A wind power prediction method and system

Abstract

The present disclosure provides a wind power prediction method and system, which relates to the field of wind power prediction, including: collecting weather forecast data and measured wind power data at each historical moment to construct a historical data set; using a BP neural network to learn the relationship between each meteorological factor and wind power in the historical data set, obtaining a power prediction model, and making a preliminary prediction of wind power; based on the historical data set, using the Seq2Seq‑AM algorithm, establishing an error inference model, and inferring the prediction error of the preliminary prediction result; using the autoencoder algorithm, establishing an error correction discrimination model, judging whether to use the prediction error to correct the preliminary prediction result, and then obtaining the final wind power of the distributed wind farm station. The present invention is based on Seq2Seq‑AM and the autoencoder to infer the prediction error and judge whether to use the prediction error for correction, thereby achieving accurate prediction of the wind power of the target station.

Description

Technical Field

[0001] The present disclosure relates to the field of wind power prediction, and in particular to a wind power prediction method and system. Background Art

[0002] The large-scale grid connection of wind power generation has increased the difficulty of real-time dispatching of the power grid, increased the operating cost of the power grid, and brought unprecedented severe challenges to the safe and stable operation of the power grid; wind power prediction can improve the predictability of wind power output and ensure the safe and stable operation of the power grid. Among them, the wind power prediction models mainly include physical models, statistical models and artificial intelligence models; the physical model is a method based on the results of numerical weather forecasts, which is converted into wind power by physical calculation or model simulation, mainly including a series of micro-meteorological models such as logarithmic wind profile model, geostrophic drag law, roughness change model, terrain change model, etc. Statistical model is a kind of mapping relationship between numerical weather forecast, historical power data and power data of the time period to be predicted directly from the data through one or more methods, and a prediction model is established on this basis. Commonly used statistical models include persistence model, autoregressive moving average (ARMA) model, differential autoregressive moving average (ARIMA) model and other artificial intelligence models have significant advantages in dealing with high-dimensional nonlinear problems. Since wind power prediction problems are often highly nonlinear problems, artificial intelligence models are commonly used prediction methods in recent years, such as BP neural network models, recurrent neural network models (RNN), etc.

[0003] Due to the limitations of the generalization ability of the artificial intelligence model itself, the use of artificial intelligence models to predict wind power will inevitably produce prediction errors. Therefore, it is necessary to correct the prediction errors of the prediction results. The existing error correction methods are generally corrected through correction algorithms such as wind speed-power curves and median corrections or machine learning correction models. However, these commonly used correction methods do not consider the temporal correlation of the correction errors and cannot correct the prediction errors of wind power prediction to the greatest extent. At the same time, the existing correction models still have the problem of achieving the best global prediction effect based only on the training data set, lacking the extraction and description of local features, so there is a problem of large local prediction errors from time to time. At the same time, in the actual prediction process, the operating environment of new energy is complex and changeable, and the environmental adaptability and prediction robustness of a single model are not strong.

[0004] Therefore, the existing artificial intelligence wind power forecasting solutions have deficiencies in prediction accuracy, adaptability and robustness. Summary of the invention

[0005] In order to solve the above problems, the present disclosure proposes a wind power prediction method and system, which, based on Seq2Seq-AM and autoencoder, infers the prediction error and determines whether to use the prediction error for correction, thereby achieving accurate prediction of the wind power of the target site.

[0006] According to some embodiments, the present disclosure adopts the following technical solutions:

[0007] A wind power prediction method, comprising:

[0008] Collect the weather forecast data and measured wind power data of the distributed wind farms to be predicted at each historical moment and construct a historical data set;

[0009] Using BP neural network, we learn the relationship between various meteorological factors and wind power in historical data sets, obtain the power prediction model, and make a preliminary prediction of wind power.

[0010] Based on the historical data set, the Seq2Seq-AM algorithm is used to establish an error inference model to infer the prediction error of the preliminary prediction results;

[0011] Using the autoencoder algorithm, an error correction discriminant model is established to determine whether to use the prediction error to correct the preliminary prediction result, and then obtain the final wind power of the distributed wind farm.

[0012] According to some embodiments, the present disclosure adopts the following technical solutions:

[0013] A wind power prediction system, comprising:

[0014] The data collection module is configured to: collect weather forecast data and measured wind power data of the distributed wind farm to be predicted at each historical moment, and construct a historical data set;

[0015] The power prediction module is configured to: use BP neural network to learn the relationship between various meteorological factors and wind power in the historical data set, obtain a power prediction model, and make a preliminary prediction of wind power;

[0016] The error prediction module is configured to: establish an error inference model based on the historical data set and use the Seq2Seq-AM algorithm to infer the prediction error of the preliminary prediction results;

[0017] The correction discrimination module is configured to: use the autoencoder algorithm to establish an error correction discrimination model to determine whether to use the prediction error to correct the preliminary prediction result, and then obtain the final wind power of the distributed wind farm station.

