Photovoltaic power short-term probability prediction system and method

Abstract

The present invention discloses a photovoltaic power short-term probability prediction system and method, comprising: a convolutional neural network, a timing network and a probability density estimation module; the timing network is composed of a self-organizing map neural network, a timing encoder and a timing decoder; the timing encoder and the timing decoder are both composed of an autocorrelation mechanism unit, a timing decomposition unit and a feedforward layer alternately connected; the timing decoder is composed of an upper branch sublayer and a lower branch sublayer; the convolutional neural network screens meteorological factors to obtain a high-dimensional mapping relationship between meteorological characteristics and optical power; the timing network performs progressive decomposition and autocorrelation extraction on the optical power timing sequence to obtain the periodicity and trend characteristics of the optical power sequence; the probability density estimation module optimizes and combines the high-dimensional mapping relationship with the periodicity and trend characteristics of the optical power sequence to obtain the output of optical power probability distribution parameters; the present invention integrates numerical weather forecast information with the weather type forecast issued by the meteorological station to realize optical power probability prediction, thereby effectively reducing the risk of false alarms from a single meteorological source.

Description

Technical Field

[0001] The present invention belongs to the field of photovoltaic prediction applications, and in particular relates to a photovoltaic power short-term probability prediction system and method. Background Art

[0002] To achieve the strategic goals of "dual carbon emissions," photovoltaic (PV) installations have significantly increased their installed capacity and share of power generation. Large-scale centralized PV installations, owing to their ability to centrally develop and efficiently manage sunlight resources, are accounting for an increasing proportion of newly installed PV capacity. However, the integration of centralized PV into the grid increases uncertainty in the power system, necessitating the day-ahead deployment of conventional generators to provide backup capacity to mitigate the impact on safe and stable operation. Therefore, accurately quantifying the uncertainty associated with centralized PV installations is crucial for the economic and stable operation of the power system. Day-ahead forecasts are a crucial basis for power generation planning and power market transactions. Day-ahead forecasts can be categorized as deterministic or probabilistic. Deterministic forecasts output a single point in the future, the expected value of the forecasted object. However, due to the chaotic nature of the atmospheric system, forecast errors are difficult to eliminate, and the forecast results often exhibit high uncertainty. Probabilistic forecasting techniques quantify forecast uncertainty by estimating its probability distribution, enabling better application in power system optimization, scheduling, market transactions, and other areas.

[0003] Currently, there is a lot of research being conducted domestically and internationally on probabilistic prediction methods, including statistical methods and deep learning methods. Statistical methods require a relatively high amount of historical prediction data and have difficulty modeling the coupled effects of multiple factors. Recurrent neural networks and their variants in deep learning are widely used in power prediction, but the structure of recurrent networks severely limits the model depth and its ability to learn from data. Based on these two methods, domestic and international scholars have achieved rich research results in the short-term probabilistic prediction of optical power. However, there are still deficiencies in balancing global and local characteristics, breaking the bottleneck of information utilization, and improving the accuracy and generalization ability of the model. Summary of the Invention

[0004] In response to the above problems, the purpose of the present invention is to overcome the shortcomings of existing optical power short-term probability prediction methods, such as their failure to fully utilize multi-source meteorological information and their weak generalization ability. First, the Pearson correlation coefficient method is used to screen meteorological factors related to optical power, and a high-dimensional mapping relationship between meteorological characteristics and optical power is established based on a convolutional neural network. Then, an improved time series autoformer network is constructed to perform progressive decomposition and autocorrelation extraction on the optical power time series, and analyze its internal structure and inherent laws. Finally, the hidden layer information output by the CNN-Autoformer network is input into the probability density estimation layer, and the optical power probability distribution parameters are obtained through the output of the fully connected layer.

