Wind power prediction method and system combining wind turbine operating status and error correction

Abstract

The present invention belongs to the technical field of power system power prediction and provides a wind power prediction method and system that combines wind turbine operating status with error correction. The technical solution is as follows: obtaining historical wind turbine operating status data and numerical weather forecast grid data; preprocessing the numerical weather forecast grid data to obtain a wind turbine conversion power sequence; combining the wind turbine historical power sequence, the wind turbine conversion power sequence, and a trained spatiotemporal joint prediction model to obtain a unified predicted power sequence for each wind turbine; adding a separate error correction model to each wind turbine, combining the wind turbine historical operating status data, the unified predicted power sequence, and the predicted meteorological elements at the corresponding location of each wind turbine to obtain an error prediction sequence; using the error prediction sequence to correct the unified predicted power sequence to obtain a corrected unified wind power prediction sequence. This method also considers the complex spatiotemporal correlations between wind turbines and the differences in the operation of each wind turbine, greatly improving the accuracy of wind power prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system power prediction, and in particular relates to a wind power prediction method and system combining wind turbine operating status and error correction. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] As global carbon reduction efforts progress, more and more wind power is being connected to power systems. Globally, 680GW of new wind power capacity will be connected to the grid over the next five years. Because wind power output is highly random, its integration into the grid inevitably introduces potential instability. Accurately forecasting wind farm output in advance can effectively mitigate the impact of wind power on the system and reduce the cost of providing spinning reserve.

[0004] Wind power prediction (WPP) tasks can be divided into ultra-short-term, short-term, and medium- to long-term predictions based on their time scales. Decades of research have shown that basic WPP methods can be categorized into three types: physical methods, statistical methods, and artificial intelligence (AI) methods.

[0005] Physical methods rely on high-precision Numerical Weather Prediction (NWP). NWP can provide forecast elements such as wind speed, wind direction, and air pressure for wind farms within the simulation calculation range, and use this as a basis to obtain the wind farm's output curve. Physical methods modeling takes into account topography and atmospheric movement, solving thermodynamic and aerodynamic equations, and have stable performance in long-term forecasts. However, physical methods have high requirements for computing resources, and the forecast geographic resolution is insufficient, which easily leads to temporal and spatial deviations in the forecast, making them unsuitable for shorter-term forecasts of 0-4 hours.

[0006] Statistical methods can extrapolate wind power forecasts based on historical data, achieving high accuracy in ultra-short-term and short-term WPPs. Commonly used statistical models include the persistence method (PM), autoregressive integrated moving average (ARIMA), regression tree (RT), and support vector regression (SVR). While statistical methods are easy to implement, they face the complex, nonlinear, and unsteady variations of wind power, and their performance degrades rapidly as the forecast scale increases.

[0007] As AI methods advance in other fields, WPP researchers have also shown great interest in them. AI not only enables extrapolated forecasts of wind power but is also being used to correct for prediction biases in physical methods. AI models possess complex nonlinear learning capabilities and can be trained using large amounts of data. Already applied AI methods include shallow artificial neural networks (ANNs), extreme learning machines (ELMs), radial basis function (RBF) neural networks, and multilayer perceptrons (MLPs). With the rise of deep learning in recent years, many specialized neural network architectures have been applied to WPP, including deep Boltzmann machines (DBNs), Elman neural networks (ENNs), recurrent neural networks (RNNs) for time series learning, and their variants, gated recurrent units (GRUs), long short-term memory networks (LSTMs), and convolutional neural networks (CNNs) for spatial feature learning. Although AI models generally outperform statistical methods, the prediction accuracy of a single model remains limited. Furthermore, a single model struggles to cope with the complex spatiotemporal correlations of WPP.

[0008] To further improve WPP performance, existing technologies have proposed the following solutions:

[0009] A variety of prediction strategies are used, including a hybrid prediction architecture, combined optimization of multi-model prediction results, and secondary prediction based on error correction (EC); a hybrid architecture of K-shape integrated GRU and K-means guided CNN (KK-CNN-GRU), which includes three modules and outperformed seven other models in prediction experiments; preliminary predictions are made using three models, namely ANN, SVR, and Gaussian process (GP), and then the prediction results of the three models are integrated through the Bayesian model average, which improves the prediction effect of a single model; the prediction sequence of historical data and WRF mode is combined, and then corrected using the CNN-BiLSTM hybrid model to achieve higher-precision multi-step sequence prediction; the recorded data sequence is multi-frequency graded, with low-frequency components used for preliminary predictions and high-frequency components used to train a dedicated EC model.