[0018] According to some embodiments, the present disclosure adopts the following technical solutions:

[0019] A computer program product comprises a computer program, wherein when the computer program is executed by a processor, the wind power prediction method is implemented.

[0020] According to some embodiments, the present disclosure adopts the following technical solutions:

[0021] A non-transitory computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by a processor, a wind power prediction method is implemented.

[0022] According to some embodiments, the present disclosure adopts the following technical solutions:

[0023] An electronic device comprises: a processor, a memory and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory so that the electronic device executes and implements the wind power prediction method.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] The present invention uses the Seq2Seq-AM algorithm to infer the prediction error and considers the temporal relationship of wind power related data through sequence prediction, thereby improving the prediction accuracy of the prediction error.

[0026] The present invention combines the attention mechanism with Seq2Seq, and calculates different semantic vectors for each moment by allocating attention weights to the decoded information, so that the semantic vector can retain information from a relatively long time ago, while increasing the amount of information retained at important moments and improving the prediction accuracy of the prediction error.

[0027] The present invention uses an autoencoder model to realize whether to use the prediction error result for each day to correct the prediction result, thereby overcoming the problem that the error correction model expands the correction error on some dates, thereby improving the prediction accuracy of new energy power. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The accompanying drawings constituting a part of the present disclosure are used to provide a further understanding of the present disclosure. The illustrative embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation on the present disclosure.

[0029] Figure 1 The present invention is a method flow chart of an embodiment of the present invention.

[0030] Figure 2 It is a schematic diagram of the structure of the BP neural network model of the embodiment of the present disclosure.

[0031] Figure 3 It is a structural diagram of the Seq2Seq model of the embodiment of the present disclosure.

[0032] Figure 4 It is a structural diagram of the GRU model of the embodiment of the present disclosure.

[0033] Figure 5 This is a structural diagram of a decoder based on the attention mechanism of an embodiment of the present disclosure.

[0034] Figure 6 This is a structural diagram of the prediction error inference model of an embodiment of the present disclosure.

[0035] Figure 7 This is a structural diagram of the autoencoder model of an embodiment of the present disclosure.

[0036] Figure 8 A correction flow chart combining the prediction error inference model and the error correction discrimination model of the embodiment of the present disclosure. DETAILED DESCRIPTION

[0037] The present disclosure is further described below in conjunction with the accompanying drawings and embodiments.

[0038] It should be noted that the following detailed descriptions are all illustrative and are intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present disclosure belongs.

[0039] 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 disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0040] Example 1

[0041] In one embodiment of the present disclosure, a wind power prediction method is provided, comprising:

[0042] Collect the weather forecast data and measured wind power data of the distributed wind farms to be predicted at each historical moment and construct a historical data set;

[0043] Using BP neural network, we learn the relationship between various meteorological factors and wind power in historical data sets, obtain the power prediction model, and make a preliminary prediction of wind power.

[0044] Based on the historical data set, the Seq2Seq-AM algorithm is used to establish an error inference model to infer the prediction error of the preliminary prediction results;

[0045] Using the autoencoder algorithm, an error correction discriminant model is established to determine whether to use the prediction error to correct the preliminary prediction result, and then obtain the final wind power of the distributed wind farm.

[0046] As an embodiment, a wind power prediction method disclosed in the present invention is based on Seq2Seq-AM and autoencoder to infer the prediction error and determine whether to use the prediction error for correction, so as to achieve accurate prediction of the wind power of the target site. The specific implementation process is as follows: Figure 1 As shown:

[0047] Step S1: Collect the weather forecast data and measured wind power data of the distributed wind farm to be predicted at each historical moment, and construct a historical data set.

[0048] The weather forecast data here include meteorological factors such as 100m wind speed, 100m wind direction, 10m wind speed, 10m wind direction, temperature, humidity, and air pressure.

[0049] After the data collection is completed, the data must be processed for outliers and normalized.

[0050] In terms of outlier processing, the outliers of meteorological forecast data and measured wind power data mainly include outliers and missing values. For outliers, the quartile method is used to eliminate the outliers and replace them with missing values. On this basis, for missing values, the quadratic spline interpolation method is used to interpolate the missing values. The formula is as follows:

[0051] (1)

[0052] (2)

[0053] in, for The functional relationship between the independent variable and the dependent variable is for hour The coefficient of .

[0054] In order to eliminate the influence of different dimensions on prediction accuracy and speed up the convergence of the model, it is necessary to use the maximum and minimum normalization algorithm to normalize the data. The formula is as follows:

[0055] (3)

[0056] in, is the normalized sequence; is the original photovoltaic series or meteorological series sample; and are the minimum and maximum values ​​of the sample data respectively. After normalization, the data values ​​are all within [0,1].