[0005] In order to solve the problems of the prior art, the present invention adopts the following technical solutions:

[0006] A photovoltaic power short-term probability prediction system, characterized in that the prediction system includes: a convolutional neural network, a time series network and a probability density estimation module; the time series network is composed of a self-organizing map neural network, a time series encoder and a time series decoder; the time series encoder and the time series decoder are both composed of an autocorrelation mechanism unit, a time series decomposition unit and a feedforward layer alternately connected; the time series decoder is composed of an upper branch sublayer and a lower branch sublayer; wherein:

[0007] The convolutional neural network screens meteorological factors to obtain a high-dimensional mapping relationship between meteorological characteristics and optical power;

[0008] The timing network performs progressive decomposition and autocorrelation extraction on the optical power timing sequence to obtain the periodicity and trend characteristics of the optical power sequence; wherein:

[0009] The self-organizing map neural network clusters the weather forecast types to obtain generalized weather types, and encodes them in a time series encoder as discrete features of the optical power daily series;

[0010] The timing encoder alternately decomposes the optical power daily sequence to obtain periodic information;

[0011] The timing decoder refines periodic information through an autocorrelation mechanism unit; comprising:

[0012] The upper branch sub-layer extracts periodic features, and the lower branch sub-layer uses weighted addition to extract trend features.

[0013] The probability density estimation module optimizes and combines the high-dimensional mapping relationship with the periodicity and trend characteristics of the optical power sequence to obtain the optical power probability distribution parameter output.

[0014] Furthermore, the timing network performs progressive decomposition and autocorrelation extraction on the optical power timing sequence to obtain the periodicity and trend characteristics of the optical power sequence; including:

[0015] By using the discrete optical power time series {X t}Establish timing dependencies according to the following formula:

[0016]

[0017] Among them, L represents the sequence length, and the autocorrelation coefficient R xx (τ) represents the sequence {X t} with its τ delay {X t-τ}, after the calculation is completed, the most likely top k time series τ1, τ2, ..., τk ;in:

[0018] By using the discrete optical power time series {X t The timing dependency is efficiently solved by Fourier transform according to the following formula:

[0019]

[0020] Among them, S xx (f) represents the autocorrelation coefficient in the frequency domain, and represent the fast Fourier transform and its inverse transform, respectively. express The conjugate part in the frequency domain calculation; τ∈{1,2,...,L};

[0021] The latency information aggregation relationship is established based on the timing dependency according to the following formula:

[0022]

[0023] Among them, argTopK(·) aims to obtain the parameter set of the top k autocorrelations; c is a hyperparameter; R Q,K is the autocorrelation result of the query Q and key K after the softmax operation; Roll(X,τ) represents the calculation process of X with a time delay τ, and Auto-Correlation represents the calculation process of the autocorrelation coefficient.

[0024] Furthermore, the self-organizing map neural network clusters the weather forecast types to obtain generalized weather types, and encodes them as discrete features of the optical power daily series in a time series encoder, including:

[0025] By performing wavelet transform on the historical optical power daily series, we obtain an approximate signal representing the change trend as the clear sky series u for that day, and a detail signal representing the random fluctuation as the fluctuation series v for that day. On this basis, we obtain the amplitude and fluctuation characteristics of the optical power series:

[0026] The optical power amplitude is the average value of the clear sky sequence u To measure, the calculation is as follows:

[0027]

[0028] Where u i is the i-th sampling value of u; n1 and n2 are the sampling start and end times respectively;

[0029] The optical power fluctuation amplitude adopts the first-order difference index △ of the fluctuation sequence v f express:

[0030]

[0031] Where, v i is the i-th sampling value of v;

[0032] The frequency of optical power fluctuations requires fast Fourier transform of the fluctuation sequence v to determine its central frequency. Using the amplitude and fluctuation characteristics of the daily optical power series as input variables, a stepwise binary split is performed based on the SOM neural network clustering algorithm to obtain D types of generalized weather types. The number of D types is determined by the Davidson-Botting index.

[0033] Statistically classify the daily optical power series under historical weather forecast types, and use the most frequently occurring generalized weather types to correspond to the original weather forecast types, thereby reducing and classifying the weather forecast types.

[0034] One-hot encoding is used to encode the generalized weather type into discrete features of the daily optical power series in the temporal encoder.