[0010] The above research has improved the prediction accuracy of WPP, but the complex architecture and combination of multiple methods increase the difficulty of optimizing training parameters. Therefore, a unified architecture, Predictive Stacked Autoencoder (PSAE), is proposed to implement sequence-to-sequence prediction tasks in an end-to-end manner, sharing the loss function during training. A spatiotemporal prediction architecture is proposed, which first uses a 2D CNN to extract two-dimensional spatial features of a 10×10 wind turbine matrix, and then uses an LSTM to learn temporal features.

[0011] However, the above scheme still has an unresolved problem, which is to provide WPP for all wind turbines in the wind farm at the same time, and take into account the complex temporal and spatial correlations between wind turbines and the differences in the operation of each wind turbine. Summary of the Invention

[0012] In order to solve at least one technical problem existing in the above-mentioned background technology, the present invention provides a wind power prediction method and system that combines the operating status of wind turbines with error correction. It proposes a unified wind power prediction framework. The prediction framework includes a spatiotemporal joint prediction model and multiple single-machine error correction models. At the same time, it takes into account the complex spatiotemporal correlations between wind turbines and the differences in the operation of each wind turbine, thereby greatly improving the accuracy of wind power prediction.

[0013] In order to achieve the above object, the present invention adopts the following technical solutions:

[0014] A first aspect of the present invention provides a wind power prediction method combining wind turbine operating status and error correction, comprising the following steps:

[0015] Obtain historical operating status data of wind turbines and numerical weather forecast grid data;

[0016] Preprocess the numerical weather forecast grid data to obtain the wind turbine conversion power series;

[0017] Combine the wind turbine historical power series, wind turbine conversion power series and the trained spatiotemporal joint prediction model to obtain a unified predicted power series for each wind turbine;

[0018] A separate error correction model is added to each wind turbine, which combines the historical operating status data of the wind turbine, the unified predicted power sequence and the predicted meteorological factors at the corresponding location of each wind turbine to obtain the error prediction sequence;

[0019] The error prediction sequence is used to correct the unified predicted power sequence to obtain the corrected unified wind power prediction sequence.

[0020] A second aspect of the present invention provides a wind power prediction system combining wind turbine operating status and error correction, comprising:

[0021] The data acquisition module is configured to: acquire historical operating status data of the wind turbine and numerical weather forecast grid data;

[0022] The data preprocessing module is configured to: preprocess the numerical weather forecast grid data to obtain a wind turbine conversion power sequence;

[0023] The initial power prediction module is configured to combine the wind turbine historical power sequence, the wind turbine conversion power sequence, and the trained spatiotemporal joint prediction model to obtain a unified predicted power sequence for each wind turbine;

[0024] The error prediction module is configured to: add a separate error correction model for each wind turbine, combine the historical operating status data of the wind turbine, the unified predicted power sequence and the predicted meteorological factors at the corresponding location of each wind turbine, and obtain the error prediction sequence;

[0025] The error correction module is configured to correct the unified predicted power sequence using the error prediction sequence to obtain a corrected unified wind power prediction sequence.

[0026] A third aspect of the present invention provides a computer-readable storage medium.

[0027] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the wind power prediction method combining wind turbine operating status and error correction as described above.

[0028] A fourth aspect of the present invention provides a computer device.

[0029] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of the wind power prediction method combining wind turbine operating status and error correction as described above are implemented.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] The present invention takes into account the complex spatiotemporal correlations between wind turbines and the differences in the operation of each wind turbine. First, a unified predicted power sequence for each wind turbine is obtained as a whole through a spatiotemporal joint prediction model; then a separate error correction model is added to each wind turbine, and the error prediction sequence is obtained by combining the historical operating status data of the wind turbine, the unified predicted power sequence and the predicted meteorological elements at the corresponding position of each wind turbine; the error prediction sequence is used to correct the unified predicted power sequence to obtain a corrected unified wind power prediction sequence, thereby improving the wind power prediction accuracy of all wind turbines in the wind farm.

