A wind speed short-term multi-step prediction method fusing direct prediction and recursive prediction

Abstract

The application discloses a wind speed short-term multi-step prediction method combining direct prediction and recursive prediction, and relates to the technical field of wind speed short-term multi-step prediction. The method comprises the following steps: obtaining an original data set of a wind speed short-term multi-step prediction task, and performing normalization processing to obtain a training data set; constructing a multi-channel LSTM encoding layer based on a global encoder and a single-step encoder; inputting the training data set into the multi-channel LSTM encoding layer for encoding to obtain an encoding state feature; constructing a double-path decoder, and inputting the encoding state feature into the double-path decoder for global recursive decoding and single-step direct decoding to obtain a global decoding feature and a single-step decoding feature; splicing and fusing the global decoding feature and the single-step decoding feature to obtain a fusion feature; and training and predicting the fusion feature based on a residual network and a padding strategy to obtain a multi-step prediction result. By using the application, the prediction accuracy can be improved while reducing the error accumulation during prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of short-term multi-step wind speed prediction, and in particular to a short-term multi-step wind speed prediction method combining direct prediction and recursive prediction. BACKGROUND

[0002] Short-term multi-step wind speed prediction refers to learning a historical window of data to predict the changes in wind speed in a future period. The historical window of data includes wind speed values at historical time points and other related variables, such as temperature, air pressure, humidity, rainfall, and other factors. Short-term multi-step wind speed prediction faces many challenges, both external reasons such as the randomness and uncertainty of wind speed, changes in weather conditions, and the influence of terrain, and internal technical difficulties such as considering the information correlation between multiple prediction time points and reducing the error propagation of multi-step prediction. In addition, since the short-term changes in wind speed are usually small, the prediction accuracy requirement is high, and such prediction is of great significance to the operation and management of wind farms, the scheduling of power systems, and the trading of energy markets. Therefore, it is necessary to continuously improve and optimize the prediction method to meet the relevant requirements.

[0003] In traditional methods, wind speed prediction is mainly based on physical models and statistical models. Physical models are represented by CFD models and Kalman filtering methods. Although these methods can consider the influence of various physical factors, they require more time to correct wind speed, so they are more suitable for long-term prediction and less suitable for short-term prediction. Statistical models are represented by ARIMA models. The statistical model can only capture linear relationships and may not accurately capture nonlinear trends and changes in wind speed, resulting in a decrease in prediction accuracy.

[0004] Compared with traditional methods, machine learning and deep learning models have better universality and flexibility in wind speed prediction. In the model training phase, different meteorological conditions and time periods can be learned and optimized accordingly. In the prediction phase, existing prediction schemes can be divided into direct prediction and recursive prediction. Direct prediction does not consider the correlation between prediction time points and cannot accurately predict the trend of changes. Recursive prediction is prone to error accumulation, and the accuracy of subsequent step prediction is not high. Therefore, the existing technology cannot balance the improvement of prediction accuracy and the reduction of error accumulation during prediction. SUMMARY

[0005] To solve the above technical problems, the purpose of the present application is to provide a short-term multi-step wind speed prediction method combining direct prediction and recursive prediction, which combines the advantages of recursive prediction and direct prediction to improve prediction accuracy while reducing error accumulation during prediction.

[0006] The technical solution adopted by the present application is: a short-term multi-step wind speed prediction method combining direct prediction and recursive prediction, comprising the following steps:

[0007] An original data set of a short-term multi-step wind speed prediction task is obtained, and normalization processing is performed to obtain a training data set;

[0008] A multi-channel LSTM encoding layer is constructed based on a global encoder and a single-step encoder;

[0009] The training data set is input into the multi-channel LSTM encoding layer for encoding to obtain an encoding state feature;

[0010] A double-path decoder is constructed, and the encoding state feature is input into the double-path decoder for global recursive decoding and single-step direct decoding to obtain a global decoding feature and a single-step decoding feature;

[0011] The global decoding feature and the single-step decoding feature are spliced and fused to obtain a fused feature;

[0012] The fused feature is trained and predicted based on a residual network and a padding strategy to obtain a multi-step prediction result.

[0013] Further, the multi-channel LSTM encoding layer is composed of a first LSTM channel and other LSTM channels, wherein:

[0014] The first LSTM channel uses a global encoder;

[0015] The other LSTM channels use a single-step encoder.

[0016] Further, the encoding state feature has the following expression:

[0017]

[0018] wherein, x t represents the t-th sample in the training data set; represents the first LSTM channel; represents the global encoding state; represents the n-th LSTM channel; represents the single-step encoding state of the n-th LSTM channel.