[0035] Furthermore, the timing encoder alternately decomposes the optical power daily sequence to obtain period information; including:

[0036] The temporal encoder is composed of N encoding layers, where the overall equation of the lth encoder layer can be expressed as:

[0037]

[0038] The temporal encoder gradually eliminates the trend component according to the following formula. The input of the encoder module is a length of L seq Optical power timing sequence (d represents dimension); get the periodic component

[0039]

[0040] Among them, “_” represents the trend component that is eliminated. represents the output of the l-th layer encoder, They represent the periodic components output by the i-th sequence decomposition unit in the l-th layer respectively.

[0041] Furthermore, the time series decoder refines the periodic information through the autocorrelation mechanism unit; including:

[0042] The decoder consists of M decoding layers, where the overall equation of the lth decoder layer can be expressed as:

[0043]

[0044] The upper branch processing unit performs alternating decomposition according to the following formula to obtain periodic components;

[0045]

[0046] Where, represents the output of the l-th layer decoder, They represent the period component and trend component output by the i-th sequence decomposition unit in the l-th layer, SeriesDecomp represents the time series decomposition module, and FeedForward represents the feedforward module;

[0047] The lower branch sublayer obtains the trend component by weighted addition based on the output of the upper branch sublayer

[0048]

[0049] Where w l,i ,i∈{1,2,3} represents the trend component extracted for the i-th time projection.

[0050] Furthermore, the probability density estimation module optimizes and combines the high-dimensional mapping relationship with the periodicity and trend characteristics of the optical power sequence to obtain the optical power probability distribution parameter output, including:

[0051] Based on the optimization combination of high-dimensional mapping relationship and periodicity and trend characteristics of optical power sequence, assuming that the observed data is x, the maximum likelihood function of Gaussian distribution parameters is constructed according to the following formula:

[0052]

[0053] In the formula, μ and σ represent the expected value and standard deviation of Gaussian distribution respectively, n represents the number of samples, argmax represents the maximum likelihood estimate, and X i represents the i-th sample;

[0054] According to the maximum likelihood function, the optical power probability distribution parameters are obtained according to the following formula:

[0055]

[0056] Where: p is the output power data in the training data, and is the estimated parameter output by the probability density estimation layer.

[0057] The present invention is also implemented by the following technical solutions:

[0058] A photovoltaic power short-term probability prediction method comprises the following steps:

[0059] The meteorological factors are filtered through convolutional neural networks to obtain a high-dimensional mapping relationship between meteorological characteristics and optical power;

[0060] The optical power sequence level link relationship is obtained by progressively decomposing the optical power sequence and extracting the autocorrelation through the timing network; wherein:

[0061] The discrete features of the daily optical power series are obtained by clustering the generalized weather types using a self-organizing map neural network.

[0062] The optical power daily series is alternately decomposed by a time series encoder to obtain periodic information;

[0063] The periodic information is refined by a timing decoder; wherein:

[0064] The upper branch sub-layer extracts periodic features, and the lower branch sub-layer uses weighted addition to extract trend features.

[0065] The probability density estimation module is used to optimize the high-dimensional mapping relationship and the optical power sequence-level link relationship to obtain the optical power probability distribution parameter output.

[0066] Beneficial effects

[0067] Compared with the prior art, the present invention has the following advantages:

[0068] (1) The proposed short-term probability prediction method for optical power based on the improved CNN-Autoformer network takes into account the impact of weather types on optical power output, classifies weather types and discretizes them into optical power daily series features, and realizes optical power probability prediction through the Autoformer network by integrating numerical weather forecast information, thereby effectively reducing the risk of false alarms from a single meteorological source and significantly improving the prediction accuracy of the model.

[0069] (2) A progressive decomposition architecture was established to achieve the goal that during the prediction process, the model alternately optimizes the prediction results and decomposes the sequence. In combination with the autocorrelation mechanism, the dependency relationship based on different periods is captured, and the optical power sequence level link is realized, thereby improving the efficiency of information utilization.

