A photovoltaic power ultra-short-term probabilistic forecasting method based on wavelet decomposition and optimized deep belief network

Abstract

The present invention discloses a photovoltaic power ultra-short-term probability prediction method based on wavelet decomposition and optimized deep belief network. The method uses wavelet decomposition to perform three-level decomposition on original data to obtain subsequences, uses deep belief network to respectively establish power prediction models for the obtained subsequences, and uses genetic algorithm to find the optimal initial parameters for the model, and equally weights the prediction results obtained by each submodel to obtain the final prediction result; secondly, based on the constructed photovoltaic power point prediction model, the point prediction error sequence is analyzed, and a photovoltaic power ultra-short-term probability prediction model based on t distribution model is constructed to calculate the prediction interval. The present invention combines wavelet decomposition technology, deep belief network and genetic algorithm to perform photovoltaic power ultra-short-term probability prediction, realizes photovoltaic power ultra-short-term probability prediction, and provides effective decision-making information for photovoltaic electric field dispatchers.

Description

Technical Field

[0001] The present invention belongs to new energy consumption technology, and specifically relates to a photovoltaic power ultra-short-term probability prediction method based on wavelet decomposition and optimized deep belief network. Background Art

[0002] Energy is the cornerstone of social and economic development and an indispensable driving force for human survival. With people's increasing concern about energy security and ecological damage, the international community has accelerated the exploration and development of new energy. As a clean energy, solar energy has become one of the most promising green new energy sources with its advantages of abundant storage and convenient access. At the same time, with the country's strong investment, solar energy resources have become the fastest-growing clean energy in development and utilization. By the end of October 2022, my country's total installed capacity reached 280 million kilowatts, with an additional installed capacity of about 29 million kilowatts. The total installed capacity and additional installed capacity have ranked first in the world for many years. The utilization level of photovoltaic power generation has increased significantly. This year, the average utilization rate of photovoltaic power generation in my country has reached about 98%. The large-scale grid-connected photovoltaic power generation has put forward higher requirements and brought greater challenges to the safe and stable operation of the power grid. Therefore, it is of great significance to grasp the changing trend of photovoltaic power generation and make timely and accurate predictions for it to reasonably dispatch electricity and ensure the economic and stable operation of photovoltaic power stations.

[0003] There are two main methods for predicting photovoltaic power: one is point prediction, which can provide a power forecast value at a certain moment, but the photovoltaic power information it provides is relatively small, and the uncertainty of the prediction result cannot be shown, making it difficult to meet the management and decision-making requirements of the power system; the other is probabilistic prediction, which can provide the probability of the photovoltaic power output range at the next prediction moment, and therefore can provide more sufficient prediction information, which can better help the power system reduce the risks in operation.

[0004] In the ultra-short-term prediction of photovoltaic power, since the numerical weather forecast information is highly complex and has a low correlation with the corresponding photovoltaic power, the input is only the historical photovoltaic power data. Since the original photovoltaic power data shows obvious volatility and randomness, these nonlinear characteristics will affect the accuracy of the prediction model results. Using the wavelet decomposition algorithm to pre-decompose the original photovoltaic power data can effectively reduce the non-stationarity of the signal, making it smoother and easier to extract features. The traditional single prediction algorithm that was widely used in the early days generally cannot achieve the goal of deep mining of power data features, while the deep belief network algorithm has the advantage of extracting features and can meet the requirements of multi-level processing and high-level expression of a large amount of nonlinear data. At the same time, the genetic algorithm is used to help the deep belief network select a set of excellent initial parameters to overcome the randomness and blindness of manual settings. Summary of the invention

[0005] Purpose of the invention: In view of some shortcomings of the existing ultra-short-term prediction model of photovoltaic power, the present invention provides a photovoltaic power probability prediction method which combines wavelet decomposition, deep belief network and genetic algorithm.