[0057] Step S2: Using the BP neural network, learn the relationship between each meteorological factor and wind power in the historical data set, obtain a power prediction model, and make a preliminary prediction of wind power.

[0058] The power prediction model adopts the Pearson correlation coefficient method to analyze the correlation between various meteorological factors in the historical data set and the measured wind power. Based on the analysis results, four meteorological factors, namely 10m wind speed, 10m wind direction, 100m wind speed and 100m wind direction, are selected as the input of the power prediction model, and the wind power is the output of the power prediction model.

[0059] Specifically, in order to determine the input and output of the power prediction model, the correlation between the meteorological forecast data and the measured power data should be analyzed first.

[0060] The output power of a wind farm varies with the fluctuation of wind speed. The wind energy captured by a wind turbine can be expressed by the following formula:

[0061] (4)

[0062] Wherein, P is the output power of the wind turbine; C P is the wind energy utilization coefficient of the wind turbine; , is the area swept by the wind turbine impeller, R is the impeller radius; ρ is the air density; v is the wind speed.

[0063] From formula (4), it can be seen that wind speed is the most important factor affecting the output power of wind turbines. Under ideal conditions without considering the influence of physical factors such as wake effect and surface roughness, the wind power of a single wind turbine is proportional to the cube of the wind speed.

[0064] Wind direction also has a certain impact on wind farm power generation; when the wind speed is low, the wake effect and roughness effect are more obvious, resulting in lower power generation efficiency of the wind farm under certain wind directions; when the wind speed exceeds the rated wind speed by a certain amount, the output power will no longer be affected by the wake effect, and the wind farm can output the rated power in any wind direction.

[0065] At the same time, in formula (4), the air density ρ represents the energy contained in wind energy, and the calculation formula of air density is as follows:

[0066] (5)

[0067] Where p is the air pressure in hPa; t is the temperature in °C; p wis the water vapor pressure, in hPa.

[0068] According to formula (5), changes in meteorological factors such as air pressure, temperature, and humidity will cause changes in air density, which in turn affects the output power of wind turbines.

[0069] The above theoretical analysis of the relationship between the wind farm power generation and various meteorological factors. This embodiment uses the Pearson correlation coefficient method to determine the correlation between various meteorological factors and wind farm power. The Pearson correlation coefficient is widely used to measure the correlation between two variables. Its calculation formula is as follows:

[0070] (6)

[0071] In the formula, are the two variables to be measured, For sequence and sequence The correlation coefficient between The larger the value of, the stronger the correlation. and Respectively represent the average values ​​of two time series; n is the length of the time series.

[0072] According to the analysis results of the Pierman correlation coefficient, the 10 m wind speed, 10 m wind direction, 100 m wind speed and 100 m wind direction were selected as meteorological factors and input into the power prediction model.

[0073] The design of BP neural network mainly includes input layer, hidden layer, output layer, activation function, training function and transfer function between layers. These parameters need to be set in the process of initializing BP neural network.

[0074] like Figure 2 As shown, the number of input layers and output layers of the BP neural network are both 1, while the number of its hidden layers can be adjusted freely. The number of hidden layers is positively correlated with the ability of the neural network to fit the function, but too deep a number of layers will also lead to overfitting problems and make the training time too long; in this embodiment, since there are two levels of wind speed, and neither is the wind speed at the hub height of the wind turbine, the relationship between the wind power output is relatively complex, so a perceptron consisting of three hidden layers is constructed. In this embodiment, the number of nodes in the input layer is 4, which are the 10m wind speed, 10m wind direction, 100m wind speed and 100m wind direction at a certain moment; the number of nodes in the output layer is 1, which is the wind power prediction result at that moment; the number of nodes in the hidden layer is 6, 8, and 4 respectively.

[0075] The artificial neural network model must have nonlinear mapping learning ability, and the main function of the activation function is to provide the network with nonlinear modeling ability, and the activation function should be differentiable and monotonic. In this embodiment, the activation functions used are the hyperbolic tangent function (tanh function) and the sigmoid function.

[0076] The function value of the Sigmoid function is in the range of (0,1). It can be used to represent probability, link the result with the probability distribution, and normalize the output result. Its mathematical expression is shown in the following formula:

[0077] (7)

[0078] The hyperbolic tangent function is derived from the hyperbolic sine function and the hyperbolic cosine function. It converges quickly and its mathematical expression is shown in the following formula:

[0079] (8)

[0080] Through the above process, the input, output and network structure of the power prediction model are established. The training process of the prediction model is as follows:

[0081] (1) Input the original wind power data set containing historical wind power data and day-ahead numerical meteorological forecast data, and divide the data set into a training set and a test set;

[0082] (2) Iteratively train the BP neural network model using the training set until the error is lower than the preset error value, and then stop the training; otherwise, continue the training;

[0083] (3) Input the samples of the test set and output the final power prediction results.