[0070] (3) By combining maximum likelihood estimation with gradient optimization and outputting optical power probability distribution parameters, the optical power interval prediction results and deterministic prediction results can be obtained simultaneously. Compared with the heuristic-based optimization algorithm, the time complexity is reduced and the neural network is effectively prevented from falling into the local optimum. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a short-term probability prediction architecture of optical power in a specific embodiment of the present invention;

[0072] Figure 2It is a convolutional neural network structure in a specific embodiment of the present invention;

[0073] Figure 3 It is the autocorrelation mechanism structure in the specific implementation mode of the present invention;

[0074] Figure 4 is the DBI under different numbers of categories in the embodiment of the present invention;

[0075] Figure 5 is the 90% confidence interval prediction result of each model in the embodiment of the present invention;

[0076] Figure 6 It is the deterministic prediction result of each model in the embodiment of the present invention.

[0077] Table 1 shows the model setting parameters in the embodiment of the present invention.

[0078] Table 2 is the monthly average value of each model evaluation index in the embodiment of the present invention DETAILED DESCRIPTION

[0079] The following is combined with Figure 1 The present invention is described as follows:

[0080] The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0081] The present invention constructs a short-term optical power probability prediction model based on an improved CNN-Autoformer under a "deep decomposition architecture". The overall architecture is as follows: Figure 1 As shown, it mainly includes 3 parts:

[0082] 1) High-dimensional meteorological feature extraction for numerical weather forecasting based on convolutional neural networks (CNN): The Pearson correlation coefficient method is used to screen meteorological factors related to optical power, such as irradiance, temperature, and humidity. Meteorological factors with strong correlation are used as CNN input, and the convolution layer is used to capture the temporal characteristics of a single meteorological factor and the mutual influence between different meteorological factors. The pooling layer is used to reduce the dimension of the output features of the convolution layer and the number of parameters while retaining important features to prevent model overfitting. The features extracted by the convolution and pooling layers are integrated and mapped through a fully connected layer, thereby establishing a high-dimensional mapping relationship between the extracted meteorological features and optical power.

[0083] 2) Progressive Decomposition and Autocorrelation Feature Extraction Based on an Improved Time Series Autoformer Network: To deeply extract the periodicity, trend, and other characteristics of the optical power time series and reduce the impact of random noise, an improved time series autoformer network is used for feature extraction. Decomposition of the optical power time series is required to extract the variation patterns of different periods and analyze the internal structure and inherent patterns of the time series.

[0084] 3) Probability distribution parameter optimization based on maximum likelihood estimation: Combining maximum likelihood estimation with the gradient optimization method in machine learning, the hidden layer information output by the CNN-Autoformer network is input into the probability density estimation layer, and the optical power probability distribution parameters are obtained through the output of the fully connected layer.

[0085] 1. Meteorological feature extraction

[0086] The typical structure of CNN network is as follows Figure 2 As shown in Figure 2. The main function of the convolutional layer is to slide the convolution kernel across different locations in the input data to calculate the feature vector, thereby deeply extracting the characteristics of the meteorological data. The pooling layer typically follows the convolutional layer to reduce the feature dimension while retaining the most significant features, which accelerates the convergence of the CNN network. At the end of the CNN network is the fully connected layer, which flattens the features extracted by the convolutional and pooling layers and outputs the high-dimensional meteorological features for numerical weather forecasting.

[0087] 2. Temporal Autoformer Network

[0088] (1) Discrete feature coding of weather types

[0089] To improve model learning efficiency and avoid underfitting caused by too few samples for a single feature, it is necessary to reduce and categorize weather forecast types to obtain generalized weather types. A SOM neural network is used to cluster the daily optical power series. One-hot encoding is then used to encode the generalized weather types into discrete features of the daily optical power series, improving the Autoformer network's ability to identify different weather types.

[0090] (2) Deep Decomposition Architecture

[0091] The present invention embeds sequence decomposition as an internal unit of the Autoformer into the encoder-decoder. During the prediction process, the model alternately optimizes the prediction results and decomposes the sequence, gradually separating the trend term and the periodic term from the latent variables to achieve progressive decomposition.