[0032] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0034] Figure 1 This is a numerical weather forecast grid data preprocessing process provided by an embodiment of the present invention;

[0035] Figure 2 This is the data training process of the STJPM unified prediction model provided by an embodiment of the present invention;

[0036] Figure 3Schematic diagram of the temporal convolutional network structure provided by an embodiment of the present invention;

[0037] Figure 4 This is the 2D ConvGRU unit calculation process provided by the embodiment of the present invention;

[0038] Figure 5 It is the multi-time-scale prediction curve of WPCC and the proposed method provided by the embodiment of the present invention;

[0039] Figure 6 The 0-24h prediction curves and errors of the method of the present invention at different stages are provided in the embodiments of the present invention;

[0040] Figure 7 This is a scatter plot of the actual and predicted distributions of the method proposed in the present invention during the 0-6 hours period, provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0042] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0043] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0044] In response to the prior art mentioned in the background, and taking into account the complex spatiotemporal correlations between wind turbines and the operational variability of each wind turbine, the present invention proposes a unified wind power forecasting framework. This framework comprises a spatiotemporal joint prediction model (STJPM) and multiple single-machine error correction (EC) models. The STJPM rationally combines the real-time operating data of all wind turbines with the forecast data provided by the NWP grid and GIDS, achieving joint prediction of multiple sequences using a shared loss function. The Com-GRU considers the operating status of each wind turbine and performs error correction on the output sequence of the unified prediction. The experimental data comes from a wind farm in mountainous terrain, and the proposed method achieves higher prediction accuracy. The proposed framework is unaffected by the arrangement of wind turbines and completes the spatiotemporal joint training of the unified prediction model in the form of a virtual wind farm. The STJPM has excellent versatility and can transform the spatiotemporal dimensions of sequence data through a TCN network, while adjusting the sequence length to achieve power forecasting at different time scales. The proposed framework also achieves higher-frequency predictions based on the WPCC. The real-time prediction of each wind turbine can also serve the refined scheduling of wind farms.

[0045] First, meteorological element forecasts at wind turbine locations are implemented using the NWP grid and the gradient plus inverse distance squared (GIDS) method to obtain the wind speed-to-power conversion series (WPCC) for each wind turbine. Subsequently, a spatiotemporal joint prediction model (STJPM) is used to achieve preliminary WPP for all wind turbines, based on a unified prediction strategy and single-machine error correction. Finally, a separate EC module is added for each wind turbine.

[0046] Example 1

[0047] This embodiment provides a unified wind power prediction method that combines wind turbine operating status and error correction, including the following steps:

[0048] Step 1: Obtain historical wind turbine operating data and numerical weather forecast grid data based on the WRF model;

[0049] In this example, the proposed model inputs include historical wind turbine operating data provided by a Supervisory Control and Data Acquisition (SCADA) system and NWP grid data based on the WRF model. All data has a strict time series, which requires a standardized input matrix based on this time series before entering the unified prediction model.

[0050] The WRF model, developed jointly by multiple research institutions including NCAR and NCEP, can achieve high-resolution three-dimensional geographic grid NWP. The WRF model offers distinct advantages, including a high degree of modularity and a hierarchical design, which allows for parallel computing. The model integrates currently available physical parameter schemes for simulating atmospheric processes over multiple days into the future.

[0051] In this example, the WRF model is used to provide the NWP grid for the target wind farm. The three nested grids are set to 100×100 (18 km), 121×121 (6 km), and 151×151 (2 km), and the number of vertical pressure layers is set to 35, based on the location and simulation domain of the selected wind farm. Based on the literature "A unified multi-step wind speed forecasting framework based on numerical weather prediction grids and wind farm monitoring data," sensitivity tests were conducted on the physical parameters of the WRF model, and the parameter combination method was optimized based on wind speed prediction accuracy.

[0052] Step 2: Preprocess the acquired numerical weather forecast grid data based on the WRF model;

[0053] Figure 1 The data processing and sequence organization process is given, including:

[0054] Step 201: First, long-term NWP forecast data is obtained based on the WRF model, and grid forecast data near the wind turbine is extracted according to the geographical location of the selected wind farm;

[0055] Step 202: The wind turbines in a wind farm located in a mountainous terrain are distributed very irregularly. The GIDS method is used to interpolate the four NWP grid points around the wind turbine to the wind turbine hub position to obtain a wind speed prediction sequence at each wind turbine position.

[0056] GIDS is based on the inverse distance weighted interpolation method and takes into account the gradient changes of meteorological factors with longitude, latitude and altitude. It has high interpolation accuracy in mountainous areas with complex terrain. Its spatial interpolation calculation formula is as follows:

[0057]

[0058] Where Z represents the estimated value of the wind turbine location; Z i is the predicted value of the i-th NWP grid point; d i is the distance between the i-th NWP grid point and the wind turbine location; N is the number of NWP grid points involved in the interpolation calculation; X, Y, E, X i 、Y i 、E i are the longitude, latitude and altitude of the wind turbine location and the i-th NWP grid point respectively; C x 、C y 、C eare the partial regression coefficients of the meteorological element values ​​of the NWP grid points and the longitude, latitude and altitude, respectively. Their values ​​directly determine the interpolation weight of each sample point;

[0059] The C x 、C y 、C e The coefficients are solved and the regression model is as follows:

[0060] Z i =C x X i +C y Y i +C e E i +C0 (2)

[0061] Where Z i represents the estimated value of the meteorological element at the i-th NWP grid point; C0 is the regression constant.