[0084] Step S3: Based on the historical data set, the Seq2Seq-AM algorithm is used to establish an error inference model to infer the prediction error of the preliminary prediction results.

[0085] Due to the limitation of the generalization ability of the artificial intelligence model, the prediction of the BP neural network will inevitably have errors, so the prediction error should be corrected. This embodiment adopts the Seq2Seq-AM model to establish an error inference model, and makes an accurate inference of the prediction error by considering the time correlation between the meteorological factor sequence and the predicted power sequence, and on this basis, corrects the prediction results of the power prediction model to improve the prediction accuracy.

[0086] The power prediction error refers to the error between the power prediction value of the wind power short-term prediction model at each moment and the actual power value at that moment. Generally speaking, the probability distribution of the power prediction error presents a quasi-normal distribution. The power prediction error can be defined by the following formula:

[0087] (9)

[0088] In the formula, e t is the prediction error at time t, y t is the real power at time t, p t is the prediction result at time t.

[0089] The existing wind power related data in this embodiment are 100m wind speed, 100m wind direction, 10m wind speed, 10m wind direction and power prediction data output by the power prediction model. In order to determine the input of the error inference model, the 100m wind speed, 100m wind direction, 10m wind speed, 10m wind direction and power prediction data are subjected to correlation analysis using the Pearson correlation coefficient method. According to the correlation analysis results, the three meteorological factor sequences of 100m wind speed, 10m wind speed and 100m wind direction and the predicted power sequence are selected as the input of the error correction model.

[0090] Seq2Seq model, the full name of which is Sequence to Sequence, is sequence to sequence learning. Seq2Seq is an end-to-end deep learning model suitable for processing time series tasks. Seq2Seq is a highly compatible model framework with strong learning ability and adaptability. It can build different network structures according to different model requirements. The structure of the Seq2Seq model consists of two parts, namely the encoder and the decoder. Its specific structure is as follows: Figure 3 As shown in the figure, the role of the encoder is to encode the time series into a semantic vector of fixed length so that the semantic vector contains the information content of the complete time series; the role of the decoder is to decode the semantic vector, construct a new time series and output it; both the encoder and decoder use GRU as the encoding and decoding units.

[0091] GRU is a simplified LSTM network, consisting of a reset gate and an update gate. The update gate is formed by merging the input gate and the forget gate in the LSTM. This simplification allows GRU to have the advantages of LSTM while reducing the parameters involved in training, thereby increasing the computing speed and reducing the convergence time.

[0092] The structure of GRU is as follows Figure 4 As shown, h t The hidden layer state of this neuron; h t-1 The hidden layer state transmitted by the previous neuron; is the candidate hidden layer state, representing the input information of the sequence at time t; r t To reset the gate; zt It is the update gate, which controls h t from h t-1 The information retained in and the information obtained from it; is the Sigmoid activation function; b is the bias value; Wz, Wr, Wh are the time series value related parameter matrices; Uz, Ur, Uh are the hidden layer value related parameter matrices. The role of the reset gate is to control the forgetting of past information, which helps GRU to mine the short-term time correlation of time series data; the role of the update gate is to determine the retention of the information of the previous moment and the reception of the information at this moment, which is conducive to GRU to mine the long-term time correlation of time series data.

[0093] The specific operation of GRU neurons is as follows:

[0094]

[0095] In the formula, It means Hadamard product, that is, the dot product calculation is performed on the corresponding positions of the matrix; as shown in the formula, the closer the value of the update gate is to 1, the less past information is forgotten; the higher the value of the reset gate is, the less information needs to be forgotten at the previous moment.

[0096] The encoder uses GRU as the encoding unit to input the wind power related data sequence Encode, where is a vector composed of the 100m wind speed, 100m wind direction, 10m wind speed, 10m wind direction at time t and the power prediction data at time t output by the power prediction model. For time t, the decoding unit receives the hidden layer state vector from the previous moment and the input at this moment , and encode them into a new hidden layer state vector through the GRU module ; The output of the encoder is a semantic vector , which contains the information of the entire input time series. The formula of the encoding process can be written as:

[0097] (14)

[0098] (15)

[0099] in, and Both represent nonlinear mapping functions.