[0006] Technical solution: A photovoltaic power ultra-short-term probability prediction method based on wavelet decomposition and optimized deep belief network includes the following steps:

[0007] (1) Clean the historical photovoltaic power data, remove outliers and data exceeding the power threshold, and construct the model input;

[0008] (2) Perform three-level wavelet decomposition on the data to obtain sub-signals of different frequencies;

[0009] (3) Establish corresponding deep belief network prediction models for sub-signals of different frequencies, and use genetic algorithms to optimize the number of hidden layer nodes;

[0010] (4) Obtain the predictions of each sub-model, equally weight them and perform denormalization, and finally obtain the point prediction result Y pre ;

[0011] (5) Analyze the photovoltaic power prediction error, fit the prediction error with a distribution function, and use the t distribution model to fit the error with a probability density to obtain a cumulative probability distribution function;

[0012] (6) Given a confidence level, calculate the prediction interval.

[0013] Furthermore, the numerical weather forecast information in step (1) is highly complex. If inappropriate numerical weather forecasts are introduced, the model will be redundant and the accuracy of the results will be affected. Therefore, a model combining statistical models and deep learning is adopted, and the historical photovoltaic output value is used as input to preprocess the data.

[0014] Furthermore, step (2) uses wavelet decomposition to decompose the photovoltaic power data with strong nonlinearity into a low-frequency signal and three high-frequency signals. The signal subsequence processed by wavelet decomposition will be smoother. The discrete wavelet decomposition function used in step (2) is calculated as follows:

[0015]

[0016] Where p is the scale factor, q is the translation coefficient, g(t) is the sequence to be decomposed, φ(·) is the mother wave function, T is the signal length, and t is the discrete time step.

[0017] Furthermore, in step (3), the corresponding deep belief network models are trained for sub-signals of different frequencies, and the number of hidden layer nodes is optimized using a genetic algorithm. The calculation process is as follows:

[0018] (3.1) Update the state of RBM hidden layer nodes:

[0019]

[0020]

[0021] Where: N j (0,1) conforms to the standard normal distribution, σ∈[0,1]. is the sigmoid function, θ L is the lower bound, θ H This is the upper realm;

[0022] (3.2) The state of the nodes in the visible layer of the restricted Boltzmann machine is updated:

[0023]

[0024] (3.3) The node states of the hidden layer of the restricted Boltzmann machine are updated again:

[0025]

[0026] (3.4) Input the hidden node state of the first restricted Boltzmann machine obtained in step (3.3) into the second restricted Boltzmann machine, and update the state of each node of the second restricted Boltzmann machine according to steps (3.1)-(3.3). Finally, the state update of all nodes in the deep belief network is completed.

[0027] (3.5) When all data have completed the first round of training, the weight matrix W is updated using the following formula:

[0028] W ij (k+1)=W ij (k)+ΔW ij

[0029] (3.6) Return to step (3.1) for the second round of training until the preset maximum number of iterations is reached or ||ΔW(i)||<ε, then stop training. The deep belief network training is now complete.

[0030] (3.7) The fitness function used by the genetic algorithm is expressed as follows:

[0031] F i =1 / E i

[0032] Where: E i is the prediction error after training the i-th sample.

[0033] (3.8) Gene selection. Based on the roulette wheel selection method, the fitness of each individual is sorted from high to low, so as to select individuals with a higher probability of being inherited to offspring, which is proportional to the fitness. The probability calculation formula of individual i being passed on is as follows:

[0034]

[0035] Where: c represents the number of all individuals.

[0036] (3.9) Gene crossover and mutation:

[0037]

[0038] Where: g mj , g nj They each represent the gene at the jth position of the mth and nth individuals before treatment; r∈[0,1].

[0039]

[0040] Where: g min , g max They each represent the upper and lower limits of the j-th gene of the m-th individual; s is the current iteration number; s max is the maximum number of evolutions; r1 takes a random value; r2∈[0,1].

[0041] Furthermore, in step (4), the prediction results obtained by each sub-model are equally weighted to obtain the final prediction result Y pre .

[0042] Furthermore, step (5) performs distribution function fitting on the prediction error to confirm that the error basically obeys t distribution.

[0043] (5.1) The t distribution model is used to fit the probability density of the error. By analyzing the probability distribution fitting curve of the error, the corresponding cumulative probability distribution function can be obtained. The specific calculation formula is as follows:

[0044]

[0045] Where: i is the range of prediction error, m≤i≤n; p i is the probability corresponding to the prediction error.

