Wind power prediction method and system for eliminating error distribution offset

Abstract

The invention relates to the technical field of wind power prediction, in particular to a wind power prediction method and system for eliminating error distribution offset, and the method comprises the steps: S1, collecting historical wind power plant data; s2, preprocessing the data; s3, decomposing a historical wind power sequence into subsequences through a CEEMDAN algorithm; s4, constructing a prediction result sequence of each sub-sequence output by the gating circulation unit network; s5, calculating a prediction error sequence of each sub-sequence and fitting the prediction error sequence into a normal distribution probability density function; s6, calculating a standard deviation and a mean value, and constructing an initial neural network; s7, iterating the initial neural network and updating the weight and bias to obtain an adjusted neural network; and S8, obtaining pre-test set data, inputting a test set prediction result sequence into the adjustment type neural network for reconstruction, and outputting a final prediction sequence. According to the method, the error distribution deviation phenomenon in the wind power prediction process is eliminated, so that the prediction precision of the offshore wind power is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power prediction, and more specifically, to a wind power prediction method and system for eliminating error distribution offset. Background Art

[0002] With the continuous development of wind power generation technology, large-capacity wind power grid connection has brought challenges to the safe and stable operation of large power grids. Therefore, accurate prediction of wind power output can dispatch wind power grid connection capacity in advance and reduce the volatility of wind power grid connection power, which is of great significance to the safe and stable operation of the power system.

[0003] In the prior art, for example, a Chinese patent application with publication number CN114611808A discloses a short-term offshore wind power prediction method based on CEEMDAN-SSA-BiLSTM, including: using the CEEMDAN algorithm to decompose the historical offshore wind power sequence to obtain different subsequences; constructing a BiLSTM prediction model for each subsequence, and using the sparrow optimization algorithm to optimize the hyperparameters such as the number of hidden layer units, learning rate, and number of iterations of the BiLSTM model, using the optimized hyperparameters to predict different subsequences, and finally superimposing the prediction results of different subsequences to obtain the offshore wind power prediction. However, this invention only relies on the historical offshore wind power sequence for prediction. Due to the influence of meteorological characteristics such as wind speed and wind direction, the wind power prediction sequence will have the problem of error distribution offset in different environments, resulting in a decrease in prediction accuracy. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art that wind power prediction sequences may have error distribution offsets under different environments, and to provide a wind power prediction method and system for eliminating error distribution offsets, thereby eliminating the error distribution offset phenomenon in the wind power prediction process, thereby improving the prediction accuracy of offshore wind power.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: A wind power prediction method for eliminating error distribution offset is provided, comprising: S1. Collect historical wind power data and meteorological characteristic data of historical wind farms; S2. Preprocessing the historical wind power data and meteorological characteristic data to obtain a historical wind power sequence and a meteorological characteristic sequence; S3, inputting the historical wind power sequence into the CEEMDAN algorithm, and the CEEMDAN algorithm outputs a plurality of relatively independent wind power subsequences; S4, constructing a gated recurrent unit network, inputting the wind power subsequence and the meteorological feature sequence into the gated recurrent unit network for training, and the gated recurrent unit network outputs a prediction result sequence of each subsequence; S5. Calculate the prediction error sequence of each subsequence according to the prediction result sequence of each subsequence and the historical wind power data, and fit the prediction error sequence of each subsequence to a normal distribution probability density function; S6. Calculate the standard deviation of the normal distribution probability density function and mean , the standard deviation With the mean As the initial weights and initial biases of the neural network, an initial neural network is constructed; S7, inputting the prediction error sequence of each subsequence into the initial neural network for iteration, calculating the total error loss function in the iteration process of the initial neural network, and updating the weights and biases of the neurons in the initial neural network according to the total error loss function, and obtaining the adjusted neural network after the iteration is completed; S8. Obtain the predicted wind power data and meteorological characteristic data of the wind farm as test set data, and execute steps S2-S4 on the test set data to obtain a test set prediction result sequence, and input the test set prediction result sequence into the adjustment neural network described in step S7. The adjustment neural network reconstructs the test set prediction result sequence to obtain a final prediction sequence.