[0062] Step 203: After obtaining the wind speed prediction sequence, each wind turbine obtains the power sequence at the corresponding moment according to the wind speed-power conversion curve (WPCC) of the wind turbine.

[0063] Step 204: Then, read the historical wind power sequence of the SCADA system at the current time t and the previous time, and splice the historical power sequence with the converted power sequence according to the time sequence.

[0064] Take a wind farm with 33 wind turbines as an example:

[0065] For a single capacity of 3 MW, the input matrix dimension is 33×(k+q); k and q in the input matrix represent the lengths of the historical power sequence and the converted power sequence, respectively. k is determined by the prediction test results, and q determines the sequence length of the multi-step prediction.

[0066] After completing the above data processing, through STJPM unified prediction and error correction based on individual Com-GRU, the 33 wind turbines can obtain a predicted power sequence of length q.

[0067] Step 3: STJPM Unified Forecast

[0068] Figure 2 The data training process of the STJPM unified prediction model is presented. STJPM consists of a TCN feature extraction network and a 2DConvGRU network. During training, they share the loss function and weight parameters, strictly considering the dimensionality transformation of spatiotemporal features.

[0069] TCN feature extraction network

[0070] First, the 33×(k+q) input matrix enters the temporal convolutional network (TCN), and is extracted through a 4-layer dilated network to increase the number of filters layer by layer. The dilation factor is d=[1,2,4,8].

[0071] Compared to traditional convolutional neural networks, TCNs use causal dilated convolutions to extract features from time series and residual blocks to address the vanishing gradient problem. Causal convolutions are unidirectional, ensuring the forward flow of information, while dilated convolutions extend the receptive field of information extraction and reduce network depth.

[0072] Figure 3 In a TCN architecture, the convolution operations between upper-layer features and lower-layer features exhibit temporal causality. Causal convolution ensures that when extracting sequence features, the output at time t depends solely on the input information at or before time t. Causal convolution uses a 1-D full CNN structure to ensure that the length of the input and output sequences at each layer remains constant. Zero padding of length k-1 is performed before the input sequence, where k is the size of the convolution kernel.

[0073] A single-layer 1-D causal convolutional layer is calculated as follows:

[0074]

[0075] Where X=(x0,x1,x2,···,x T ) is the input time series, f is a filter: {0,1,···,k-1}→X, F(·) represents a convolution operation.

[0076] In ordinary causal convolution, a sequence that is too long will significantly increase the number of network layers, increase training time and computational cost. Therefore, dilated convolution is used to speed up the flow of information. The convolution formula can be rewritten as:

[0077]

[0078] The expansion factor d can expand the receptive field of each layer to (k-1)d. In order to ensure that the filters in the deep network do not produce input omissions when obtaining a very large effective history, d is generally chosen to be 2 in the rth hidden layer. r-1 .

[0079] Residual blocks are proposed to skip layers in deep neural networks to solve optimization problems such as deep network degradation and gradient vanishing. Assuming the output of the last layer is F(x) and the input is x, the final output of a residual block is:

[0080] H(x)=ReLU(F(x)+x) (5)

[0081] A residual block consists of two layers of dilated causal convolutions, followed by batch normalization using rectified linear units (ReLU). In addition, dropout is added after each dilated convolution within the residual block for regularization.

[0082] The final layer outputs a 1024×(k+q) feature matrix. While maintaining the time series length, the feature dimensions are converted to a three-dimensional matrix to meet the input requirements of the 2D ConvLSTM network layer. The input matrix for the 2D ConvLSTM network is 32×32×1×(k+q), corresponding to length, width, number of channels, and time series length, respectively.

[0083] 2D ConvGRU unit

[0084] like Figure 4 The figure shows the calculation process of a 2D ConvGRU unit. ConvGRU changes the calculation method of a single GRU unit from fully connected weights to convolutional feature extraction, reducing the calculation cost.

[0085] 2D ConvGRU can simultaneously learn temporal and spatial features during training. Each time unit ( Figure 4 ) is a 3D matrix consisting of the spatial dimensions length × width and the number of channels, where the number of cells represents the time dimension.

