Wind power interval prediction method based on deep learning and Markov chain Monte Carlo

Abstract

The invention relates to a wind power interval prediction method based on deep learning and Markov chain Monte Carlo. The method comprises the steps of carrying out feature extraction on historical wind power generation data by utilizing LSTM, constructing a Gaussian likelihood function based on a mean value and a variance according to hidden layer information representation, carrying out modeling processing on the Gaussian likelihood function by utilizing negative binomial distribution, optimizing hidden layer parameters of the LSTM by utilizing a loss function of a probability density function obtained through design, and obtaining hidden layer parameters of the LSTM. Obtaining a mean value and a variance of the optimized Gaussian likelihood function; a Markov chain is constructed according to the distribution characteristics of data characteristics represented by the mean value and variance of the optimized Gaussian likelihood function by adopting a Markov chain Monte Carlo algorithm to randomly generate a large number of samples, so that the stable distribution of the samples is the same as the known distribution; and taking a prediction interval with a predetermined size as a final prediction result for the distribution of the samples which most accord with the historical data characteristics. By adopting the method, the problem that a large amount of historical data cannot be effectively utilized by probability prediction can be solved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a wind power interval prediction method based on deep learning and Markov chain Monte Carlo. Background Art

[0002] Under the guidance of policies and the promotion of enterprises, wind power generation, as a green, clean and convenient renewable energy source, has seen its installed capacity grow rapidly every year. However, wind power generation relies on wind energy with strong randomness, weak periodicity and poor stability, which brings huge difficulties to the safe and stable operation of the power system after it is connected to the grid. In order to improve the power system's ability to undertake and absorb wind power generation connected to the grid, academia and industry have proposed a variety of wind power generation prediction methods to achieve high-precision prediction. According to different prediction time scales, wind power generation prediction can be divided into short-term, medium-term and long-term predictions; according to different model principles, early wind power generation predictions were mainly physical prediction models that analyzed numerical weather information and statistical models based on a small amount of historical data. With the improvement of computer performance and the development of big data analysis technology, models based on deep learning and complex hybrid prediction models have become the main research directions of wind power generation prediction. According to the different forms of prediction results, wind power generation prediction can be divided into deterministic point prediction and probabilistic prediction.

[0003] At present, wind power generation prediction is mainly achieved by analyzing and extracting the characteristics of historical data, and realizing the deterministic point prediction of future wind power generation based on deep learning neural networks. Although this prediction can obtain the predicted value, it cannot reflect the uncertainty level of the predicted value, and it is difficult to meet the long-term overall planning of the power system dispatch after wind power is connected to the grid by power companies. Compared with the deterministic point prediction method, the interval prediction method can generate a prediction interval that reflects the uncertainty level, provide more and more reasonable wind power generation prediction information, and thus provide strong support for the overall robust dispatch of the power system. Wind power generation prediction methods based on probabilistic prediction models have been widely studied, mainly divided into parametric methods and non-parametric methods. The method of estimating parameters by determining the probability distribution function based on prior knowledge or assumptions about the data distribution is called the parametric method. The parametric method is simple, effective and low-cost, but when the actual distribution of the parameters is different from that of the data, its prediction performance may be poor. Non-parametric methods that represent uncertainty are mainly divided into three forms: interval method, quantum method and probability density method. These methods require more complex estimation processes, but have better prediction performance and adaptability. At present, probabilistic prediction based on non-parametric methods makes little use of data and requires a lot of time for parameter estimation. Summary of the invention

[0004] Based on this, it is necessary to provide a wind power interval prediction method based on deep learning and Markov Chain Monte Carlo that can overcome the problem that probabilistic prediction cannot effectively utilize a large amount of historical data in order to address the above technical problems.

[0005] A wind power interval prediction method based on deep learning and Markov chain Monte Carlo, the method comprising:

[0006] Obtain historical wind power generation data; use LSTM to extract features from the historical wind power generation data to obtain hidden layer information representation; the hidden layer information representation includes data features;

[0007] According to the hidden layer information representation, a Gaussian likelihood function based on mean and variance is constructed, and the Gaussian likelihood function is modeled using negative binomial distribution to obtain a probability density function; the probability density function can be used to represent the maximum approximation process of the prediction target;

[0008] The loss function of the probability density function is designed to optimize the hidden layer parameters of the LSTM to obtain the mean and variance of the optimized Gaussian likelihood function;

[0009] The Markov Chain Monte Carlo algorithm is used to construct a Markov chain to randomly generate a large number of samples based on the distribution characteristics of the data features represented by the mean and variance of the optimized Gaussian likelihood function, so that its stable distribution is the same as the known distribution, and the sample distribution that best matches the characteristics of the historical data is used as the final prediction result with a prediction interval of a predetermined size.

