence is smoothed and corrected by physical constraints through the post-processing module to generate the final power prediction result.

[0016] Preferably, the data cleaning in step two includes the following steps:

[0017] Step 201: Use the sliding interquartile range method to identify and remove outliers;

[0018] Step 202: Fill in the missing data by linear interpolation between adjacent time points;

[0019] Step 203: Perform Z-score standardization on wind speed and temperature data.

[0020] Preferably, the nonlinear combination in step three is achieved through a feature cross-layer, specifically including the following steps:

[0021] Step 301: Multiply the wind speed and the wind turbine pitch angle.

[0022] Step 302: Construct second-order polynomial features for air density and altitude;

[0023] Step 303: Use the hyperbolic tangent function to fuse temperature and humidity parameters.

[0024] Preferably, the LSTM unit in step four includes a forget gate, an input gate, and an output gate, wherein:

[0025] The forget gate uses the Sigmoid activation function;

[0026] The input gate includes a sigmoid and hyperbolic tangent dual-path structure;

[0027] Set the gradient clipping threshold for the output gate.

[0028] Preferably, the attention mechanism module in step five performs the following operations:

[0029] Step 501: Calculate the attention score for the hidden state at each time step;

[0030] Step 502: Normalize the score using the Softmax function to generate a weight vector;

[0031] Step 503: Perform a weighted summation of the hidden states to generate a context vector.

[0032] Preferably, the adaptive moment estimation algorithm in step six includes an initial learning rate, a first-order moment estimation decay factor, and a second-order moment estimation decay factor, and has the following characteristics:

[0033] The initial learning rate was set to 0.001-0.005;

[0034] The first-order moment estimates the attenuation factor β1 = 0.9;

[0035] The second-order moment estimate attenuation factor β2 = 0.999.

[0036] Preferably, step seven further includes quantifying the prediction uncertainty, specifically:

[0037] Step 701: Use Monte Carlo Dropout to generate probability prediction intervals;

[0038] Step 702: Calculate the 90% confidence interval using quantile regression;

[0039] Step 703: Output the predicted value and the upper and lower boundary curves.

[0040] Preferably, the physical constraint correction in step eight includes the following steps:

[0041] Step 801: Apply the zero power lower limit constraint;

[0042] Step 802: Set the upper limit constraint for rated power;

[0043] Step 803: Perform slope correction based on the wind turbine power curve.

[0044] Compared with the prior art, this application has at least the following beneficial effects:

[0045] When this invention is implemented, by performing physical constraint correction, a large number of theoretical infeasibility points can be eliminated, so that the prediction results meet the wind turbine ramp rate limit (≤3% / min). Thus, under the premise of ensuring grid safety, the reserve power generation capacity reserved by the system to cope with emergencies (such as load fluctuations, unit failures, and fluctuations in new energy output) is reduced.

[0046] By employing a multi-dimensional feature fusion mechanism and a dynamic weight optimization strategy, the prediction accuracy is significantly improved. First, weather forecasts, wind turbine operating status, and terrain feature data are fused to construct physically meaningful feature data, enhancing the model's ability to represent complex power grid engineering environments. Second, an adaptive feature selection module is introduced to dynamically optimize input weights and reduce redundant data interference. Finally, an attention mechanism is used to enhance feature extraction at key time nodes, thereby reducing the volatility of the prediction curve. Attached Figure Description

[0047] Figure 1 This is a flowchart of a wind power prediction method based on long short-term memory networks according to this application. Detailed Implementation

[0048] The present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0049] like Figure 1 As shown, this application discloses a wind power prediction method based on long short-term memory networks, including the following steps:

[0050] Step 1: Data Acquisition. A multidimensional dataset of the target wind farm is acquired in real-time through a data acquisition system. This dataset includes at least historical power data, numerical weather forecast data, wind turbine operating status parameters, and geographic information data. By synchronously acquiring numerical weather forecasts, wind turbine operating parameters, and geographic information for the wind farm, a multidimensional data input system is constructed. This solves the problem of single feature dimensions in traditional methods, enabling subsequent predictions to obtain more comprehensive environmental and equipment status data. When collecting historical power data, historical power data of wind turbines within a predetermined time period is collected to facilitate subsequent prediction of wind turbine power. For example, historical power data of wind turbines is collected every 10 minutes, continuously collecting data from the past 2 hours.