[0006] The present invention first collects historical wind power data and meteorological characteristic data of historical wind farms, then pre-processes the data to obtain historical wind power sequences and meteorological characteristic sequences, then uses the CEEMDAN algorithm, that is, fully integrates empirical mode decomposition and adaptive noise algorithm, to decompose the historical wind power sequence into a plurality of independent wind power subsequences, and inputs the obtained wind power subsequences and meteorological characteristic sequences into a gated recurrent unit network for training, thereby obtaining prediction result sequences of each subsequence; then calculates a prediction error sequence according to the prediction result sequence of each subsequence and the true value of power in the historical wind power data, fits the obtained prediction error sequence to a normal distribution probability density function, and inputs the prediction error sequence into a constructed initial neural network for iteration, and calculates a total error loss function in the iteration process; the iteration process continues for multiple times, and each iteration needs to update the weight and bias of the initial neural network according to the loss function, and in the process of dynamically adjusting the error distribution, the normal distribution of the prediction error sequence of each subsequence tends to be symmetrical, thereby reducing the distribution deviation of the prediction error, and obtaining an adjusted neural network after multiple iterations. When actual prediction is needed, the test set data is collected, and after preprocessing, CEEMDAN algorithm decomposition, and gated recurrent unit network prediction, the obtained test set prediction result sequence is input into the adjustment neural network for reconstruction to obtain the final prediction sequence. The final prediction sequence eliminates the error distribution offset phenomenon in the wind power prediction process, thereby improving the prediction accuracy of offshore wind power.

[0007] Preferably, in step S5, the method for calculating the prediction error sequence of each subsequence is:

[0008] In the formula, For the The prediction error sequence of subsequences, For the The predicted value of the subsequence, For the The predicted value comes from the prediction result sequence of each subsequence output by the gated recurrent unit network, and the true value comes from the historical wind power data. The difference between the predicted value of the subsequence and the true value of the subsequence is the prediction error value of the subsequence. The prediction error value of each subsequence is calculated to pave the way for the subsequent fitting to the normal distribution curve.

[0009] Preferably, in step S6, each of the subsequences corresponds to a separate neural network-like unit, and the standard deviation of the prediction error sequence of each subsequence is and mean As the initial weight of the neural network unit and initial bias . Standard Deviation and mean It fully reflects the distribution characteristics of the prediction error sequence of each subsequence. Taking this as the initial weight and bias can make the neural network closer to the optimal solution in the early stage of training. Due to the different error statistical characteristics of each subsequence, based on the standard deviation and mean The initialization method can specifically adjust the weights and biases of each neural network-like unit to better meet the characteristics of each subsequence, thereby improving the overall prediction accuracy.

[0010] Preferably, the mathematical expression of the neural network-like unit is:

[0011] In the formula, is a neural network-like unit, is the prediction error sequence of each subsequence. The weights in the neural network-like unit are used to adjust the importance of the input features, determine the transmission and transformation process of the signal, scale the input prediction error sequence, and change the shape of the normal distribution curve. The bias is used to adjust the activation threshold of the neuron and enhance the flexibility of the model, thereby changing the overall position of the normal distribution curve. The weights and biases work together to enable the neural network to fit complex nonlinear relationships, and continuously optimize through training to ultimately achieve accurate predictions.

[0012] Preferably, the initial neural network-like unit comprises a plurality of groups of neural network-like units:

[0013] In the formula, is the initial class neural network, It is a neural network-like unit. Multiple groups of neural network-like units can capture the features of different levels and scales in the input error sequence feature data, thereby improving the expression ability of the initial neural network-like unit and increasing the prediction accuracy.