[0100] The decoder uses GRU as the decoding unit, and uses GRU to decode the semantic vector, the hidden state vector inside the decoder, and the output of the decoding unit at the previous moment, and finally forms an output sequence. At time t, the GRU decoding unit receives the hidden state vector constructed by the decoding unit at the previous moment , the output of the decoding unit at the previous moment The output semantic vector of the encoder , and maps it to the hidden layer state vector at this moment For each moment, , and semantic vector After the feedforward neural network (FFN), it is transformed into the output at this moment , that is, the prediction error at time t; FFN is part of the decoder. The encoder and decoder parts of Seq2Seq-AM only have GRU and attention mechanism parts. However, in order to convert the hidden layer and semantic vector into the final prediction error output, Seq2Seq must pass through FFN after decoding.

[0101] Finally, construct the output prediction error inference sequence ; The formula expression of the decoding process can be written as:

[0102] (16)

[0103] (17)

[0104] in, Represents a nonlinear mapping function.

[0105] In the Seq2Seq model, the semantic vector The length of is fixed, so the semantic vector The amount of information contained is limited; affected by this, the performance of the Seq2Seq model will decrease as the length of the input sequence increases, and it is impossible to establish an effective long-term temporal dependency, which affects the learning ability of the model; by introducing the attention mechanism (AM), the information of all moments in the entire time series can be connected through weights, so that the semantic vector can retain information from a relatively long time, and at the same time, the retention of important moment information can be increased; therefore, the introduction of the attention mechanism can improve the prediction accuracy of the Seq2Seq model. This embodiment introduces the attention mechanism in the decoder module, and calculates different semantic vectors for each moment by allocating attention weights to the decoded information. On this basis, the final output prediction error inference sequence is obtained.

[0106] The decoder based on the attention mechanism is as follows Figure 5As shown in the figure, From time 1 to n The corresponding hidden layer is for t The principle of the attention mechanism is to calculate the degree of association between the current input sequence and the output sequence. When the degree of association is higher, the relative attention score is higher. The higher the attention score, the greater the weight assigned to it by the attention mechanism. There are three calculation formulas for the score function of the attention score, which are:

[0107] (18)

[0108] This embodiment uses the dot score function as the score function calculation formula.

[0109] For the decoder at time t, the corresponding semantic vector The hidden state of each neuron in the encoder The calculation formula is:

[0110] (19)

[0111] (20)

[0112] (twenty one)

[0113] (twenty two)

[0114] in, Refers to the decoder hidden layer state vector at time t and the encoder's hidden layer state The similarity score between .

[0115] For the Seq2Seq-AM model, at time t, the GRU receives the semantic vector of the previous time , the hidden layer state at the previous moment To construct the hidden layer state vector at the moment At the same time, the semantic vector of this moment is constructed through the attention mechanism On this basis, the hidden layer state vector at this moment is With semantic vector Horizontal connections become new information vectors , by taking the new information vector By inputting the feedforward neural network, the wind power prediction error at time t can be inferred.

[0116] Based on the above analysis, this embodiment designs a wind power error inference model based on Seq2Seq-AM, such as Figure 6 As shown in the figure, it consists of three parts: encoder, attention layer and decoder.

[0117] Since the time resolution of wind power related data such as meteorological data and predicted power data is 15 minutes, there are 96 moments in a day. For the Seq2Seq-AM model, if the input time step is short, it cannot provide complete macro time series features, making it difficult for the model to capture the time correlation within the data and unable to learn historical data information. If the input time step is too long, the amount of computation and time consumption of the encoding process of the encoder and the decoding process of the decoder will increase significantly, reducing the training efficiency. Therefore, it is necessary to find the most suitable time step. According to the results of the comparative experiment in this embodiment, the time step of the encoder is set to 192, and the meteorological factors and predicted power data of the current day and the previous day, a total of two days, are input, that is, a wind power related data sequence composed of meteorological factors and predicted power data at 192 moments in 2 days. The time step of the decoder is also set to 96, that is, the prediction error value of 96 moments in the next day is predicted. Therefore, the Seq2Seq-AM model uses the wind power related data sequence including the 2-day meteorological factor sequence and the predicted power sequence as input, and the prediction error inference result as output.

[0118] The specific process of the error inference model based on Seq2Seq-AM is as follows:

[0119] (1) The elements of the 100-meter wind speed sequence, 100-meter wind direction sequence, 10-meter wind speed sequence, and predicted power sequence at each moment after data preprocessing are concatenated to form the input samples at that moment, and a three-dimensional input sequence is constructed.

[0120] (2) According to the requirements of the experiment, set the hyperparameters of the Seq2Seq-AM model, including the number of network layers, number of neurons, L2 adjustment factor, etc.

[0121] (3) Input the input sequence formed in (1) into the Seq2Seq-AM model and infer the prediction error.

[0122] (4) By evaluating the prediction results of the model, the model parameters are modified; by modifying the model parameters, the prediction accuracy of the model is improved, thereby further improving the correction effect of the error correction model.