[0010] The above-mentioned wind power interval prediction method based on deep learning and Markov chain Monte Carlo, this application uses LSTM model to extract time series features from a large amount of historical data, and realizes the probability density description of data through Gaussian likelihood function and negative binomial distribution modeling, so that each historical data feature can have an impact on the prediction result. Compared with the traditional interval prediction method, the method described in this application can combine the deep learning model to realize the capture of historical wind power generation data characteristics, and realizes the combination of deep learning and probability prediction by taking the parameters of the deep learning model as the input of the probability density function parameter characteristic equation. Then, the Markov chain Monte Carlo algorithm is used to construct a Markov chain to randomly generate a large number of samples according to the distribution characteristics of the data characteristics characterized by the mean and variance of the optimized Gaussian likelihood function, so that its stable distribution is the same as the known distribution, and the sample distribution that best meets the historical data characteristics is used as the final prediction result with a predetermined prediction interval, which solves the problem that the previous probability prediction is completely dependent on historical data. The distribution characteristics of historical data are captured by the Markov chain Monte Carlo algorithm, and at the same time, a new distribution is randomly constructed based on the current input data, and the historical data distribution is fitted to achieve an acceptable error range, thereby obtaining a more accurate interval prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 A schematic diagram of a flow chart of a wind power interval prediction method based on deep learning and Markov chain Monte Carlo in one embodiment;

[0012] Figure 2 It is a framework diagram of the interval prediction process in one embodiment;

[0013] Figure 3 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0014] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0015] In one embodiment, Figure 1 and Figure 2 As shown, a wind power interval prediction method based on deep learning and Markov chain Monte Carlo is provided, comprising the following steps:

[0016] Step 102, obtaining historical wind power generation data; using LSTM to extract features from the historical wind power generation data to obtain hidden layer information representation; the hidden layer information representation includes data features.

[0017] Traditional probabilistic forecasting methods, such as forecasting techniques based on physical models and statistical models, mainly rely on numerical weather information or a small amount of historical data. These models have obvious deficiencies in capturing the long-term time series characteristics and complex nonlinear characteristics of wind power data. This application introduces LSTM. Through the loop structure and memory gate mechanism, LSTM can effectively capture the long-term dependence and short-term variability in historical wind power data, and provide rich feature expressions for subsequent forecasts. LSTM is used to process a large amount of historical wind power data, and generates hidden layer information representations by extracting features from the data. These hidden layer information representations not only contain the basic statistical characteristics of the data, but also capture the time series patterns and potential regularities of the data, and can provide more comprehensive data support for probability density modeling. Compared with traditional methods, the introduction of LSTM enables the wind power forecasting model to adapt to complex time series data, thereby improving the accuracy of the forecast.

[0018] Step 104, construct a Gaussian likelihood function based on mean and variance according to the hidden layer information representation, use negative binomial distribution to model the Gaussian likelihood function, and obtain a probability density function; the probability density function can be used to represent the maximum approximation process of the prediction target.

[0019] After the hidden layer information obtained by LSTM is represented, the probability distribution of the data is modeled through the Gaussian likelihood function. The Gaussian likelihood function uses the mean and variance as parameters, and can express the central tendency and distribution width of the data, which is very important for processing the volatility and uncertainty in wind power data. Gaussian distribution usually has a good fitting effect, but when the distribution of the predicted results does not completely conform to the characteristics of the normal distribution, it may be difficult to obtain sufficiently accurate results by relying solely on the Gaussian distribution. The present application further uses the negative binomial distribution to model the Gaussian likelihood function. This modeling method not only retains the centralization of the Gaussian distribution, but also improves the model's ability to fit non-normal characteristic data. This adjustment based on the negative binomial distribution can make up for the limitations of the Gaussian distribution, making the final probability density function more flexible and applicable, and thus more robust in dealing with complex fluctuations in wind power data.

[0020] Step 106, designing a loss function of a probability density function to optimize the hidden layer parameters of the LSTM, and obtaining the mean and variance of the optimized Gaussian likelihood function.

[0021] In order to make the prediction model fit the data distribution characteristics more accurately, this application designs a probability density function loss function to optimize the hidden layer parameters of LSTM. By optimizing the mean and variance, the Gaussian likelihood function is made to fit the actual data distribution as closely as possible, which greatly improves the accuracy of the prediction results.