[0051] Step 2, spatiotemporal alignment processing: The multidimensional dataset is subjected to spatiotemporal alignment processing, which includes data cleaning, missing value imputation, and timestamp standardization to generate a standardized input matrix. By performing timestamp unification, spatial mapping, and outlier cleaning operations on heterogeneous data, the data distortion problem caused by sensor failure or coordinate system differences is solved, thereby eliminating spatiotemporal reference misalignment interference, ensuring the spatiotemporal consistency of the input matrix, and providing a reliable data source for model training.

[0052] Step 3: Feature fusion of data. A feature fusion module is constructed to nonlinearly combine meteorological elements, equipment status and spatial location information in the standardized input matrix to generate high-dimensional feature vectors.

[0053] Step 4: Network construction. Design the main structure of the Long Short-Term Memory Network. The main structure of the Long Short-Term Memory Network includes an input layer, a hidden layer and an output layer connected in sequence. The hidden layer is composed of multiple stacked LSTM units. By configuring LSTM units, it is possible to effectively capture the wind power change pattern within a 72-hour period and ensure the stability of the prediction results under sudden environmental changes.

[0054] Step 5: Attention enhancement. An attention mechanism module is cascaded after the hidden layer to dynamically calculate the weight distribution of feature vectors at different time steps and generate context-aware feature maps.

[0055] Step 6: Model training. A sliding time window strategy is used to divide the training set and the test set. The network parameters are optimized based on the adaptive moment estimation algorithm to obtain the network model. Through sliding window data partitioning and adaptive moment estimation optimization, the model aging problem caused by climate model changes is solved, and the prediction error between summer and winter seasons is reduced.

[0056] Step 7: Power prediction. The trained network model receives the high-dimensional feature vector from Step 3 and outputs a wind power prediction sequence for the next 12-72 hours.

[0057] Step 8, Result Correction: The post-processing module performs smoothing filtering and physical constraint correction on the predicted sequence to generate the final power prediction result. Material constraint correction can eliminate theoretically feasible but practically infeasible data (such as data with negative power or data exceeding the rated value), ensuring more accurate prediction results and reducing the risk of misjudgment.

[0058] In its implementation, this solution addresses the insufficient representational power of single features through multi-dimensional data fusion. Spatiotemporal alignment ensures the comparability of heterogeneous data; the feature fusion module overcomes the limitations of traditional single-dimensional modeling; the LSTM main structure captures long-term temporal patterns; the attention mechanism enhances feature extraction at key time nodes; the sliding window strategy optimizes the timely utilization of data; and the post-processing module generates the final prediction results.

[0059] In the implementation of this scheme, steps one to three are used to collect and process the data for prediction. Steps four to six are used to build and train the prediction model. In step seven, the data processed in step three is input into the model trained in step six to predict the wind turbine power. Step eight is used to correct the prediction results, which further improves the accuracy of the prediction results.

[0060] The data cleaning in step two includes the following steps:

[0061] Step 201: Use the sliding interquartile range method to identify and remove outliers;

[0062] Step 202: Fill in the missing data by linear interpolation between adjacent time points;

[0063] Step 203: Perform Z-score standardization on wind speed and temperature data.

[0064] The sliding interquartile range (IIL) method dynamically adapts to changes in data distribution, avoiding false rejection caused by fixed thresholds; linear interpolation maintains data continuity and is computationally efficient; Z-score standardization eliminates the influence of dimensions while preserving the data distribution pattern. This combined approach significantly reduces the data anomaly rate and substantially improves the stability of subsequent model training.