[0014] Preferably, in step S7, the total error loss function is obtained by adding all sub-error loss functions, and the sub-error loss function is obtained by calculating the standard deviation of the prediction error sequence of each sub-sequence. and mean This error loss function calculation method assigns an independent error loss function to each subsequence. By analyzing the loss function value of each subsequence, it can identify which subsequences contribute more to the overall prediction error, thereby improving the model in a targeted manner, so that the prediction model can more accurately optimize the prediction results of each subsequence, thereby improving the overall prediction accuracy.

[0015] Preferably, the sub-error loss function is calculated as follows:

[0016] In the formula, is the sub-error loss function, For the The mean value of the subsequences. Standard deviation and mean It has clear statistical significance and can intuitively reflect the distribution characteristics of the prediction errors of each subsequence. Calculating the loss function based on these two characteristic parameters can enhance the interpretability of the model optimization process. During the optimization process, the weight of each subsequence in the total loss function can be dynamically adjusted. For subsequences with large error fluctuations or large systematic deviations, the loss function will assign higher weights, thereby balancing the contribution of each subsequence.

[0017] Preferably, the weights and biases are updated in the iterative process as follows:

[0018]

[0019]

[0020] in For the The subsequence is The weight at the iteration, For the The subsequence is The weight at the iteration, For the The subsequence is The bias at the iteration, For the The subsequence is The bias at the iteration, and are the learning rates of the neural network units described above, For the The total error loss function difference at the iteration, For the The total error loss function value at the iteration, For the The total error loss function at the iteration. By continuously updating weights and biases, the model neural network can better fit the training data, capture the complex patterns and relationships in the data, and thus improve accuracy. The loss function is used to measure the difference between the model prediction results and the actual results. The process of iteratively updating weights and biases is the process of minimizing the loss function. As the iteration proceeds, the normal distribution curves of the error sequences of each subsequence tend to be symmetrical, and the prediction accuracy gradually improves.

[0021] Preferably, in step S8, the reconstruction process of the test set prediction result sequence is:

[0022]

[0023]

[0024] in For the The test set predicts the result sequence, For the The subsequences are used to adjust the bias of the neural network. For the The weights of the subsequences in the neural network are adjusted. To adjust the first A neural network-like unit, To adjust the neural network, is the final prediction sequence, For the After multiple iterations of updating and optimization, the adjustment neural network makes the normal distribution curve of each subsequence error sequence tend to be symmetrical, eliminating the phenomenon of error distribution deviation. It only needs to input the test set prediction result sequence into the adjustment neural network, and the adjustment neural network reconstructs the test set subsequence according to the weights and biases generated by the last iteration to obtain the final prediction sequence.

[0025] The present invention also provides a wind power prediction system, which is used to implement the wind power prediction method according to any one of claims 1 to 9, and is characterized by comprising: Data acquisition module: used to acquire power characteristic data and meteorological characteristic data of historical wind farms and predicted wind farms, and transmit the power characteristic data and the meteorological characteristic data to the data processing module; Data processing module: used for preprocessing the power characteristic data and the meteorological characteristic data and performing CEEMDAN algorithm decomposition to obtain a wind power subsequence; A gated recurrent unit network prediction module: based on the input wind power subsequence and the meteorological feature sequence, outputs a prediction result sequence of the subsequence; Neural network prediction module: used to fit the normal distribution probability density curve to construct the initial neural network, and iterate based on the prediction error sequence of the input subsequence. During the iteration, the initial neural network is updated and the output is the adjustment neural network. The power characteristic data and meteorological characteristic data of the predicted wind farm are used as the input of the adjustment neural network, and the adjustment neural network predicts and outputs the final prediction sequence.