[0022] The optimization of loss function plays a vital role in deep learning, because through this process, the model not only obtains the ability to fit the input data, but also has the tolerance for the random volatility and uncertainty unique to wind power generation. The optimized Gaussian likelihood function can more flexibly adjust the mean and variance to cope with abnormal fluctuations or short-term instability that may occur in the input data, thereby ensuring the reliability of wind power forecasting.

[0023] Step 108, using the Markov Chain Monte Carlo algorithm to construct a Markov chain to randomly generate a large number of samples based on the distribution characteristics of the data features represented by the mean and variance of the optimized Gaussian likelihood function, so that its stable distribution is the same as the known distribution, and the sample distribution that best matches the historical data characteristics is used as the final prediction result with a prediction interval of a predetermined size.

[0024] After obtaining the optimized Gaussian likelihood function, sampling using the Markov Chain Monte Carlo algorithm (MCMC) is a key step. The MCMC algorithm can generate a large number of samples that meet the distribution characteristics of historical data from the probability density function based on the optimized mean and variance. These samples not only have the characteristic distribution of historical data, but also make the sample distribution closer to the known data distribution through the stationary distribution characteristics of the Markov chain. In this way, by constructing a large number of samples, the model can more accurately fit the distribution characteristics of historical data and form a stable and reliable prediction interval. Compared with the traditional interval prediction method based on static data, the MCMC algorithm can dynamically adapt to the changing characteristics of input data and continuously generate new samples, so that the prediction results can more effectively cope with the uncertainty of wind energy resources. This dynamic adaptability significantly improves the flexibility of the wind power prediction model and ensures the accuracy and reliability of the prediction interval. At the same time, it solves the problem that the previous probability prediction is completely dependent on historical data. The Markov Chain Monte Carlo algorithm captures the distribution characteristics of historical data. At the same time, a new distribution is randomly constructed based on the current input data, and it is used to fit the distribution of historical data until an acceptable error range is achieved, thereby obtaining a more accurate interval prediction.

[0025] In the above-mentioned wind power interval prediction method based on deep learning and Markov chain Monte Carlo, the present application uses the LSTM model to extract time series features from a large amount of historical data, and realizes the probability density description of the data through Gaussian likelihood function and negative binomial distribution modeling, so that each historical data feature can have an impact on the prediction result. Compared with the traditional interval prediction method, the method described in the present application can combine the deep learning model to realize the capture of historical wind power generation data characteristics, and realizes the combination of deep learning and probability prediction by taking the parameters of the deep learning model as the input of the probability density function parameter characteristic equation. Then, the Markov chain Monte Carlo algorithm is used to construct a Markov chain to randomly generate a large number of samples according to the distribution characteristics of the data features represented by the mean and variance of the optimized Gaussian likelihood function, so that its stable distribution is the same as the known distribution, and the sample distribution that best meets the historical data characteristics is used as the final prediction result with a prediction interval of a predetermined size, which solves the problem that the previous probability prediction is completely dependent on historical data. The Markov Chain Monte Carlo algorithm is used to capture the distribution characteristics of historical data. At the same time, a new distribution is randomly constructed based on the current input data, and this is used to fit the historical data distribution until an acceptable error range is achieved, thereby obtaining a more accurate interval forecast.

[0026] In a specific embodiment, the present application adopts the following two evaluation indicators: prediction interval coverage probability (PICP) and prediction interval normalized average width (PINAW).

[0027] Prediction interval coverage probability (PICP) refers to the probability that the true value falls within the prediction interval in multiple predictions. It reflects the coverage of the prediction interval to the true value, and the calculation formula is as follows:

[0028]

[0029] The closer PICP is to 1, the stronger the coverage of the prediction interval to the true value and the more reliable the prediction result.

[0030] The normalized average width of prediction intervals (PINAW) refers to the ratio of the average width of the prediction interval to the range of the data set. It reflects the compactness of the prediction interval and is calculated as shown in formula (2):

[0031]

[0032] Among them, R represents the range of the prediction interval, which is used to standardize the width of the prediction interval, so as to make the evaluation index more objective. The smaller the PINAW is, the more compact the prediction interval is and the more accurate the prediction result is.

[0033] In the above formulas, N is the number of samples, y i is the predicted value, UB i and LB i are the upper and lower limits of the prediction interval.

[0034] In one embodiment, the LSTM includes an input gate, a forget gate and an output gate; the input gate controls whether to retain the previous hidden layer state and the current input state, the forget gate controls the forgetting and discarding of information in the data stream, and the output gate controls the output value of the memory unit.