[0046] (5.2) Set the confidence level to (1-α)×100%, and the prediction interval of PV power is:

[0047]

[0048] Where: is the inverse function of F(x), and p is the predicted power value obtained by point prediction.

[0049] Furthermore, in step (6), the upper and lower limits of each moment are sequentially connected, and the resulting interval is the prediction interval.

[0050] Beneficial effect: Compared with the existing technology, the method of the present invention uses wavelet decomposition technology to decompose the original photovoltaic power generation sequence, establishes an ultra-short-term prediction model for photovoltaic power generation based on a deep belief network model, and uses a genetic algorithm to optimize the model parameters. On this basis, a parameterized photovoltaic power ultra-short-term probability prediction method based on the t-distribution model is proposed, which effectively improves the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a schematic diagram of the process of the method of the present invention;

[0052] Figure 2 This is a diagram of the prediction results of the method described in the present invention in an example application.

[0053] Figure 3 is a probability prediction plot (95% confidence level). DETAILED DESCRIPTION

[0054] In order to illustrate the technical solution disclosed by the present invention in detail, it is further described below in conjunction with the accompanying drawings and specific embodiments.

[0055] like Figure 1 As shown, the ultra-short-term probability prediction method of photovoltaic power based on wavelet decomposition and optimized deep belief network of the present invention comprises the following steps:

[0056] S1. Clean the historical photovoltaic power data, remove abnormal values ​​and data exceeding the power threshold, and build the model input;

[0057] S2, perform three-level wavelet decomposition on the data to obtain sub-signals of different frequencies;

[0058] S3. Establish corresponding deep belief network prediction models for sub-signals of different frequencies, and use genetic algorithm to optimize the number of hidden layer nodes;

[0059] S4, get the predictions of each sub-model, weight them equally and perform denormalization, and finally get the point prediction result Y pre ;

[0060] S5. Analyze the photovoltaic power prediction error, fit the prediction error with a distribution function, and use the t distribution model to fit the error with a probability density to obtain a cumulative probability distribution function;

[0061] S6. Given the confidence level, calculate the prediction interval.

[0062] Among them, due to the excellent feature extraction ability of the deep belief network, it can better obtain and learn the internal nonlinear characteristics of the photovoltaic power generation series; on the other hand, the original photovoltaic power data is processed using wavelet decomposition technology, and the non-stationarity of the processed signal is effectively weakened. After input into the model, its characteristics can be better analyzed, which significantly improves the prediction effect.

[0063] The following is a detailed process of a photovoltaic power ultra-short-term probability prediction method and steps based on wavelet decomposition and optimized deep belief network.

[0064] 1) Setting a threshold comparison to eliminate abnormal zero values ​​and data beyond the power limit range in the original photovoltaic power data;

[0065] 2) Divide the training set and the test set, normalize the data set, and perform three-level wavelet decomposition on the data to obtain sub-signals of different frequencies;

[0066] 3) Train the corresponding deep belief network model for sub-signals of different frequencies respectively, and use the genetic algorithm to find the optimal initial parameters for the model. The calculation process is as follows:

[0067] Restricted Boltzmann machine hidden layer node state update:

[0068]

[0069] Where: N j (0,1) conforms to the standard normal distribution, σ∈[0,1]. is the sigmoid function, θ L is its lower bound, θ H is its upper bound.

[0070] Restricted Boltzmann visible layer node state update:

[0071]

[0072] The node states of the restricted Boltzmann hidden layer are updated again:

[0073]

[0074] Get the hidden node state S' of the first restricted Boltzmann machine j The input is fed into the second restricted Boltzmann machine, and then the states of the nodes of the second restricted Boltzmann machine are updated. Finally, the states of all nodes in the deep belief network are updated.

[0075] When all the training data are input, the first round of training is completed, and the weight matrix W is updated using the following formula:

[0076] W ij (k+1)=W ij (k)+ΔW ij

[0077] Return to the first step for the second round of training until the preset maximum number of iterations is reached or ||ΔW(i)||<ε, then stop training. The deep belief network training is now complete.

[0078] The fitness function used by the genetic algorithm is expressed as follows:

[0079] F i =1 / E i

[0080] Where: E i is the prediction error after training the i-th sample.