[0026] The wind power prediction system of the present invention first collects power characteristic data and meteorological characteristic data of historical wind farms through a data acquisition module, then the data processing module pre-processes the power characteristic data and meteorological characteristic data, and uses the CEEMDAN algorithm to decompose the pre-processed power characteristic data to obtain a wind power subsequence, and the gated recurrent unit network prediction module inputs the wind power subsequence and the meteorological characteristic sequence into the gated recurrent unit network, and finally obtains the prediction result of the subsequence, and the neural network prediction module fits the normal distribution probability density curve according to the error result of the subsequence and the power characteristic data, and iterates based on the prediction error sequence of the input subsequence, updates the initial neural network during the iteration process, outputs the adjustment neural network, and then the power characteristic data and meteorological characteristic data of the predicted wind farm are collected by the data acquisition module, and after passing through the data processing module and the gated recurrent unit network prediction module in turn, they are input into the adjustment neural network in the neural network unit prediction module, and the adjustment neural network reconstructs the prediction result sequence, adjusts the prediction error distribution of different subsequences, eliminates the error distribution offset phenomenon in the wind power prediction process, and thus improves the prediction accuracy of offshore wind power.

[0027] Compared with the prior art, the present invention has the following beneficial effects: 1. By fitting the normal distribution probability density function and combining it with a neural network iteration method, the error distribution deviation phenomenon in the wind power prediction process is eliminated, thereby improving the prediction accuracy of offshore wind power; 2. Realize the decomposition prediction of wind power, adjust the prediction error distribution of different subsequences, and reconstruct the prediction results to keep the error distribution function symmetrical. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 A flow chart of a wind power prediction method for eliminating error distribution offset; Figure 2 Schematic diagram of the effect of bias on error distribution offset; Figure 3 Schematic diagram of the effect of weight on error distribution offset; Figure 4 It is the prediction error curve of the common prediction method; Figure 5 is a prediction error curve diagram of the wind power prediction method of the present invention; Figure 6 It is the probability distribution diagram of prediction error of common prediction methods; Figure 7 It is a prediction error probability distribution diagram of the wind power prediction method of the present invention. DETAILED DESCRIPTION

[0029] The present invention will be further described below in conjunction with specific implementation modes.

[0030] Embodiment 1 This embodiment is a first embodiment of a wind power prediction method for eliminating error distribution offset, comprising: S1. Collect historical wind power data and meteorological characteristic data of historical wind farms; S2. Preprocess the historical wind power data and meteorological characteristic data to obtain a historical wind power sequence and a meteorological characteristic sequence; S3, inputting the historical wind power sequence into the CEEMDAN algorithm, and the CEEMDAN algorithm outputs several relatively independent wind power subsequences; S4, constructing a gated recurrent unit network, inputting the wind power subsequence and the meteorological feature sequence into the gated recurrent unit network for training, and the gated recurrent unit network outputs a prediction result sequence of each subsequence; S5. Calculate the prediction error sequence of each subsequence according to the prediction result sequence of each subsequence and the historical wind power data, and fit the prediction error sequence of each subsequence to a normal distribution probability density function; S6. Calculate the standard deviation of the normal distribution probability density function and mean , the standard deviation With the mean As the initial weights and initial biases of the neural network, an initial neural network is constructed; S7, inputting the prediction error sequence of each subsequence into the initial neural network for iteration, calculating the total error loss function in the iterative process of the initial neural network, and updating the weights and biases of the neurons in the initial neural network according to the total error loss function, and obtaining the adjusted neural network after the iteration is completed; S8. Obtain the predicted wind power data and meteorological characteristic data of the wind farm as the test set data, and execute steps S2-S4 on the test set data to obtain the test set prediction result sequence, and input the test set prediction result sequence into the adjustment class neural network in step S7. The adjustment class neural network reconstructs the test set prediction result sequence to obtain the final prediction sequence.