[0035] In a specific embodiment, based on historical wind power generation data [y i,1 ,y i,2 ,…,y i,t-1 ]=y i,1:t-1 To estimate the future wind power generation situation [y i,t ,y i,t+1 ,…,y i,t+T ]=y i,t:T , the wind power generation data i at time t is represented as y i,t , t represents a point with known time, and t:T represents a point with unknown time. The approximation process of this prediction target is expressed as follows:

[0036] f(y i,1:t-1 ,x i,1:t-1 )→y i,t:T (3)

[0037] where f(·) represents the approximation function, x∈R m×nDenotes a dependent variable.

[0038] This application uses LSTM as an autoregressive convolutional network to learn and extract data features. The input gate of the LSTM model unit controls whether to retain the previous hidden layer state and the current input state, the forget gate controls the forgetting and discarding of information in the data stream, and the output gate controls the output value of the memory unit. The calculation formula at time step t is as follows:

[0039] i t =σ(W ix x t +W ih h t-1 +W ic c t-1 +b i ) (4)

[0040] f t =σ(W fx x t +W fh h t-1 +W fc c t-1 +b f ) (5)

[0041]

[0042] o t =σ(W ox x t +W oh h t-1 +W oc c t-1 +b o ) (7)

[0043] where x t ,h t-1 and c t-1 are the input vector of LSTM at time t, the output of the previous LSTM storage unit at time step t-1, and the state information of the storage unit at time step t-1. W is the corresponding weight matrix, W ·x represents the connection matrix between the input and storage modules of the previous hidden layer, W ·h is the connection matrix between the output of the previous hidden layer and the storage module, W ·c is the diagonal matrix connecting the three gates and storage units in the storage module. b is the bias term, σ and are two different nonlinear activation functions, namely sigmoid and hyperbolic tangent function. ⊙ represents the Hadamard product of two matrices, that is, the element-by-element multiplication between matrices. Based on the LSTM’s ability to learn and extract features of historical data and current input data, its hidden layer h t-1Contains the final learning results. For better description, the hidden layer information of the i-th group input at time t is represented as h i,t .

[0044] In one embodiment, the constructed model can directly predict the probability distribution of future wind power generation data. The Gaussian likelihood function is used to characterize the distribution characteristics of historical wind power generation data. The parameter matching method based on the Gaussian likelihood function of mean and variance can be expressed as:

[0045]

[0046] μ(h i,t )=w μ h i,t

[0047] σ(h i,t )=log(1+exp(w σ h i,t +b σ ))

[0048] Among them, h i,t represents the hidden layer information representation of the i-th group input at time t, y represents the predicted value, μ represents the mean, σ represents the variance, and w μ represents the weight of the mean, w σ represents the weight of the variance, b σ Represents the bias term of the variance.

[0049] In one embodiment, the Gaussian likelihood function is modeled using negative binomial distribution to obtain a probability density function, including:

[0050] The Gaussian likelihood function is modeled using negative binomial distribution, and the probability density function is obtained as follows:

[0051] μ(h i,t )=log(1+exp(w μ h i,t +b μ ))

[0052] σ(h i,t )=log(1+exp(w σ h i,t +b σ ))

[0053]

[0054] Among them, h i,t represents the hidden layer information representation of the i-th group input at time t, y represents the predicted value, μ(h i,t ) represents the mean value constructed based on the neural network, σ(hi,t ) represents the variance based on the neural network, w μ represents the weight of the mean, w σ represents the weight of the variance, b σ represents the bias term of the variance, b μ Represents the bias term about the mean.

[0055] In one embodiment, the probability density function can be used to represent the maximum approximation process of the predicted target as

[0056]

[0057] Among them, T represents the length of the data set, y i,1:t-1 represents historical wind power generation data, x i,1:t-1 Denotes a dependent variable.

[0058] In one embodiment, the loss function of the probability density function is designed to be

[0059]

[0060] Where N represents the number of data sets, T represents the length of the data set, and Y i,t represents the real data of the i-th group of input at time t, and l(y|μ,σ) represents the probability density function.

[0061] In a specific embodiment, for the real data corresponding to the LSTM training data, a loss function between the probability density function and the real data is constructed. First, in the historical wind power generation data {y i,1:T} i＝1,...,N and the related variable x i,1:T Under known conditions, the real data Y corresponding to the predicted result can be determined i,1:T According to the principle of gradient descent in deep learning models, the above parameters are all based on the parameters of the deep learning model LSTM. The parameters that constitute the model, such as h(·) of LSTM, μ(·) and σ(·) of Gaussian likelihood estimation function, can be learned and updated through the following loss function:

[0062]

[0063] Based on the above loss, the parameters of the LSTM deep neural network are updated, and then the hidden layer h(·) is updated. Then, based on the hidden function h(·), the parameters of the Gaussian likelihood function are reconstructed, so that all parameters can be optimized by the gradient descent method.