[0086] The calculation process of a 2D ConvGRU unit is as follows:

[0087] Z t =σ(W zx *X t +W zh *H t-1 +b z ) (6)

[0088] R t =σ(W rx *X t +W rh *H t-1 +b r ) (7)

[0089]

[0090]

[0091] Where: W zx 、W zh 、W rx 、W rh 、W hx 、W hh is the weight matrix; b z 、br 、b h is the bias vector; * represents the convolution operation; ⊙ represents the dot product operation; σ(·) represents the sigmoid function; F(·) represents the activation function; it is passed to the hidden state H at the next moment. t , candidate hidden states Update gate vector Z t and reset gate vector R t are all three-dimensional tensors.

[0092] Transforming the input data's dimensionality enables simultaneous associative learning of spatial and temporal features. Essentially, this involves rearranging wind turbines in a virtual wind farm. Joint learning of spatiotemporal features is accomplished using a three-layer 2D ConvLSTM network, with convolution kernels of 3×3. A 2×2 max pooling operation is applied after the first two ConvLSTM layers to expand the receptive field and reduce training costs.

[0093] After obtaining the output of the last 2D ConvLSTM network layer, all features are flattened using a Dense layer and mapped into a 33×q output matrix at the output layer. At this point, each wind turbine receives a uniform predicted power sequence.

[0094] During the STJPM model training process, we achieved end-to-end training using a unified architecture while maintaining a constant sequence length. Based on the unified prediction results, we calculated the prediction error for each wind turbine and used it to train the EC model for each wind turbine.

[0095] Step 4: Single fan error correction

[0096] During the error correction phase, the model considers the impact of multiple historical turbine operating variables on forecast accuracy, including historical power, generator speed, blade speed, and generator operating temperature. Furthermore, NWP forecast factors such as temperature, wind direction, humidity, and pressure, which are considered to affect power output, are also included in the EC model inputs, similarly obtained using the NWP grid and GIDS interpolation method.

[0097] The EC model input sequence is divided into t- and t+ variables based on the current time t, taking into account the timing characteristics of wind turbine operation. The EC model consists of two parallel GRU layers and a unified GRU layer, which is used as a Com-GRU model to perform error correction.

[0098] The wind turbine's historical operating status variable is treated as a t-variable. After Max-Min normalization, it enters a GRU layer, resulting in a 16×m output feature, where m represents the length of the read sequence. Each wind turbine's unified predicted power sequence and its geographically-based NWP prediction factor are treated as t+ variables. After normalization, they enter another parallel GRU layer, resulting in a 16×q output feature. The outputs of the GRU layers for the t- and t+ variables are then concatenated in time series and entered into a unified GRU layer, where the output is mapped into a single wind turbine error prediction sequence of length q.

[0099] Finally, the error prediction sequence is used to correct the power sequence uniformly predicted by STJPM to improve the WPP accuracy of all wind turbines in the wind farm.

[0100] In this embodiment, the evaluation index is used to measure the power prediction performance of the model. MAE and RMSE are widely used in error evaluation to reflect the long-term prediction performance of the model. The formula is as follows:

[0101]

[0102]

[0103] and p(j) t+τ is the predicted wind power and the actual wind power, t is the current time, τ is the prediction benchmark, and N is the number of samples in the test set.

[0104] In order to reflect the performance gap between different models under various prediction strategies, the performance improvement (PI) indicator is proposed. The calculation formula is as follows:

[0105]

[0106] Where R is the error value, j represents the current model, and each PI value is the average decrease ratio of MAE and RMSE between the current model and the previous model.

[0107] In order to verify the prediction performance of the proposed prediction method, the present invention conducted a prediction test in a real wind farm.

[0108] First, the case wind farm and the collected data set are introduced. Then, the experimental results of 11 models with five prediction strategies are provided, and the error analysis and discussion are carried out.

[0109] Take, for example, a wind farm in a certain city, located in mountainous terrain. The entire wind farm consists of a wind tower, a substation, and 33 3MW wind turbines. The spatial arrangement of the turbines, influenced by the terrain, is highly irregular.

[0110] The dataset contains two years of wind farm operational data, totaling 70,176 time points. The collected data includes output power and operating status data for 33 wind turbines at a 15-minute resolution, as well as NWP grid-forecast meteorological data at the same resolution. Because SCADA data often contain omissions or errors, eight consecutive points (i.e., two hours) of zero or unchanged data are considered abnormal data. After data processing and organization, abnormal data sequences are removed. Table 1 presents the basic power generation statistics of the wind farm.

[0111] Table 1 Wind farm operation statistics

[0112]

[0113]

[0114] This invention performs a one-day-ahead multi-step power series forecast for all wind turbines in a wind farm. In this embodiment, q is set to 96. Testing has shown that when k ≥ 6, the accuracy of the unified forecast model no longer improves. Considering the randomness of data records, k is set to 8. In single-turbine error correction, the value of m is determined by the actual forecast performance. In the present invention's experiments, m ranged from 7 to 12.