[0031] like Figure 1As shown, the wind power prediction method in this embodiment first collects historical wind power data and meteorological characteristic data of historical wind farms, then preprocesses the data to obtain historical wind power sequences and meteorological characteristic sequences, and then uses the CEEMDAN algorithm to decompose the historical wind power sequences into a number of independent wind power subsequences, and inputs the obtained wind power subsequences and meteorological characteristic sequences into the gated recurrent unit network for training, thereby obtaining prediction result sequences of each subsequence; then, the prediction error sequence is calculated according to the prediction result sequence of each subsequence and the true value of the power in the historical wind power data, the obtained prediction error sequence is fitted to a normal distribution probability density function, and the prediction error sequence is input into the constructed initial neural network for iteration, and the total error loss function in the iterative process is calculated; the iterative process continues for multiple times, and each iteration needs to update the weights and biases of the initial neural network according to the loss function, and in the process of dynamically adjusting the error distribution, the normal distribution of the prediction error sequence of each subsequence tends to be symmetrical, thereby reducing the distribution offset of the prediction error, and obtaining the adjusted neural network after multiple iterations. When actual prediction is needed, the test set data is collected, and after preprocessing, CEEMDAN algorithm decomposition, and gated recurrent unit network prediction, the obtained test set prediction result sequence is input into the adjustment neural network for reconstruction to obtain the final prediction sequence. The final prediction sequence eliminates the error distribution offset phenomenon in the wind power prediction process, thereby improving the prediction accuracy of offshore wind power.

[0032] In step S2 of this embodiment, the historical wind power data and meteorological characteristic data are preprocessed, including the power sequence and the wind speed sequence, temperature, and air pressure sequence at different heights are processed by min-max normalization to obtain the processed power sequence P, wind speed sequence WS, temperature sequence Tem, and air pressure sequence Pre. In step S3, the power sequence P, wind speed sequence WS, temperature sequence Tem, and air pressure sequence Pre are input into the gated recurrent unit for training.

[0033] In step S3, the CEEMDAN algorithm decomposes the historical wind power sequence into wind power subsequences, and the decomposition method is:

[0034] in The decomposed Wind power subsequences, is the historical wind power sequence, and N represents the added noise intensity. In wind power forecasting, CEEMDAN can effectively separate signal components of different time scales by adding adaptive noise multiple times and integrating and averaging, significantly reducing the modal aliasing phenomenon. The decomposed IMF subsequence is purer and suitable for processing nonlinear and non-stationary signals. It can significantly improve the decomposition quality and provide a higher quality data foundation for subsequent modeling and forecasting.

[0035] In step S5, the prediction error sequence of each subsequence is calculated as follows:

[0036] in For the The prediction error sequence of subsequences, For the The predicted value of the subsequence, For the The difference between the predicted value of a subsequence and the true value of the subsequence is the prediction error value of the subsequence. The prediction error value of each subsequence is calculated to pave the way for the subsequent fitting into a normal distribution curve.

[0037] In step S6, each subsequence corresponds to a separate neural network-like unit, and the standard deviation of the prediction error sequence of each subsequence is and mean As the initial weights of the neural network-like unit and initial bias . Standard Deviation and mean It fully reflects the distribution characteristics of the prediction error sequence of each subsequence. Taking this as the initial weight and bias can make the neural network closer to the optimal solution in the early stage of training. Due to the different error statistical characteristics of each subsequence, based on the standard deviation and mean The initialization method can specifically adjust the weights and biases of each neural network-like unit to better meet the characteristics of each subsequence, thereby improving the overall prediction accuracy.

[0038] The mathematical expression of the neural network-like unit in this embodiment is:

[0039] in is a neural network-like unit, is the prediction error sequence of each subsequence. The weights in the neural network-like unit are used to adjust the importance of the input features, determine the transmission and transformation process of the signal, scale the input prediction error sequence, and change the shape of the normal distribution curve. The bias is used to adjust the activation threshold of the neuron and enhance the flexibility of the model, thereby changing the overall position of the normal distribution curve. The weights and biases work together to enable the neural network to fit complex nonlinear relationships, and continuously optimize through training to ultimately achieve accurate predictions.

[0040] The initial class neural network includes multiple groups of class neural network units:

[0041] in is the initial class neural network, It is a neural network-like unit. Multiple groups of neural network-like units can capture the features of different levels and scales in the input error sequence feature data, thereby improving the expression ability of the initial neural network-like unit and increasing the prediction accuracy.

[0042] In this embodiment, step S5 fits the prediction error sequence of each subsequence to a normal distribution probability density function, and the mathematical expression is as follows:

[0043]

[0044]

[0045] In the formula, is the standard deviation of the forecast error series, is the mean of the forecast error series, Indicates the sample size.