[0081] Gene selection: Based on the roulette wheel selection method, the fitness of each individual is sorted from high to low, and individuals with a higher probability of being inherited to offspring are selected, which is proportional to the fitness. The probability calculation formula for individual i being passed on is as follows:

[0082]

[0083] Where: c represents the number of all individuals.

[0084] Genetic crossover and mutation:

[0085]

[0086] Where: g mj , g nj They each represent the gene at the jth position of the mth and nth individuals before treatment; r∈[0,1].

[0087]

[0088] Where: g min , g max They each represent the upper and lower limits of the j-th gene of the m-th individual; s is the current iteration number; s max is the maximum number of evolutions; r1 takes a random value; r2∈[0,1].

[0089] 4) Bring the test set into the trained deep belief network model to obtain the photovoltaic power output prediction value Y pre .

[0090] 5) Analyze the photovoltaic power prediction error, fit the prediction error with a distribution function, use the t distribution model to fit the error with a probability density, and obtain the corresponding cumulative probability distribution function;

[0091] 6) Set the confidence level to (1-α)×100% and calculate the prediction interval.

[0092] At this point, the ultra-short-term power probability prediction of photovoltaic power can be realized by using the above wavelet decomposition and optimized deep belief network. The prediction flow chart is as attached to the manual. Figure 1 shown.

[0093] Example 1

[0094] In order to verify the effectiveness of the method of the present invention, the following experiment was conducted: a total of 960 photovoltaic power data from a photovoltaic power station in Ningxia from 8:00 to 18:00 every day for 24 consecutive days in March were used for prediction, with a time resolution of 15 minutes. The training set and the test set were divided into a ratio of 9:1, and the test results at 90% and 95% confidence levels were output. The prediction results are shown in the attached manual. Figure 2 , Figure 3 shown.

[0095] The deterministic prediction performance is generally evaluated based on the following two indicators: prediction interval coverage probability, prediction interval average width, and comprehensive score.

[0096] The prediction interval coverage probability is defined as follows:

[0097]

[0098] Where: P c is the prediction interval coverage probability; N represents the total number of samples; is the indicative function, when The value is 1, otherwise it is 0.

[0099] The average width of the prediction interval is defined as follows:

[0100]

[0101] Where: P INAW is the average width of the interval prediction; R is the range of the target value, which is normalized using R. INAW The smaller it is, the more sensitive the prediction model is.

[0102] The composite score is defined as follows:

[0103]

[0104] Where: and Represent the upper and lower limits at point t, respectively. is the interval width at t at the confidence level (1-α)×100%, s t is the comprehensive score at t.

[0105] Table 1 Evaluation indicators of prediction results

[0106]

[0107] The distribution of the predicted results is as shown in the attached manual. Figure 2 (Probability prediction results at 90% confidence level), Figure 3 (95% confidence level probability prediction result) is shown. Figure 2 , 3 It can be seen that the photovoltaic output power predicted by the method of the present invention has a good prediction effect. In summary, the present invention can realize the prediction of photovoltaic power and can be used for practical engineering applications.