[0019] Further, the global decoding feature has the following expression:

[0020]

[0021] wherein, represents the global encoding state; represents the first LSTM channel decoder; represents the global recursive decoding feature of the n-th step.

[0022] Further, the single-step decoding feature, the expression is as follows:

[0023]

[0024] Wherein, Indicates the nth LSTM channel decoder; Indicates the single-step encoding state of the nth LSTM channel; Indicates the nth single-step decoding feature.

[0025] Further, the fusion feature, the expression is as follows:

[0026]

[0027] Wherein, Indicates the global recursive decoding feature of the nth step; Indicates the nth single-step decoding feature; Indicates the nth fusion feature.

[0028] Further, the padding strategy is to fill the left boundary value on the left side of the fusion feature sequence, so as to ensure that the sequence length is unchanged.

[0029] The method has the advantages that: the multi-channel LSTM encoding and decoding network is designed, the advantages of recursive prediction and direct prediction are fused in the encoding and decoding process, more comprehensive hidden information can be extracted; the final prediction result is converted through the residual network, the information correlation during prediction is fully considered, the prediction accuracy is further improved and the error accumulation during prediction is reduced; The strategy of filling the left boundary value maintains the sequence length, avoids interference caused by learning sequence information; ultimately, the prediction accuracy is improved while the error accumulation during prediction is reduced. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 is a step flow chart of a wind speed short-term multi-step prediction method fusing direct prediction and recursive prediction of the application;

[0031] Figure 2 is a training and prediction schematic diagram of a wind speed short-term multi-step prediction method fusing direct prediction and recursive prediction of the application;

[0032] Figure 3 is an LSTM-based encoding and decoding network flow chart of a wind speed short-term multi-step prediction method fusing direct prediction and recursive prediction of the application;

[0033] Figure 4 is a residual network prediction flow chart of a wind speed short-term multi-step prediction method fusing direct prediction and recursive prediction of the application. DETAILED DESCRIPTION

[0034] The application will be further described in detail below in combination with the drawings and specific embodiments. For the step numbers in the following embodiments, they are only set for the convenience of description, and the order between the steps is not limited in any way, and the execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0035] With reference to Figure 1 and Figure 2 , the application provides a wind speed short-term multi-step prediction method combining direct prediction and recursive prediction, which comprises the following steps:

[0036] S1, obtaining an original data set of a wind speed short-term multi-step prediction task and performing normalization processing to obtain a training data set; specifically, the original data set of the wind speed short-term multi-step prediction task is normalized, and the expression is as follows:

[0037]

[0038] Wherein, x represents the original data set; x min represents the minimum value in the original data set; x max represents the maximum value in the original data set; x n represents the result after normalization.

[0039] Next, according to the maximum step length n required for prediction and the historical window length t for prediction, sliding is performed to divide the training set and the test set in an 8:2 manner. For example, [x1, x2…x t , x t+1 , x t+2 …x t+n ] as a training sample, [x2, x3…x t+1 , x t+2 , x t+3 …x t+n+1 ] as the next training sample.

[0040] S2, constructing a multi-channel LSTM encoding layer based on a global encoder and a single-step encoder;

[0041] Specifically, the multi-channel LSTM encoding layer is composed of a first LSTM channel and other LSTM channels, wherein:

[0042] The first LSTM channel adopts a global encoder;

[0043] The other LSTM channels adopt a single-step encoder.

[0044] S3, inputting the training data set into the multi-channel LSTM encoding layer for encoding to obtain an encoding state feature;

[0045] Specifically, with reference toFigure 3 The input training data set is encoded using a multi-channel LSTM encoding layer to fully mine the hidden information of the sequence. t The input training data set sequence (x1, x2…xt) is input into the first LSTM channel, and the output at the last time of the LSTM is taken as the encoding result of the first LSTM channel.

[0046]

[0047] Wherein, x t represents the t-th sample in the training data set; represents the first LSTM channel; represents the global encoding state.

[0048] The present application opens a new channel for the second step (the first step is global encoding, and the second step is single-step encoding of the second LSTM channel) and the wind speed prediction thereafter to directly learn the prediction information, and separately encodes and decodes to obtain the corresponding single-step encoding state. In order to maintain the same time interval between the historical data and the prediction time, the input training data set sequence needs to be appropriately cropped. t-2i …x t-2 ,x t ); The new sequence is input into the second LSTM channel network, and the single-step encoding state of the second-step prediction is output. Similarly, the subsequent LSTM channel network crops the corresponding historical data input according to the predicted time step, and outputs the corresponding encoding state.

[0049]

[0050] Wherein, x t represents the t-th sample in the training data set; represents the n-th LSTM channel; represents the single-step encoding state of the n-th LSTM channel.