[0115] In both unified prediction and single-turbine error correction, the dataset was divided into training, validation, and test sets in a 4:1:1 ratio. The training set was used for model parameter training, with hyperparameters adjusted based on the validation set. The test set was used to compare the prediction performance of different methods. The proposed STJPM and Com-GRU models initialized their parameters using a normal distribution. The loss function used was Mae, which is included in the Python Keras module, and the parameter update used the Adaptive Moment Estimation (Adam) optimization algorithm with a dynamically adjustable learning rate.

[0116] In order to evaluate the performance improvement of the proposed method, another 10 different baseline methods are used for case wind farm power prediction, and the 11 models are divided into the following five implementation strategies.

[0117] 1) Wind Speed-Power Conversion Curve (WPCC): The power sequence is obtained based on the WPCC, which serves as the most basic prediction benchmark of the present invention.

[0118] 2) Single wind turbine single-step prediction (SVR, LightGBM): SVR

[0119] The kernel function is selected from poly, rbf, linear and sigmoid according to the prediction effect. The setting of the penalty factor C is determined by trial and error according to the reference.

[0120] LightGBM uses a grid search method to determine the learning rate and early_stop_round parameters.

[0121] 3) Multi-step prediction of a single wind turbine (ANN, RNN, LSTM): A sequence-to-sequence prediction method is used. Based on experimental results, the number of network layers is 2-4. Before the output layer, the number of neurons in each layer increases exponentially by 2.

[0122] 4) Unified multi-step prediction (TCN-3DCNN, STJPM): Use 3DCNN to replace the 2D ConvGRU network layer and regard the sequence time length as the height dimension of 3DCNN.

[0123] 5) Unified multi-step prediction and error correction (LightGBM, LSTM, Com-GRU): LightGBM is used for single-step error correction of a single wind turbine, while LSTM and Com-GRU implement multi-step sequence error correction of a single wind turbine.

[0124] Table 2 shows the multi-step forecast metric statistics for 11 models with forecast timeframes ranging from 15 minutes to 6 hours. Table 3 shows the average forecast metric statistics for four timeframes for one-day-ahead forecasts. WPCC performs worse than other models at all timescales, with the gap being more pronounced in the 0-6 hour timeframe. This is because WPCC relies on a 6-hour NWP forecast frequency, while other models can rely on monitoring data for high-frequency 15-minute forecasts. Furthermore, other models, based on WPCC, have been trained using actual power output, retaining the WPCC's stable forecast performance in the 6-24 hour timeframe.

[0125] Table 2 Single-point statistics of predicted MAE and RMSE indicators (MW)

[0126]

[0127] Table 3 Forecast MAE, RMSE indicator segment statistics (MW)

[0128]

[0129] The above analysis shows that for single-step prediction of a single wind turbine, LightGBM performs similarly to SVR, but significantly inferior to ANN, RNN, and LSTM for multi-step prediction. Multi-step prediction considers power trend changes, while RNN and LSTM have the ability to learn temporal features, making them superior to ANN. LSTM, on the other hand, builds on RNNs by solving the optimization problem of gradient training, and its performance has been validated in multiple publications. Unified multi-step prediction considers the mutual influence of wind turbine operation, further reducing prediction error. The proposed STJPM model outperforms TCN-3DCNN, evidently due to the 2D ConvGRU's more efficient handling of the temporal dimension during training than the 3DCNN. After adding the EC model, the Com-GRU model outperforms both the LightGBM with single-step error correction and the LSTM with multi-step error correction. This demonstrates that the error series still has strong temporal correlation, and the Com-GRU architecture is more suitable for combining the real-time operating characteristics of a single wind turbine with unified prediction results.

[0130] Figure 5 Sample curves of the actual and predicted power of the entire wind farm are presented, including advance predictions at multiple time scales using WPCC and the proposed method. The WPCC curve is relatively smooth, and the overall prediction trend does not deviate significantly, but the error increases significantly when the actual power fluctuates rapidly. The proposed method performs multi-step predictions for 0-24 hours based on WPCC. The predicted curves are close to the actual curves in the 15-minute and 30-minute predictions, and accurate predictions can be achieved even in areas with strong fluctuations. As the prediction scale increases, the prediction curve gradually smoothes, and obvious distortion of the prediction trend first appears in areas with large fluctuations. In the 0-6 hour prediction, the decline in prediction performance is more obvious than in the 6-24 hour prediction. This is because the 0-6 hour period is greatly affected by historical power, while as the prediction scale increases, the WPCC sequence begins to dominate.