Claims

1. A photovoltaic power ultra-short-term probability prediction method based on wavelet decomposition and optimized deep belief network, characterized in that: The following steps are involved: (1) Clean the historical photovoltaic power data, remove outliers and data exceeding the power threshold, and construct the model input; (2) Decomposing the photovoltaic power data into a low-frequency signal and three high-frequency signals based on the discrete wavelet decomposition function of the input data to obtain sub-signals of different frequencies; (3) For the four sub-signals of different frequencies, four corresponding deep belief network prediction models composed of multi-layer restricted Boltzmann machines are established respectively, and the number of neuron nodes of each layer of restricted Boltzmann machines is optimized by genetic algorithm to obtain the optimal deep belief network prediction model; The genetic algorithm is used to optimize the number of hidden layer nodes. The calculation process is as follows: (3.1) Restricted Boltzmann machine hidden layer node state update: Where: N j (0,1) conforms to the standard normal distribution, σ∈[0,1], is the sigmoid function, θ L is the lower bound, θ H is the upper bound; (3.2) The state of the nodes in the visible layer of the restricted Boltzmann machine is updated: (3.3) The node states of the hidden layer of the restricted Boltzmann machine are updated again: (3.4) Put step (3.3) to get the hidden node state S' of the first restricted Boltzmann machine j Input into the second restricted Boltzmann machine, and update the status of each node of the second restricted Boltzmann machine according to steps (3.1)-(3.3), and finally complete the status update of all nodes in the deep belief network; (3.5) When all data have completed the first round of training, the weight matrix W is updated using the following formula: IN ij (k+1)=In ij (k)+ΔW ij ; (3.6) Return to step (3.1) for the second round of training until the preset maximum number of iterations is reached or ||ΔW(i)|| < ε, then stop training. This is the end of the deep belief network training. (3.7) The fitness function used by the genetic algorithm is expressed as follows: F i =1 / E i Where: E i is the prediction error after training the i-th sample; (3.8) Gene selection, based on the roulette wheel selection method, sorts the fitness of each individual from high to low, and selects individuals with a higher probability of being inherited to offspring, which is proportional to the fitness. The probability calculation formula for individual i being passed on is as follows: In the formula: c is the number of all individuals; (3.9) Gene crossover and mutation: Where: g mj , g nj are the genes at the jth position of the mth and nth individuals before treatment; r∈[0,1], Where: g min , g max are the upper and lower limits of the j-th gene of the m-th individual; s is the current iteration number; s max is the maximum number of evolutions; r1 is random; r2∈[0,1]; (4) Obtain the predictions of each sub-model, equally weight them and perform denormalization, and finally obtain the point prediction result Y pre ; (5) Analyze the photovoltaic power prediction error, fit the prediction error with a distribution function, use the t distribution model to fit the error probability density, and obtain the cumulative probability distribution function; (6) Given a confidence level, calculate the prediction interval.

2. The photovoltaic power ultra-short-term probability prediction method based on wavelet decomposition and optimized deep belief network according to claim 1 is characterized in that: In step (1), the historical data includes weather forecast numerical information. A statistical model and a deep learning model are combined to preprocess the data using the historical photovoltaic output value as input.

3. The photovoltaic power ultra-short-term probability prediction method based on wavelet decomposition and optimized deep belief network according to claim 1 is characterized in that: Step (2) performs three-level wavelet decomposition on the data to obtain sub-signals of different frequencies, including using wavelet decomposition to decompose the photovoltaic power data with strong nonlinearity into a low-frequency signal and three high-frequency signals. The signal sub-sequence processed by wavelet decomposition will be smoother. The discrete wavelet decomposition function used in step (2) is calculated as follows: Where p is the scale factor, q is the translation coefficient, g(t) is the sequence to be decomposed, φ(·) is the mother wave function, T is the signal length, and t is the discrete time step.

4. The photovoltaic power ultra-short-term probability prediction method based on wavelet decomposition and optimized deep belief network according to claim 1 is characterized in that: Step (4) obtains the predictions of each sub-model, weights them equally and performs denormalization, and finally obtains the point prediction result Y pre , including equally weighting the prediction results obtained by each sub-model and performing denormalization to obtain the final point prediction result Y pre .

5. The photovoltaic power ultra-short-term probability prediction method based on wavelet decomposition and optimized deep belief network according to claim 1 is characterized in that: Step (5) analyzes the photovoltaic power prediction error, performs distribution function fitting on the prediction error, uses the t distribution model to perform probability density fitting on the error, and obtains a cumulative probability distribution function, which specifically includes: (5.1) The t Location-Scale distribution model is used to fit the probability density of the error. By analyzing the probability distribution fitting curve of the error, the corresponding cumulative probability distribution function can be obtained. The specific calculation formula is as follows: Where: i is the range of prediction error, m≤i≤n; p i is the probability corresponding to the prediction error; (5.2) Set the confidence level to (1-α)×100%, and the prediction interval of PV power is: Where: is the inverse function of F(x), and p is the predicted power value obtained by point prediction.

6. The photovoltaic power ultra-short-term probability prediction method based on wavelet decomposition and optimized deep belief network according to claim 1 is characterized in that: Step (6) Given the confidence level, the prediction interval is calculated, which specifically includes connecting the upper and lower limits of each moment in sequence, and the resulting interval is the prediction interval.

Images