[0123] Step S4: using the autoencoder algorithm, establish an error correction discrimination model to determine whether to use the prediction error to correct the preliminary prediction result, and then obtain the final wind power of the distributed wind farm.

[0124] The error correction discrimination model inputs the meteorological matrix composed of seven meteorological factors, namely, 100m wind speed, 100m wind direction, 10m wind speed, 10m wind direction, temperature, humidity, and air pressure, into the autoencoder, further reconstructs the meteorological matrix, calculates the Euclidean distance between the meteorological matrix and the reconstructed meteorological matrix, and determines whether to use the prediction error inferred by the error inference model to correct the preliminary prediction result based on a preset distance threshold.

[0125] Since the prediction results are corrected using the error inference model, the prediction error at that moment will increase on certain dates. Therefore, this embodiment uses the autoencoder principle to extract the meteorological matrix at the moment when the correction effect is good, and uses the pre-trained autoencoder model to judge whether error correction is needed on a certain day; through the error correction discrimination model, intelligent error correction of the prediction results of the prediction model can be achieved, which effectively improves the prediction accuracy of short-term wind power generation prediction.

[0126] Auto encoder (AE) is an unsupervised learning model that can be used for data feature extraction, data dimensionality reduction, data reconstruction and other tasks. Its function is to encode unlabeled information so that it has effective label features. Its role is to learn the low-dimensional feature representation of high-dimensional information, and to learn and reconstruct data by extracting low-dimensional features of high-dimensional information.

[0127] The main structure of the autoencoder is an "encoder-decoder" structure, which consists of an encoder and a decoder; the encoder is used to extract the features of the input high-dimensional vector and compress it into a low-dimensional vector; the decoder is used to learn the feature expression of the low-dimensional vector and re-decode it into a high-dimensional vector that is similar to the original input data; the ultimate goal of the autoencoder model is to complete the reconstruction of the input vector in order to obtain a reconstructed vector that is most similar to the input vector.

[0128] The structure of the autoencoder model is as follows Figure 7 As shown, the encoder converts the input high-dimensional vector Encoded as a low-dimensional feature vector , the decoder converts the low-dimensional feature vector Decode into a high-dimensional vector of the same dimension as the input vector , the encoding process of the encoder and the decoding process of the decoder can be described by the formula:

[0129] (twenty three)

[0130] (twenty four)

[0131] in, and are the nonlinear transformation activation functions of the encoder and decoder respectively, and The weight settings for the encoder and decoder respectively; and They are the bias settings for the encoder and decoder respectively; commonly used nonlinear transformation activation functions include Sigmoid, Tanh, Relu and other functions.

[0132] Through the above autoencoder model, from the input meteorological matrix Reconstructing the meteorological matrix , and then use the Euclidean distance to calculate and The distance, Euclidean distance, also known as Euclidean distance, is one of the most classic distance algorithms, and its formula is shown as follows:

[0133] (25)

[0134] Finally, based on the preset distance threshold, it is determined whether to use the prediction error to correct the preliminary prediction result.

[0135] In order to determine whether an error needs to be corrected at a certain moment by learning the meteorological matrix, it is necessary to construct a meteorological matrix data set at a moment with a good correction effect for training the error correction discrimination model, that is, the meteorological matrix data set is composed of meteorological matrices screened out according to the correction effect evaluated according to the comprehensive evaluation standard. Therefore, it is first necessary to evaluate the correction effect after the error correction of the inference result using the error inference model. This embodiment constructs a comprehensive evaluation standard as shown in the following formula:

[0136] (26)

[0137] in, It refers to the difference between the normalized mean square error (NRMSE) of the forecast result before correction and the NRMSE of the forecast result after correction. It refers to the difference between the normalized mean absolute error (NMAE) of the forecast results before correction and the NMAE of the forecast results after correction. Refers to the R-squared (R) of the forecast result before correction on a certain day. 2 ) and the R of the revised prediction results 2 The difference between the normalized maximum error ( ) refers to the difference between the NMAXE of the forecast result before correction and the NMAXE of the forecast result after correction on a certain day; NRMSE, NMAE, R 2 ,The calculation formula of NMAXE is as follows:

[0138] (27)

[0139] (28)

[0140] (29)

[0141] (30)

[0142] in, for i The actual value of wind power at the moment, for i The wind power forecast value at the time, is the predicted installed capacity of the wind farm.

[0143] By selecting the meteorological matrix of the date with better correction results, constructing the meteorological matrix data set of the moment with good correction effect is a necessary preparation for training the error discrimination model based on the autoencoder. The meteorological matrix is ​​constructed by combining the meteorological data vectors composed of 7 elements, namely, 100m wind speed, 100m wind direction, 10m wind speed, 10m wind direction, temperature, humidity, and air pressure at 96 moments in a day to form a (96, 7) meteorological matrix. This embodiment selects the meteorological matrix of the first 10% of the dates of EFF as the sample of the meteorological matrix data set to construct a data set for training the error correction discrimination model.