[0092] In the encoder part, assuming there are N encoder layers, the overall equation of the lth encoder layer can be expressed as: The encoder obtains the periodic component by gradually eliminating the trend component

[0093]

[0094] In the formula, “_” represents the trend component that is eliminated, represents the output of the l-th layer encoder, They represent the periodic components output by the i-th sequence decomposition unit in the l-th layer respectively.

[0095] In the decoder part, assuming there are M decoder layers, the overall equation of the lth decoder layer can be expressed as: Each decoder layer contains an internal autocorrelation mechanism and an encoder-decoder autocorrelation mechanism, which can respectively refine the prediction and utilize the decoder's periodic component information. The decoder uses a two-way processing mode. The upper branch processes the period component, and the lower branch uses weighted addition based on the output of the upper branch sub-layer to obtain the trend component. The specific process is as follows:

[0096]

[0097] Where, represents the output of the l-th layer decoder; They represent the period component and trend component of the output of the i-th sequence decomposition unit in the l-th layer, respectively, and w l,i ,i∈{1,2,3} represents the trend component extracted for the i-th time projection.

[0098] (3) Autocorrelation mechanism

[0099] Optical power sequences usually exhibit similar sub-processes under similar phases. Therefore, the autocorrelation mechanism of the present invention replaces the traditional self-attention mechanism to enhance the feature extraction capability of similar weather processes and adjacent days.

[0100] Time series dependency: The modeling method of the present invention is based on autonomous modeling of historical data, where the current trend is highly correlated with the historical time series. The current time pattern can be frequently observed in the historical subsequence. For a discrete optical power time series {X t}, in order to model the correlation between its time series, the autocorrelation coefficient R is proposed xx (τ), can be calculated by the following formula:

[0101]

[0102] In the formula, the autocorrelation coefficient R xx (τ) represents the sequence {X t} with its τ delay {X t-τ}, after the calculation is completed, the most likely top k time series τ1, τ2, ..., τ k .

[0103] Delay information aggregation: The proposed time series dependency feature can quantify the correlation between the historical subsequence and the current estimated sequence. On this basis, the delay information aggregation module is used to select the first k delay sequences τ1, τ2, ..., τ k The sequence with a time delay of τ is rolled over to the current phase period. Finally, all subsequences are aggregated and merged to perform a softmax operation. This whole process is different from the traditional dot product aggregation mechanism based on the self-attention mechanism.

[0104] To calculate the autocorrelation coefficient, a multi-head mechanism is also used, similar to self-attention. For one head, assuming a time series of length L, similar to the self-attention mechanism, after the high-dimensional projection process, Q, K, and V can be obtained. Self-attention can be replaced, and the autocorrelation can be expressed as follows:

[0105]

[0106] Where argTopK(·) is intended to obtain the parameter set of the top k autocorrelations. In the present invention, Where c is a hyperparameter. Q,K is the autocorrelation result of the query Q and key K after the softmax operation. Roll(X,τ) represents the calculation process of X with a time delay τ, where the current sequence being processed will be shifted by one time step as the similarity calculation is performed until the sequence moves to the τ lag time step.

[0107] In order to speed up the calculation process, the time lag information of Roll(X,τ) will be stored in a matrix, and the similarity calculation is only performed once through the matrix dot product. As for the autocorrelation between the encoder and decoder, similar to self-attention, the key K and value V are the encoder The output of the query Q comes from the previous block of the decoder.

[0108] Therefore, suppose the query, key, and value of the i-th head are where i∈{1,2,...,h}, L seq Indicates the sequence length, with d model The multi-head autocorrelation of channels and h heads is calculated as follows:

[0109] Muti-Head(Q,K,V)=W output *Concat(head1,head2,...,head n ) (9)

[0110] Among them, W output Indicates output projection, Concat indicates connection operation, head i Calculate according to the following formula:

[0111] head i =Auto-Correlation(Q i ,K i ,V i ) (10)

[0112] Fast calculation: To improve computational efficiency, as shown in Equations (11)-(12), the proposed autocorrelation mechanism and Fast Fourier Transforms (FFT) calculation mechanism can reduce the computational complexity to O(LlogL). Similarly, a memory cost of O(LlogL) can also be achieved. FFT can transform continuous data in the time domain into simpler discrete frequency features in the frequency domain, allowing the Autoformer to process long time series. For the time series {X t}, its autocorrelation coefficient R xx (τ) can be expressed as Figure 3 The calculation is performed by FFT, and the process is as follows:

[0113]

[0114] Among them, S xx (f) represents the autocorrelation coefficient in the frequency domain, and represent the fast Fourier transform and its inverse transform, respectively. express The conjugate part of the frequency domain calculation. τ∈{1,2,...,L}. Furthermore, due to the matrix dot product, the autocorrelation process only needs to be calculated once by using the Fast Fourier Transform (FFT). Therefore, the complexity of the autocorrelation mechanism is O(LlogL).

[0115] 3. Probability Density Estimation Layer

[0116] The present invention combines maximum likelihood estimation with gradient optimization of deep learning. By training sample observations, the most likely parameter values ​​of the probability density distribution function are inferred. The probability density distribution function can also be used to obtain the prediction interval and deterministic prediction value under the specified coverage probability.

[0117] Taking Gaussian distribution as an example, the observed data is x, and the maximum likelihood function of Gaussian distribution parameters can be constructed as:

[0118]

[0119] In the formula, μ and σ represent the expected value and standard deviation of Gaussian distribution respectively, n represents the number of samples, argmax represents the maximum likelihood estimate, and X i represents the i-th sample;

[0120] By finding the extreme value of the above formula, we can obtain the optimal Gaussian distribution parameters under the sample observation value. After substituting the probability density function into (13), we take the logarithm of both sides to obtain the final loss function:

[0121]

[0122] Where p is the output power data in the training data, and is the estimated parameter output by the probability density estimation layer.

[0123] 4. Examples

[0124] This example is based on a photovoltaic power plant. The plant's energy management system contains optical power data, numerical weather forecast data, and weather type forecast data from January 1, 2022, to December 31, 2023. The first 70% of this dataset is used as the training set, the next 20% as the validation set, and the remaining 10% as the test set. The parameter settings are shown in Table 1.

[0125] Table 1

[0126]

[0127] (1) General classification of weather types

[0128] The optical power daily series features are extracted and the SOM neural network clustering algorithm is used for step-by-step classification. The number of classifications N is determined by DBI. The smaller the DBI, the better the classification effect. Figure 5 shown.

[0129] Theoretically, the more types of generalized weather types there are, the more refined the classification model will be, but at the same time, the complexity of the model will increase. Figure 4 It can be seen that when N=5, DBI has a significant inflection point, so the weather forecast weather types are further divided into five generalized weather types.

[0130] (2) Model evaluation

[0131] Two models are used for comparative experiments, namely LSTM and GRU.

[0132] Interval prediction usually uses the prediction interval coverage probability (PICP) to measure the coverage quality of the prediction interval, which means the probability that the measured value falls within the interval, and is defined as follows:

[0133]

[0134] PICE=|PINC-PICP| (16)

[0135] Where n represents the number of calculation points, [L i ,U i ] represents the interval predicted by point i, y i represents the measured value of the i-th point, PINC is the given confidence level, and PI CE is the interval coverage error. The smaller the PI CE, the better the interval prediction effect.

[0136] The normalized average width (PI NAW) is used to measure the quality of interval width, which is defined as follows:

[0137]

[0138] Where R represents the difference between the maximum and minimum power values ​​measured on that day;

[0139] In addition, the normalized root mean square error (NRMSE) and normalized mean absolute error (NMAE) are used to evaluate the output deterministic prediction results:

[0140]

[0141] In the above formula, P i is the actual power, Y i is the predicted power, P capcity The optical power data is normalized to the installed capacity, P capcity Take it as 1.

[0142] (3) Comparative experimental results

[0143] Compared to traditional interval prediction methods, the probability density estimation method proposed in this paper not only obtains the probability density distribution but also obtains deterministic prediction results through the predicted expected value. The proposed CNN-Autoformer model, along with the GRU model and the LSTM model, was used to perform 90% confidence interval and deterministic predictions for a consecutive month, with a randomly selected day being visualized for analysis.