[0065] The nonlinear combination in step three is achieved through a feature cross layer, specifically including the following steps:

[0066] Step 301: Multiply the wind speed and the wind turbine pitch angle.

[0067] Step 302: Construct second-order polynomial features for air density and altitude;

[0068] Step 303: Use the hyperbolic tangent function to fuse temperature and humidity parameters.

[0069] The product of wind speed and blade pitch angle reflects aerodynamic energy capture efficiency; the polynomial correlation between air density and altitude corrects for the influence of air pressure; and the nonlinear fusion of temperature and humidity characterizes changes in atmospheric state. This design makes the physical meaning of the input features clear and reduces the prediction error compared to the original feature combination.

[0070] The LSTM unit described in step four includes a forget gate, an input gate, and an output gate, wherein:

[0071] The forget gate uses the Sigmoid activation function;

[0072] The input gate includes a sigmoid and hyperbolic tangent dual-path structure;

[0073] Set the gradient clipping threshold for the output gate.

[0074] The forget gate controls the decay rate of historical information to avoid interference from outdated data; the dual-path input gate enhances feature extraction capabilities; and gradient clipping suppresses the risk of training divergence. This scheme reduces the fluctuation of prediction errors in the model under sudden events such as typhoons.

[0075] The attention mechanism module in step five performs the following operations:

[0076] Step 501: Calculate the attention score for the hidden state at each time step;

[0077] Step 502: Normalize the score using the Softmax function to generate a weight vector;

[0078] Step 503: Perform a weighted summation of the hidden states to generate a context vector.

[0079] Attention scores are used to quantify the importance of features, softmax ensures the interpretability of weights, and weighted summation is used to focus on key information. This approach significantly improves the model's prediction accuracy during periods of rapid wind speed changes, particularly enhancing the prediction accuracy of peak and trough power points.

[0080] The adaptive moment estimation algorithm in step six includes an initial learning rate, a first-order moment estimation decay factor, and a second-order moment estimation decay factor, and has the following characteristics:

[0081] The initial learning rate was set to 0.001-0.005;

[0082] The first-order moment estimates the attenuation factor β1 = 0.9;

[0083] The second-order moment estimate attenuation factor β2 = 0.999.

[0084] A specific initial learning rate range is used to balance convergence speed and stability; the dual decay factor adaptively adjusts parameter updates. This technique significantly reduces training time and allows the loss function to converge to a lower level.

[0085] Step seven also includes quantifying the prediction uncertainty, specifically:

[0086] Step 701: Use Monte Carlo Dropout to generate probability prediction intervals;

[0087] Step 702: Calculate the 90% confidence interval using quantile regression;

[0088] Step 703: Output the predicted value and the upper and lower boundary curves.

[0089] Monte Carlo Dropout is used to simulate network parameter distribution; quantile regression calculations can accurately construct confidence intervals; and multiple output modes are used to support risk management. This solution enables dispatching departments to assess the risks of extreme scenarios and improve grid security under high-proportion wind power integration.

[0090] The physical constraint correction in step eight includes the following steps:

[0091] Step 801: Apply the zero power lower limit constraint;

[0092] Step 802: Set the upper limit constraint for rated power;

[0093] Step 803: Perform slope correction based on the wind turbine power curve.

[0094] When implementing the above scheme, zero-power constraints prevent negative values ​​from occurring; rated power limits are consistent with equipment capabilities; and slope correction is used to match the wind turbine's ramping characteristics. Furthermore, during implementation, physically infeasible prediction points are eliminated, significantly improving the applicability of the prediction results.

[0095] The technical features of the above embodiments can be combined in any way (as long as there is no contradiction in the combination of these technical features). For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; these embodiments not explicitly written should also be considered to be within the scope of this specification.