[0144] The autoencoder is used to learn the meteorological matrix of the date with better error correction effect, and a pre-trained autoencoder model is constructed. On this basis, through the pre-trained autoencoder model, first, the meteorological matrix of each date in the forecast time period is input to reconstruct the meteorological matrix, and then the Euclidean distance between the meteorological matrix and the reconstructed meteorological matrix is ​​calculated.

[0145] For a certain day, if the Euclidean distance between the meteorological matrix of that day and the reconstructed meteorological matrix is ​​higher than the threshold, it is considered that the day does not meet the meteorological matrix that needs error correction; on the contrary, if the Euclidean distance between the meteorological matrix of that day and the reconstructed meteorological matrix is ​​lower than the threshold, it is considered that the day meets the meteorological matrix that needs error correction; the judgment logic here is: if it is higher than the threshold, it means that the day does not meet the meteorological matrix of the date with better correction effect learned by the autoencoder model. If correction is made to this day, the correction effect will be poor, so no correction is made. On the contrary, if it is lower than the threshold, it means that the meteorological matrix of that day meets the meteorological matrix of the date with better correction effect learned by the autoencoder model. If correction is made to this day, the correction effect will be better; therefore, correction is made if it is lower than the threshold, and no correction is made if it is higher than the threshold.

[0146] During the unsupervised training process, the autoencoder model is trained to learn the input meteorological matrix of the date with good correction conditions to complete the pre-training of the autoencoder model; then, the meteorological matrix of each day in the forecast period is input into the model to obtain the output matrix; by calculating the Euclidean distance between the input matrix and the output matrix, it is determined whether the Euclidean distance exceeds the threshold, so as to determine whether the prediction error inferred by the error inference model is used to correct the prediction result at each moment.

[0147] Finally, the error inference model and the error correction discriminant model are combined to achieve high-precision correction of the prediction results. The specific process is as follows: Figure 8 As shown:

[0148] (1) Construct a wind power-related data sequence consisting of the screened meteorological factors and power forecast data. At the same time, construct a meteorological matrix consisting of all meteorological factors for each moment. The length of the sequence and the matrix are both the prediction time step, that is, 96 moments.

[0149] (2) The wind power related data sequence is input into the attention Seq2Seq-M model to obtain the inference results of the power prediction error at each moment. At the same time, the meteorological matrix of each day in the prediction period is input into the autoencoder to reconstruct the meteorological matrix.

[0150] (3) Based on whether the Euclidean distance between the meteorological matrix and the reconstructed meteorological matrix is ​​greater than a threshold, it is determined whether each day in the forecast period needs to be corrected.

[0151] Example 2

[0152] In one embodiment of the present disclosure, a wind power prediction system is provided, comprising:

[0153] The data collection module is configured to: collect weather forecast data and measured wind power data of the distributed wind farm to be predicted at each historical moment, and construct a historical data set;

[0154] The power prediction module is configured to: use BP neural network to learn the relationship between various meteorological factors and wind power in the historical data set, obtain a power prediction model, and make a preliminary prediction of wind power;

[0155] The error prediction module is configured to: establish an error inference model based on the historical data set and use the Seq2Seq-AM algorithm to infer the prediction error of the preliminary prediction results;

[0156] The correction discrimination module is configured to: use the autoencoder algorithm to establish an error correction discrimination model to determine whether to use the prediction error to correct the preliminary prediction result, and then obtain the final wind power of the distributed wind farm station.

[0157] Example 3

[0158] In one embodiment of the present disclosure, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the wind power prediction method is implemented.

[0159] Example 4

[0160] In one embodiment of the present disclosure, a non-transitory computer-readable storage medium is provided, wherein the non-transitory computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the wind power prediction method is implemented.

[0161] Example 5

[0162] In one embodiment of the present disclosure, an electronic device is provided, including: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes and implements the wind power prediction method.

[0163] The present disclosure is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present disclosure. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, 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 generate 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 flowchart and / or block diagram. 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.

[0164] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0165] Although the above describes the specific implementation methods of the present disclosure in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present disclosure. Technical personnel in the relevant field should understand that on the basis of the technical solution of the present disclosure, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present disclosure.