[0144] Figure 5The 90% confidence interval prediction results of the three models are shown. Figure 6 The deterministic prediction results of three models are shown.

[0145] from Figure 5 and Figure 6 As can be seen, the CNN-Autoformer model has the narrowest prediction interval width and the best tracking performance during the PV ramp-up and ramp-down phases. During the stable PV output phase, the GRU model has the narrowest 90% confidence interval width, followed by the CNN-Autoformer model, and the LSTM model has the widest. However, the GRU model's narrow interval width causes the measured optical power values ​​to exceed the lower limit of the 90% confidence interval between 11:30 and 1:00 PM. The CNN-Autoformer model's measured optical power briefly exceeds the lower limit of the 90% confidence interval at 11:30 and 12:15 PM, while the LSTM model maintains the widest interval width during the noon period and never exceeds the 90% confidence interval. After 2:00 PM, PV output began to slowly decline. The CNN-Autoformer model's deterministic prediction results were the best, and the measured optical power closely followed the deterministic prediction results. In contrast, the measured optical power curves of the GRU and LSTM models both exceeded the lower limit of the 90% confidence interval. The reason is that the CNN-Autoformer model, on the one hand, captures the pattern of irradiance drop in the evening period in meteorological information through CNN, and on the other hand, learns the pattern of light power change on adjacent and similar days through the autocorrelation mechanism, thereby improving the model's prediction performance.

[0146] Table 2 shows the monthly root mean square error, monthly average absolute error, and monthly average coverage error and monthly average coverage width of the deterministic prediction of the model proposed in this invention and the two comparative models within one month. As can be seen from Table 2, in the comparison of the four evaluation indicators, the model proposed in this invention performed the best, followed by the GRU model, and the LSTM model performed the worst. In terms of deterministic prediction, the monthly root mean square error of the model proposed in this invention was reduced by 3.9% compared with the GRU model, and the monthly average absolute error was reduced by 3.91%; in terms of 90% confidence interval prediction, the monthly average coverage error of the model proposed in this invention was reduced by 3.55% compared with the GRU model, and the monthly average coverage width was reduced by 11.32%. As shown below:

[0147] Table 2

[0148]

[0149]

[0150] (4) Conclusion

[0151] The case results show that:

[0152] 1) The proposed short-term probabilistic optical power prediction method based on the improved CNN-Autoformer network takes into account the impact of weather types on optical power output, classifies weather types and discretely encodes them into daily optical power series features, integrates numerical weather forecast information, effectively reduces the risk of false alarms from a single meteorological source, and significantly improves the prediction accuracy of the model.

[0153] 2) The 30-day simulation results show that compared with the GRU model, the proposed method improves the NRMSE, NMAE, PICE, and PINAW evaluation indicators by 3.90%, 3.91%, 3.55%, and 11.32%, respectively; compared with the LSTM model, the proposed method improves the NRMSE, NMAE, PICE, and PINAW evaluation indicators by 5.08%, 5.06%, 18.99%, and 19.33%, respectively, verifying the effectiveness of the proposed method.

[0154] 3) The method proposed in this invention can be used for short-term power prediction of centralized photovoltaic power stations.

[0155] Although the present invention has been described above, the present invention is not limited to the above-mentioned specific embodiments. The above-mentioned specific embodiments are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can make many variations without departing from the purpose of the present invention, and these are all protected by the present invention.