[0131] The test results are analyzed for errors and further discussed. Figure 6 The prediction curves for the proposed method at different stages and the error distribution at different time points for a specific case day are presented. The basic WPCC curve exhibits a large error of 20MW in the first few steps starting at time t. Furthermore, an error value of up to 40MW is observed at the end of the case day, and the prediction trend significantly lags behind the actual power. However, the prediction value provided by STJPM significantly reduces the initial prediction error and effectively corrects the amplitude of the large error. After Com-GRU error correction, the overall prediction accuracy is further improved. It is worth noting that when the actual power is high during the forecast case day, the resulting error also increases.

[0132] Since the performance of the proposed method is relatively stable within 6-24h, Figure 7The scatter plot shows the actual, predicted, and error values ​​for 0-6 hours. As the forecast scale increases, the sample points gradually deviate from the central axis, and the distribution becomes more divergent. In the 15-minute forecast, the larger errors primarily come from higher actual power values. In the 6-hour forecast, even lower actual power values ​​can result in large errors. This demonstrates that as the forecast scale increases, the predicted power becomes increasingly conservative. The error scatter plot also shows an overall trend of overestimating low power values ​​and underestimating high power values.

[0133] To further demonstrate the superiority of the proposed method, Table 4 presents the PI statistics for 10 models, calculated with WPCC as the initial benchmark and referenced to the previous model. Compared to WPCC, even SVR achieves performance improvements of over 15% in the 0-24 hour forecast period, with an improvement of approximately 50% in the 15-minute forecast period. This demonstrates that after training with extensive real-world data, the prediction model effectively corrects for the spatiotemporal bias introduced by NWP forecast data. Notably, within the same prediction strategy, the performance improvements of different models range from 1% to 4%. However, after changing the prediction strategy, the performance improvements range from 4% to 10%. This demonstrates that the application of the new strategy is more effective than simply replacing the model, providing new insights for further research.

[0134] Table 4 Prediction PI index statistics (%)

[0135]

[0136] Experimental results show that the proposed method achieves the highest prediction accuracy, primarily due to two distinct advantages. 1) The proposed method effectively combines deep learning with physical methods. Deep learning incorporates real-time wind turbine operating data, while physical methods consider large-scale atmospheric motion. 2) STJPM's architecture is designed to address the spatiotemporal correlations of data, enabling coupled learning of complex spatiotemporal features. Com-GRU considers the real-time operating status of individual wind turbines, effectively correcting for unified prediction errors.

[0137] Example 2

[0138] This embodiment provides a wind power prediction system that combines wind turbine operating status with error correction, including:

[0139] The data acquisition module is configured to: acquire historical operating status data of the wind turbine and numerical weather forecast grid data;

[0140] The data preprocessing module is configured to: preprocess the numerical weather forecast grid data to obtain a wind turbine conversion power sequence;

[0141] The initial power prediction module is configured to combine the wind turbine historical power sequence, the wind turbine conversion power sequence, and the trained spatiotemporal joint prediction model to obtain a unified predicted power sequence for each wind turbine;

[0142] The error prediction module is configured to: add a separate error correction model for each wind turbine, combine the historical operating status data of the wind turbine, the unified predicted power sequence and the predicted meteorological factors at the corresponding location of each wind turbine, and obtain the error prediction sequence;

[0143] The error correction module is configured to correct the unified predicted power sequence using the error prediction sequence to obtain a corrected unified wind power prediction sequence.

[0144] Example 3

[0145] This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the steps of the wind power prediction method combining wind turbine operating status and error correction as described above are implemented.

[0146] Example 4

[0147] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the wind power prediction method combining wind turbine operating status and error correction as described above are implemented.

[0148] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage and optical storage, etc.) containing computer-usable program code.