Claims

1. A wind power prediction method, characterized in that: include: Collect the weather forecast data and measured wind power data of the distributed wind farms to be predicted at each historical moment and construct a historical data set; Using BP neural network, we learn the relationship between various meteorological factors and wind power in historical data sets, obtain the power prediction model, and make a preliminary prediction of wind power. Based on the historical data set, the Seq2Seq-AM algorithm is used to establish an error inference model to infer the prediction error of the preliminary prediction results; Using the autoencoder algorithm, an error correction discrimination model is established to determine whether to use the prediction error to correct the preliminary prediction results, and then obtain the final wind power of the distributed wind farm station; The error correction discrimination model is trained based on a constructed meteorological matrix data set, wherein the meteorological matrix data set is composed of meteorological matrices selected based on correction effects evaluated according to a comprehensive evaluation standard; the comprehensive evaluation standard is specifically: ； in, It refers to the difference between the normalized mean square error of the forecast result before correction on a certain day and the NRMSE of the forecast result after correction. It refers to the difference between the normalized mean absolute error of the forecast results before correction and the NMAE of the forecast results after correction on a certain day. Refers to the R of the forecast result before correction on a certain day. 2 The R of the revised prediction results 2 The difference between the normalized maximum error It refers to the difference between the NMAXE of the forecast result before correction and the NMAXE of the forecast result after correction on a certain day; The error correction discriminant model inputs a meteorological matrix consisting of seven meteorological factors, namely, 100m wind speed, 100m wind direction, 10m wind speed, 10m wind direction, temperature, humidity, and air pressure, into an autoencoder; inputs the meteorological matrix of each date in the prediction time period, reconstructs the meteorological matrix, and then calculates the Euclidean distance between the meteorological matrix and the reconstructed meteorological matrix, and based on a preset distance threshold, determines whether to use the prediction error inferred by the error inference model to correct the preliminary prediction results of each date.

2. A wind power prediction method according to claim 1, characterized in that: The construction of the data set also includes preprocessing of outliers and normalization of the collected weather forecast data and measured wind power data.

3. A wind power prediction method according to claim 1, characterized in that: The power prediction model adopts the Pearson correlation coefficient method to analyze the correlation between various meteorological factors in the historical data set and the measured wind power. Based on the analysis results, four meteorological factors, namely 10m wind speed, 10m wind direction, 100m wind speed and 100m wind direction, are selected as the input of the power prediction model, and the wind power is the output of the power prediction model.

4. A wind power prediction method according to claim 1, characterized in that: The error inference model adopts the Pearson correlation coefficient method to analyze the correlation between the four meteorological factors of 10m wind speed, 10m wind direction, 100m wind speed and 100m wind direction in the data set and the power prediction value. According to the analysis results, 10m wind speed, 100m wind speed and 100m wind direction are selected to form a meteorological factor sequence, together with the predicted power sequence, to form a wind power related data sequence as the input of the error inference model, and the predicted error sequence is the output of the error inference model.

5. A wind power prediction system, characterized in that: include: The data collection module is configured to: collect weather forecast data and measured wind power data of the distributed wind farm to be predicted at each historical moment, and construct a historical data set; The power prediction module is configured to: use BP neural network to learn the relationship between various meteorological factors and wind power in the historical data set, obtain a power prediction model, and make a preliminary prediction of wind power; The error prediction module is configured to: establish an error inference model based on the historical data set and use the Seq2Seq-AM algorithm to infer the prediction error of the preliminary prediction results; The correction discrimination module is configured to: use the autoencoder algorithm to establish an error correction discrimination model to determine whether to use the prediction error to correct the preliminary prediction result, and then obtain the final wind power of the distributed wind farm station; The error correction discrimination model is trained based on a constructed meteorological matrix data set, wherein the meteorological matrix data set is composed of meteorological matrices selected based on correction effects evaluated according to a comprehensive evaluation standard; the comprehensive evaluation standard is specifically: ； in, It refers to the difference between the normalized mean square error of the forecast result before correction on a certain day and the NRMSE of the forecast result after correction. It refers to the difference between the normalized mean absolute error of the forecast results before correction and the NMAE of the forecast results after correction on a certain day. Refers to the R of the forecast result before correction on a certain day. 2 The R of the revised prediction results 2 The difference between the normalized maximum error It refers to the difference between the NMAXE of the forecast result before correction and the NMAXE of the forecast result after correction on a certain day; The error correction discriminant model inputs a meteorological matrix consisting of seven meteorological factors, namely, 100m wind speed, 100m wind direction, 10m wind speed, 10m wind direction, temperature, humidity, and air pressure, into an autoencoder; inputs the meteorological matrix of each date in the prediction time period, reconstructs the meteorological matrix, and then calculates the Euclidean distance between the meteorological matrix and the reconstructed meteorological matrix, and based on a preset distance threshold, determines whether to use the prediction error inferred by the error inference model to correct the preliminary prediction results of each date.

6. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, a wind power prediction method according to any one of claims 1 to 4 is implemented.

7. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, a wind power prediction method according to any one of claims 1 to 4 is implemented.

8. An electronic device, characterized in that: include: A processor, a memory and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory so that the electronic device executes a wind power prediction method as described in any one of claims 1 to 4.

Images