[0046] The gated recurrent unit network in this embodiment is divided into two layers, with 4 and 8 neurons respectively, and the mathematical expression is:

[0047] In the formula, , , is the weight matrix, corresponding to the update gate , Reset Gate and candidate activation The input part, , , For the current input The corresponding weight matrices are used to calculate the update gate , reset gate and candidate activation The hidden state of , , is the bias parameter matrix, corresponding to the update gate , reset gate and candidate activation The weight bias of is matrix multiplication, is the Sigmod function, To reset the gate, To update the gate, is the candidate state of the hidden layer at the current moment, is the current implicit state, is the implicit state at the previous moment, The input status at the current moment.

[0048] Embodiment 2 This embodiment is a second embodiment of a wind power prediction method for eliminating error distribution offset. This embodiment is similar to the first embodiment, except that in step S7, the total error loss function is obtained by adding all sub-error loss functions, and the sub-error loss function is obtained by calculating the standard deviation of the prediction error sequence of each sub-sequence. and mean Calculated.

[0049] The sub-error loss function is calculated as:

[0050] in is the sub-error loss function, For the The mean value of the subsequences. Standard deviation and mean It has clear statistical significance and can intuitively reflect the distribution characteristics of the prediction errors of each subsequence. Calculating the loss function based on these two characteristic parameters can enhance the interpretability of the model optimization process. During the optimization process, the weight of each subsequence in the total loss function can be dynamically adjusted. For subsequences with large error fluctuations or large systematic deviations, the loss function will assign higher weights, thereby balancing the contribution of each subsequence.

[0051] In step S7, the weights and biases are updated in multiple iterations as follows:

[0052]

[0053]

[0054] in For the The weight of a subsequence at the t+1th iteration, For the The weight of a subsequence at the tth iteration, For the The bias of the subsequence at the t+1th iteration, For the The bias of the subsequence at the tth iteration, and are the learning rates of neural network-like units, is the total error loss function difference at the t+1th iteration, is the total error loss function value at the t+1th iteration, is the total error loss function at the tth iteration. By continuously updating weights and biases, the model neural network can better fit the training data, capture the complex patterns and relationships in the data, and thus improve accuracy. The loss function is used to measure the difference between the model prediction results and the actual results. The process of iteratively updating weights and biases is the process of minimizing the loss function. Figure 2 The effect of bias on the error distribution shift is shown, Figure 3 The influence of weights on the error distribution offset is shown, where the solid line on the left is the error distribution curve at the beginning of the iteration, the solid line on the right is the error distribution curve after the iteration ends, and the dotted line on the right is the error distribution curve at the beginning of the iteration. It can be seen that as the iteration proceeds, the normal distribution curves of the error sequences of each subsequence tend to be symmetrical, and the prediction accuracy gradually improves.

[0055] Embodiment 3 This embodiment is a third embodiment of a wind power prediction method for eliminating error distribution offset. This embodiment is similar to the first embodiment, except that in step S8, the reconstruction process of the test set prediction result sequence is:

[0056]

[0057]

[0058] in For the The test set predicts the result sequence, For the The subsequences are used to adjust the bias of the neural network. For the The weights of the subsequences in the neural network are adjusted. To adjust the first A neural network-like unit, To adjust the neural network, is the final prediction sequence, For the After multiple iterations of updating and optimization, the adjustment neural network makes the normal distribution curve of each subsequence error sequence tend to be symmetrical, eliminating the phenomenon of error distribution deviation. It only needs to input the test set prediction result sequence into the adjustment neural network, and the adjustment neural network reconstructs the test set subsequence according to the weights and biases generated by the last iteration to obtain the final prediction sequence.