Claims

1. A wind power prediction method based on long short-term memory networks, characterized in that, Includes the following steps: Step 1, Data Acquisition: The multidimensional dataset of the target wind farm is acquired in real time through the data acquisition system. The multidimensional dataset includes at least historical power data, numerical weather forecast data, wind turbine operating status parameters, and geographic information data. Step 2, Spatiotemporal alignment processing: The multidimensional dataset is subjected to spatiotemporal alignment processing, which includes data cleaning, missing value imputation, and timestamp standardization to generate a standardized input matrix; Step 3, feature fusion of data: Construct a feature fusion module to nonlinearly combine meteorological elements, equipment status and spatial location information in the standardized input matrix to generate high-dimensional feature vectors of historical power data, numerical weather forecast data, wind turbine operating status parameters and geographic information data; Step 4, Network Construction: Design the main structure of the Long Short-Term Memory Network. The main structure of the Long Short-Term Memory Network includes an input layer, a hidden layer, and an output layer connected in sequence. The hidden layer is composed of multiple stacked LSTM units. Step 5, Attention Enhancement: An attention mechanism module is cascaded after the hidden layer to dynamically calculate the weight distribution of feature vectors at different time steps and generate context-aware feature maps; Step 6, Model Training: The training set and test set are divided using a sliding time window strategy, and the network parameters are optimized based on the adaptive moment estimation algorithm to obtain the network model; Step 7, Power Prediction: Utilize the trained network model to receive high-dimensional feature vectors of historical power data, numerical weather forecast data, wind turbine operating status parameters, and geographic information data, and output a wind power prediction sequence for the next 12-72 hours. Step 8, Result Correction: The prediction sequence is smoothed and corrected by physical constraints through the post-processing module to generate the final power prediction result.

2. The wind power prediction method based on long short-term memory networks according to claim 1, characterized in that, The data cleaning in step two includes the following steps: Step 201: Use the sliding interquartile range method to identify and remove outliers; Step 202: Fill in the missing data by linear interpolation between adjacent time points; Step 203: Perform Z-score standardization on wind speed and temperature data.

3. The wind power prediction method based on long short-term memory networks according to claim 1, characterized in that, The nonlinear combination in step three is achieved through a feature cross layer, specifically including the following steps: Step 301: Multiply the wind speed and the wind turbine pitch angle. Step 302: Construct second-order polynomial features for air density and altitude; Step 303: Use the hyperbolic tangent function to fuse temperature and humidity parameters.

4. The wind power prediction method based on long short-term memory networks according to claim 1, characterized in that, The LSTM unit described in step four includes a forget gate, an input gate, and an output gate, wherein: The forget gate uses the Sigmoid activation function; The input gate includes a sigmoid and hyperbolic tangent dual-path structure; Set the gradient clipping threshold for the output gate.

5. The wind power prediction method based on long short-term memory networks according to claim 1, characterized in that, The attention mechanism module in step five performs the following operations: Step 501: Calculate the attention score for the hidden state at each time step; Step 502: Normalize the score using the Softmax function to generate a weight vector; Step 503: Perform a weighted summation of the hidden states to generate a context vector.

6. The wind power prediction method based on long short-term memory networks according to claim 1, characterized in that, The adaptive moment estimation algorithm in step six includes an initial learning rate, a first-order moment estimation decay factor, and a second-order moment estimation decay factor, and has the following characteristics: The initial learning rate was set to 0.001-0.005; The first-order moment estimates the attenuation factor β1 = 0.9; The second-order moment estimate attenuation factor β2 = 0.

999.

7. The wind power prediction method based on long short-term memory networks according to claim 1, characterized in that, Step seven also includes quantifying the prediction uncertainty, specifically: Step 701: Use Monte Carlo Dropout to generate probability prediction intervals; Step 702: Calculate the 90% confidence interval using quantile regression; Step 703: Output the predicted value and the upper and lower boundary curves.

8. The wind power prediction method based on long short-term memory network according to claim 1, characterized in that, The physical constraint correction in step eight includes the following steps: Step 801: Apply the zero power lower limit constraint; Step 802: Set the upper limit constraint for rated power; Step 803: Perform slope correction based on the wind turbine power curve.

Images