[0051] In order to effectively learn the historical data information when single-step state encoding is performed, the present application requires at least two historical data input LSTM channel networks to participate in training, so the prediction step length n should not exceed half of the original sequence window length t.

[0052] S4, a two-way decoder is constructed, and the encoded state features are input into the two-way decoder for global recursive decoding and single-step direct decoding to obtain global decoding features and single-step decoding features;

[0053] Specifically, the two-way decoder is composed of two LSTM networks, which respectively perform global recursive decoding operation and single-step direct decoding operation. The input of the global recursive decoding operation is the global encoded state feature The corresponding n-step recursive decoding features are output Recursion refers to the use of the decoding result of the previous time in the LSTM decoding operation. The global decoding feature is obtained by recursion, and its expression is as follows:

[0054]

[0055] Wherein, The global encoded state is represented by G; The first LSTM channel decoder is represented by L1; The n-step global recursive decoding feature is represented by G n.

[0056] The single-step direct decoding operation outputs the decoding feature of the corresponding time step according to the input single-step encoded feature. Since the first-step prediction does not perform single-step encoding, the global encoded feature is used instead, because the first-step prediction does not need to crop the sequence, so its effect is the same as the global encoding. The expression of the single-step decoding feature is as follows:

[0057]

[0058] Wherein, The n-th LSTM channel decoder is represented by Ln; The single-step encoded state of the n-th LSTM channel is represented by Sn; The n-th single-step decoding feature is represented by Sn.

[0059] S5, the global decoding features and the single-step decoding features are spliced and fused to obtain fusion features;

[0060] Specifically, the fusion is performed in a splicing manner, and its expression is as follows:

[0061]

[0062] Wherein, The n-step global recursive decoding feature is represented by G n. The n-th single-step decoding feature is represented by Sn. The n-th fusion feature is represented by Fn.

[0063] S6, the fusion features are trained and predicted based on the residual network and the padding strategy to obtain multi-step prediction results.

[0064] Specifically, referring to Figure 4 , the residual network is composed of one convolution layer and one ReLU activation function, since the time sequence convolution needs to add padding to ensure the time sequence length unchanged, and the common zero padding will interfere with the learning of sequence information. Therefore, the convolution kernel size of the convolution layer in the embodiment of the present application is set to 2, and the padding strategy is used to pad the sequence of the fusion features. The padding strategy is to pad the left boundary value on the left side of the sequence of the fusion features to ensure the sequence length unchanged. The expression of the residual network is as follows:

[0065] y=f(F(x)+x)

[0066] Wherein, x represents the padded fusion feature sequence, F(·) represents the residual function composed of one convolution layer and one ReLU activation function, f(·) represents the full connection operation, and y represents the multi-step prediction result.

[0067] The above is a specific description of the preferred embodiment of the present application, but the present application is not limited to the above-mentioned embodiments, and those skilled in the art can make various equivalent modifications or replacements without departing from the spirit of the present application. These equivalent modifications or replacements are all included in the scope defined by the claims of the present application.

Claims

1. A wind speed short-term multi-step prediction method that fuses direct prediction and recursive prediction, characterized in that, The method comprises the following steps: obtaining an original data set of a short-term multi-step wind speed prediction task and performing normalization processing to obtain a training data set; constructing a multi-channel LSTM encoding layer based on a global encoder and a single-step encoder; inputting the training data set into the multi-channel LSTM encoding layer for encoding to obtain an encoding state feature; constructing a double-path decoder and inputting the encoding state feature into the double-path decoder for global recursive decoding and single-step direct decoding to obtain a global decoding feature and a single-step decoding feature; splicing and fusing the global decoding feature and the single-step decoding feature to obtain a fused feature; training and predicting the fused feature based on a residual network and a padding strategy to obtain a multi-step prediction result; the multi-channel LSTM encoding layer is composed of a first LSTM channel and other LSTM channels, wherein: the first LSTM channel adopts a global encoder; the other LSTM channels adopt single-step encoders; the expression of the encoding state feature is as follows: wherein x t denotes the t-th sample in the training data set; denotes the first LSTM channel; denotes the global encoding state; denotes the n-th LSTM channel; denotes the one-step encoding state of the n-th LSTM channel; the expression of the global decoding feature is as follows: wherein, denotes the global encoding state; denotes the first LSTM channel decoder; denotes the global recurrent decoding feature at the n-th step; the expression of the single-step decoding feature is as follows: wherein, denotes the nth LSTM channel decoder; denotes the single-step encoding state of the nth LSTM channel; denotes the nth single-step decoding feature; the expression of the fused feature is as follows: wherein represents the n-th fused feature.

2. The method according to claim 1, wherein, the padding strategy is to fill the left boundary value on the left side of the fused feature sequence to ensure that the sequence length remains unchanged.

Images