[0059] like Figure 4 , Figure 5 As shown, when the common prediction method is used to predict wind power, the error mean is -0.9874 when there is an error distribution offset. When the method in this embodiment is used to predict, the error mean is reduced to -0.0087 after considering the error distribution offset. Figure 6 , Figure 7 As shown, the prediction error value obtained by the common prediction method and the prediction error value obtained by the prediction method in this embodiment are both fitted into a normal distribution curve, the horizontal axis represents the error value, and the vertical axis represents the probability density of the error value. The normal distribution curve obtained by the prediction method in this embodiment is more symmetrical, the distribution is more uniform, and the error offset value is significantly lower. The wind power prediction method in this embodiment takes into account the impact of the error distribution offset on the wind power prediction result. By fitting to a normal distribution probability density function combined with a neural network iteration method, the error distribution offset phenomenon in the wind power prediction process is eliminated, and the wind power prediction error value is reduced, thereby improving the prediction accuracy of offshore wind power.

[0060] Embodiment 4 This embodiment is the first embodiment of the wind power prediction system, comprising: a data acquisition module: used to acquire power characteristic data and meteorological characteristic data of historical wind farms and predicted wind farms, and transmit the power characteristic data and meteorological characteristic data to a data processing module; Data processing module: used to pre-process power characteristic data and meteorological characteristic data and decompose them using CEEMDAN algorithm to obtain wind power subsequences; Gated recurrent unit network prediction module: Based on the input wind power subsequence and meteorological feature sequence, it outputs the prediction result sequence of the subsequence; Neural network prediction module: used to fit the normal distribution probability density curve to construct the initial neural network, and iterate based on the prediction error sequence of the input subsequence. During the iteration, the initial neural network is updated and the output is the adjustment neural network. The power characteristic data and meteorological characteristic data of the predicted wind farm are used as the input of the adjustment neural network, and the adjustment neural network predicts and outputs the final prediction sequence.

[0061] The wind power prediction system in this embodiment first collects power characteristic data and meteorological characteristic data of historical wind farms through the data acquisition module, and then the data processing module preprocesses the power characteristic data and meteorological characteristic data, and uses the CEEMDAN algorithm to decompose the preprocessed power characteristic data to obtain a wind power subsequence. The gated recurrent unit network prediction module inputs the wind power subsequence and the meteorological characteristic sequence into the gated recurrent unit network, and finally obtains the prediction result of the subsequence. The neural network prediction module fits the normal distribution probability density curve according to the error result of the subsequence and the power characteristic data, and iterates based on the prediction error sequence of the input subsequence. During the iteration, the initial neural network is updated and the adjustment neural network is output. The power characteristic data and meteorological characteristic data of the predicted wind farm are collected by the data acquisition module, and after passing through the data processing module and the gated recurrent unit network prediction module in turn, they are input into the adjustment neural network in the neural network unit prediction module. The adjustment neural network reconstructs the prediction result sequence, adjusts the prediction error distribution of different subsequences, eliminates the error distribution offset phenomenon in the wind power prediction process, and thus improves the prediction accuracy of offshore wind power.

[0062] In the specific contents of the above-mentioned specific implementation methods, the various technical features can be combined in any non-contradictory manner. In order to make the description concise, not all possible combinations of the above-mentioned technical features are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0063] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A wind power prediction method for eliminating error distribution offset, characterized in that: include: S1. Collect historical wind power data and meteorological characteristic data of historical wind farms; S2. Preprocessing the historical wind power data and meteorological characteristic data to obtain a historical wind power sequence and a meteorological characteristic sequence; S3, inputting the historical wind power sequence into the CEEMDAN algorithm, and the CEEMDAN algorithm outputs a plurality of relatively independent wind power subsequences; S4, constructing a gated recurrent unit network, inputting the wind power subsequence and the meteorological feature sequence into the gated recurrent unit network for training, and the gated recurrent unit network outputs a prediction result sequence of each subsequence; S5. Calculate the prediction error sequence of each subsequence according to the prediction result sequence of each subsequence and the historical wind power data, and fit the prediction error sequence of each subsequence to a normal distribution probability density function; S6. Calculate the standard deviation of the normal distribution probability density function and mean , the standard deviation With the mean As the initial weights and initial biases of the neural network, an initial neural network is constructed; S7, inputting the prediction error sequence of each subsequence into the initial neural network for iteration, calculating the total error loss function in the iteration process of the initial neural network, and updating the weights and biases of the neurons in the initial neural network according to the total error loss function, and obtaining the adjusted neural network after the iteration is completed; S8. Obtain the predicted wind power data and meteorological characteristic data of the wind farm as test set data, and execute steps S2-S4 on the test set data to obtain a test set prediction result sequence, and input the test set prediction result sequence into the adjustment neural network described in step S7. The adjustment neural network reconstructs the test set prediction result sequence to obtain a final prediction sequence.