[0149] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0150] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0151] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0152] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware through a computer program. The program can be stored in a computer-readable storage medium, and when executed, the program can include the processes in the above-described method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0153] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A wind power prediction method combining wind turbine operating status and error correction, characterized in that: The steps include: Obtain historical operating status data of wind turbines and numerical weather forecast grid data; Preprocess the numerical weather forecast grid data to obtain the wind turbine conversion power series; Combine the wind turbine historical power series, wind turbine conversion power series and the trained spatiotemporal joint prediction model to obtain a unified predicted power series for each wind turbine; A separate error correction model is added to each wind turbine, which combines the historical operating status data of the wind turbine, the unified predicted power sequence and the predicted meteorological factors at the corresponding location of each wind turbine to obtain the error prediction sequence; The error prediction sequence is used to correct the unified predicted power sequence to obtain the corrected unified wind power prediction sequence; The above method adds a separate error correction model to each wind turbine, combines the historical operating status data of the wind turbine, the unified predicted power sequence and the predicted meteorological factors at the corresponding location of each wind turbine, and obtains the error prediction sequence, including: The error correction model consists of two parallel GRU layers and a unified GRU layer. The input sequence of the EC model is based on the current moment t Divided into t -and t + variable, the fan historical operating status variable as t -variable, after normalization, enters the first GRU layer to obtain the first output feature; the unified predicted power sequence of each wind turbine and the predicted meteorological elements at the corresponding position of each wind turbine are used as t + variable, after normalization, enters the parallel second GRU layer to obtain the second output feature; The first output feature and the second output feature are concatenated according to the time series, enter the unified GRU layer, and output a single wind turbine error prediction sequence.

2. The wind power prediction method combining wind turbine operating status and error correction according to claim 1, characterized in that: The preprocessing of the numerical weather forecast grid data to obtain a wind turbine conversion power sequence includes: Based on the numerical weather forecast grid data, according to the geographical location of the selected wind farm, the grid forecast data near the wind turbine is extracted; Combined with the grid prediction data near the wind turbine, the GIDS method is used to interpolate the four NWP grid points around the wind turbine to the wind turbine hub position to obtain the wind speed prediction sequence at each wind turbine position; According to the wind speed-power conversion curve of the wind turbine, the wind speed prediction sequence at each wind turbine location is converted to obtain the power sequence at the corresponding moment.

3. The wind power prediction method combining wind turbine operating status and error correction according to claim 1, characterized in that: The method combines the wind turbine historical power sequence, the wind turbine conversion power sequence and the trained spatiotemporal joint prediction model to obtain a unified predicted power sequence for each wind turbine, including: The wind turbine historical power sequence and the wind turbine conversion power sequence are spliced ​​and input into the TCN feature extraction network. When using causal convolution for sequence feature extraction, the expansion factor is introduced and causal expansion convolution is used to extract features from the time series. After feature extraction is completed, the 2D ConvGRU network is entered to perform joint learning of spatiotemporal features to predict the unified power sequence of each wind turbine.

4. The wind power prediction method combining wind turbine operating status and error correction according to claim 3, characterized in that: When training the TCN feature extraction network, a residual block is introduced, and different layers of the TCN feature extraction network are connected by skip layers. A residual block includes two layers of expanded causal convolution, and batch normalization is performed after the convolution operation.

5. The wind power prediction method combining wind turbine operating status and error correction according to claim 3, characterized in that: When the 2D ConvGRU network performs joint learning of spatiotemporal features, the calculation method of a single GRU unit is changed from fully connected weights to convolutional feature extraction.

6. The wind power prediction method combining wind turbine operating status and error correction according to claim 1, characterized in that: The historical operating status data of the wind turbine includes historical power, generator speed, blade speed and generator operating temperature.

7. A wind power prediction system combining wind turbine operating status and error correction, characterized in that: include: The data acquisition module is configured to: acquire historical operating status data of the wind turbine and numerical weather forecast grid data; The data preprocessing module is configured to: preprocess the numerical weather forecast grid data to obtain a wind turbine conversion power sequence; The initial power prediction module is configured to combine the wind turbine historical power sequence, the wind turbine conversion power sequence, and the trained spatiotemporal joint prediction model to obtain a unified predicted power sequence for each wind turbine; The error prediction module is configured to add a separate error correction model to each wind turbine, combine the historical operating status data of the wind turbine, the unified predicted power sequence, and the predicted meteorological factors at the corresponding location of each wind turbine, and obtain the error prediction sequence; including: The error correction model consists of two parallel GRU layers and a unified GRU layer. The input sequence of the EC model is based on the current moment t Divided into t -and t + variable, the fan historical operating status variable as t -variable, after normalization, enters the first GRU layer to obtain the first output feature; the unified predicted power sequence of each wind turbine and the predicted meteorological elements at the corresponding position of each wind turbine are used as t + variable, after normalization, enters the parallel second GRU layer to obtain the second output feature; The first and second output features are concatenated in time series, fed into a unified GRU layer, and a single wind turbine error prediction sequence is output; The error correction module is configured to correct the unified predicted power sequence using the error prediction sequence to obtain a corrected unified wind power prediction sequence.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the wind power prediction method combining wind turbine operating status and error correction as described in any one of claims 1 to 6 are implemented.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the wind power prediction method combining wind turbine operating status and error correction are implemented as described in any one of claims 1 to 6.

Images