Claims

1. A photovoltaic power short-term probability prediction system, characterized in that: The prediction system includes: a convolutional neural network, a time series network and a probability density estimation module; the time series network is composed of a self-organizing map neural network, a time series encoder and a time series decoder; the time series encoder and the time series decoder are both composed of an autocorrelation mechanism unit, a time series decomposition unit and a feedforward layer alternately connected; the time series decoder is composed of an upper branch sublayer and a lower branch sublayer; wherein: The convolutional neural network screens meteorological factors to obtain a high-dimensional mapping relationship between meteorological characteristics and optical power; The timing network performs progressive decomposition and autocorrelation extraction on the optical power timing sequence to obtain the periodicity and trend characteristics of the optical power sequence; wherein: By using the discrete optical power time series {X t }Establish timing dependencies according to the following formula: Among them, L represents the sequence length, and the autocorrelation coefficient R xx (τ) represents the sequence {X t } with its τ delay {X t-τ }, after the calculation is completed, the most likely top k time series τ1, τ2, ..., τ k ;in: By using the discrete optical power time series {X t The timing dependency is efficiently solved by Fourier transform according to the following formula: Among them, S xx (f) represents the autocorrelation coefficient in the frequency domain, and represent the fast Fourier transform and its inverse transform, respectively. express The conjugate part in the frequency domain calculation; τ∈{1,2,...,L}; The latency information aggregation relationship is established based on the timing dependency according to the following formula: Among them, argTopK(·) aims to obtain the parameter set of the top k autocorrelations; c is a hyperparameter; R Q,K is the autocorrelation result of the query Q and key K after the softmax operation; Roll(X,τ) represents the calculation process of X with a time delay τ, and Auto-Correlation represents the calculation process of the autocorrelation coefficient; The self-organizing map neural network clusters the weather forecast types to obtain generalized weather types, and encodes them in a time series encoder as discrete features of the optical power daily series; comprising: By performing wavelet transform on the historical daily optical power series, we obtain an approximate signal representing the change trend as the clear sky series u, and a detail signal representing random fluctuations as the fluctuation series v. On this basis, we obtain the amplitude and fluctuation characteristics of the optical power series: The optical power amplitude is the average value of the clear sky sequence u To measure, the calculation is as follows: Where u i is the i-th sampling value of u; n1 and n2 are the sampling start and end times respectively; The optical power fluctuation amplitude adopts the first-order difference index △ of the fluctuation sequence v f express: Where, v i is the i-th sampling value of v; The frequency of optical power fluctuations requires fast Fourier transform of the fluctuation sequence v to determine its central frequency. Using the amplitude and fluctuation characteristics of the daily optical power series as input variables, a stepwise binary split is performed based on the SOM neural network clustering algorithm to obtain D types of generalized weather types. The number of D types is determined by the Davidson-Botting index. Statistically classify the daily optical power series under historical weather forecast types, and use the most frequently occurring generalized weather types to correspond to the original weather forecast types, thereby reducing and classifying the weather forecast types. One-hot encoding is used to encode the generalized weather type into discrete features of the daily optical power series in the temporal encoder. The time series encoder alternately decomposes the optical power daily sequence to obtain periodic information; including: The temporal encoder is composed of N encoding layers, where the overall equation of the lth encoder layer can be expressed as: The temporal encoder gradually eliminates the trend component according to the following formula. The input of the encoder module is a length of L seq Optical power timing sequence Get the periodic component The temporal decoder consists of M decoding layers, where the overall equation of the lth decoder layer can be expressed as: The upper branch sublayer is alternately decomposed according to the following formula to obtain periodic components; Where, represents the output of the l-th layer decoder, They represent the period component and trend component output by the i-th sequence decomposition unit in the l-th layer, SeriesDecomp represents the time series decomposition module, and FeedForward represents the feedforward module; The lower branch sublayer obtains the trend component by weighted addition based on the output of the upper branch sublayer Where w l,i ,i∈{1,2,3} represents the trend component extracted for the i-th time projection; The probability density estimation module optimizes and combines the high-dimensional mapping relationship with the periodicity and trend characteristics of the optical power sequence to obtain the optical power probability distribution parameter output; including: Based on the optimization combination of high-dimensional mapping relationship and periodicity and trend characteristics of optical power sequence, the observation data is x, and the maximum likelihood function of Gaussian distribution parameters is constructed according to the following formula: In the formula, μ and σ represent the expected value and standard deviation of Gaussian distribution respectively, n represents the number of samples, arg max represents the maximum likelihood estimate, X i represents the i-th sample; According to the maximum likelihood function, the optical power probability distribution parameters are obtained according to the following formula: Where: p is the output power data in the training data, and is the estimated parameter output by the probability density estimation layer.

Images