[0064] After processing and training by the aforementioned method, more accurate data characteristics of historical wind power data can be obtained. The Markov chain Monte Carlo algorithm can construct a Markov chain to randomly generate a large number of samples based on the distribution characteristics of known data, and finally obtain the result whose stable distribution is closest to the known distribution. Through the training process combining deep neural network with Gaussian likelihood function, parameters μ and σ that can effectively characterize historical wind power generation data are obtained. The parameters μ and σ are used to construct a logarithmic function that conforms to the Gaussian distribution as follows:

[0065]

[0066] The above logarithmic function can improve the computational efficiency of the Markov Chain Monte Carlo algorithm, the applicability of the optimization algorithm, the simplification of the maximum likelihood estimation, and the efficiency of the gradient calculation.

[0067] Then, a random walk algorithm is used to construct a Markov chain so that its stable distribution (that is, the state distribution after a sufficiently long time) is the same as the target distribution. The specific process is as follows:

[0068]

[0069] The sample distribution that best fits the characteristics of historical data is taken as the final prediction result with a 90% prediction interval, and then interval prediction is performed on all data in the test set. The corresponding interval evaluation indicators PICP and PINAW are 0.72 and 1.1326, respectively.

[0070] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0071] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 3As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a wind power interval prediction method based on deep learning and Markov chain Monte Carlo is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a button, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.

[0072] Those skilled in the art will understand that Figure 3 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0073] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing related hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in the present application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRS

[0074] DRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0075] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments 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.

[0076] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A wind power interval prediction method based on deep learning and Markov chain Monte Carlo, characterized in that: The method comprises: Acquire historical wind power generation data; use LSTM to extract features from the historical wind power generation data to obtain hidden layer information representation; the hidden layer information representation includes data features; A Gaussian likelihood function based on mean and variance is constructed according to the hidden layer information representation, and the Gaussian likelihood function is modeled using negative binomial distribution to obtain a probability density function; the probability density function can be used to represent the maximum approximation process of the prediction target; Designing a loss function of the probability density function to optimize the hidden layer parameters of the LSTM to obtain the mean and variance of the optimized Gaussian likelihood function; The Markov Chain Monte Carlo algorithm is used to construct a Markov chain to randomly generate a large number of samples based on the distribution characteristics of the data features represented by the mean and variance of the optimized Gaussian likelihood function, so that its stable distribution is the same as the known distribution, and the sample distribution that best matches the characteristics of the historical data is used as the final prediction result with a prediction interval of a predetermined size.

2. The method according to claim 1, characterized in that The LSTM includes an input gate, a forget gate and an output gate; the input gate controls whether to retain the previous hidden layer state and the current input state, the forget gate controls the forgetting and discarding of information in the data stream, and the output gate controls the output value of the memory unit.

3. The method according to claim 1, characterized in that A Gaussian likelihood function based on mean and variance is constructed according to the hidden layer information representation, including: μ(h i,t )=w μ h i,t σ(h i,t )=log(1+exp(w σ h i,t +b σ )) Among them, h i,t represents the hidden layer information representation of the i-th group input at time t, y represents the predicted value, μ represents the mean, σ represents the variance, and w μ represents the weight of the mean, w σ represents the weight of the variance, b σ Represents the bias term of the variance.

4. The method according to any one of claims 1 to 3, characterized in that: The Gaussian likelihood function is modeled using negative binomial distribution to obtain a probability density function, including: The Gaussian likelihood function is modeled using negative binomial distribution, and the probability density function is obtained as follows: μ(h i,t )=log(1+exp(w μ h i,t +b μ )) σ(h i,t )=log(1+exp(w σ h i,t +b σ )) Among them, h i,t represents the hidden layer information representation of the i-th group input at time t, y represents the predicted value, μ(h i,t ) represents the mean, σ(h i,t ) represents the variance, w μ represents the weight of the mean, w σ represents the weight of the variance, b σ represents the bias term of the variance, b μ Represents the bias term about the mean.

5. The method according to claim 4, characterized in that The probability density function can be used to represent the maximum approximation process of the predicted target: Among them, T represents the length of the data set, y i,1:t-1 represents historical wind power generation data, x i,1:t-1 Denotes a dependent variable.

6. The method according to claim 1, characterized in that The loss function for designing the probability density function is Among them, N represents the number of samples, T represents the length of the data set, and Y i,t represents the real data of the i-th group of input at time t, and l(y|μ,σ) represents the probability density function.

Images