2. The method for wind power prediction for eliminating error distribution offset according to claim 1, characterized in that: In step S5, the prediction error sequence of each subsequence is calculated as follows: In the formula, For the The prediction error sequence of subsequences, For the The predicted value of the subsequence, For the The true value of a subsequence.

3. The method for wind power prediction for eliminating error distribution offset according to claim 1, characterized in that: In step S6, each subsequence corresponds to a separate neural network-like unit, and the standard deviation of the prediction error sequence of each subsequence is and mean As the initial weight of the neural network unit and initial bias .

4. The method for wind power prediction for eliminating error distribution offset according to claim 3, characterized in that: The mathematical expression of the neural network-like unit is: In the formula, is a neural network-like unit, is the prediction error sequence of each subsequence.

5. The method for wind power prediction for eliminating error distribution offset according to claim 4, characterized in that: The initial neural network-like unit comprises a plurality of groups of neural network-like units: In the formula, is the initial class neural network, A neural network-like unit.

6. The method for wind power prediction for eliminating error distribution offset according to claim 1, characterized in that: In step S7, the total error loss function is obtained by adding all sub-error loss functions, and the sub-error loss function is obtained by calculating the standard deviation of the prediction error sequence of each sub-sequence. and mean Calculated.

7. The method for wind power prediction for eliminating error distribution offset according to claim 6, characterized in that: The sub-error loss function is calculated as follows: In the formula, is the sub-error loss function, For the The average value of the subsequences.

8. The method for wind power prediction for eliminating error distribution offset according to claim 7, characterized in that: The weights and biases are updated in the iterative process as follows: In the formula, For the The subsequence is The weight at the iteration, For the The subsequence is The weight at the iteration, For the The subsequence is The bias at the iteration, For the The subsequence is The bias at the iteration, and are the learning rates of the neural network units described above, For the The total error loss function difference at the iteration, For the The total error loss function value at the iteration, For the The total error loss function at iterations.

9. The method for wind power prediction for eliminating error distribution offset according to claim 1, characterized in that: In step S8, the reconstruction process of the test set prediction result sequence is: In the formula, For the The test set predicts the result sequence, For the The subsequences are used to adjust the bias of the neural network. For the The weights of the subsequences in the neural network are adjusted. To adjust the first A neural network-like unit, To adjust the neural network, is the final prediction sequence, For the A test set of subsequences.

10. A wind power prediction system, used to implement the wind power prediction method according to any one of claims 1 to 9, characterized in that: include: Data acquisition module: used to acquire power characteristic data and meteorological characteristic data of historical wind farms and predicted wind farms, and transmit the power characteristic data and the meteorological characteristic data to the data processing module; Data processing module: used for preprocessing the power characteristic data and the meteorological characteristic data and performing CEEMDAN algorithm decomposition to obtain a wind power subsequence; A gated recurrent unit network prediction module: based on the input wind power subsequence and the meteorological feature sequence, outputs a prediction result sequence of the subsequence; Neural network prediction module: used to fit the normal distribution probability density curve to construct an initial neural network, and iterate based on the prediction error sequence of the input subsequence, update the initial neural network during the iteration process, and output an adjustment neural network; predict the power characteristic data and meteorological characteristic data of the wind farm as the input of the adjustment neural network, and the adjustment neural network predicts and outputs the final prediction sequence.

Images