Low weather resolution photovoltaic prediction method and system based on similar day training set

Abstract

The present invention proposes a low weather resolution photovoltaic prediction method and system based on a similar day training set, belonging to the technical field of photovoltaic prediction, including: obtaining weather forecast data of a day to be predicted; constructing the nearest similar day similar to the weather forecast data of the day to be predicted, and obtaining the meteorological data and photovoltaic power generation data of this day; reconstructing a similar day training set; decomposing the similar day training set into two modal components, establishing a hybrid deep learning model for the two modal components for learning and training. When predicting, use the photovoltaic output of the modal components of the last set number of days of the similar day training set plus the weather forecast data of the day to be predicted to infer the modal component output corresponding to the two modes, and add the two to obtain a trend photovoltaic output prediction curve; obtain a volatility prediction curve based on the similar day training set; perform weighted averaging on the obtained volatility prediction curve and trend prediction curve to obtain the final prediction result.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic prediction, and particularly relates to a low weather resolution photovoltaic prediction method and system based on a similar day training set. Background Art

[0002] The statements in this part only provide background technical information related to the present invention, and do not necessarily constitute prior art.

[0003] Fossil fuels have long been the cornerstone of global economic growth, but their overuse has caused serious environmental problems, forcing the world to turn to exploring cleaner, lower-carbon and more sustainable energy sources. Among these energy sources, photovoltaic power generation stands out due to its advantages such as low cost, strong sustainability and good environmental benefits, and has achieved rapid growth globally. The global photovoltaic installed capacity has increased from 400GW in 2010 to 843GW in 2021, and is expected to reach 1700GW by 2030.

[0004] Although photovoltaic energy has many advantages, its high dependence on weather makes its output highly unstable, posing a huge challenge to the safe and stable operation of the power grid. This situation has become more serious as more and more distributed photovoltaic systems are connected to the power grid. Accurate prediction of current-day photovoltaic output is crucial for maintaining the security, stability and reliability of the power grid. Compared with 1-hour prediction, 15-minute interval prediction provides a finer granularity, which is more conducive to improving the scheduling of power systems, energy management plans and assisting photovoltaics to participate in power market transactions. However, photovoltaic power generation is highly sensitive to frequent and rapid changes in weather, which increases the uncertainty and volatility of photovoltaic output, and this phenomenon is particularly obvious in small distributed photovoltaic power generation stations. In addition, due to cost considerations, a large number of distributed stations do not have access to high-precision, fine-grained weather information, and the weather information that can be obtained at low cost is mostly at an hourly interval, which further increases the difficulty of predicting the current-day 96-point photovoltaic output.

[0005] The current-day photovoltaic output prediction is essentially a time series prediction problem. The current existing methods generally train and tune parameters based on deep learning models after dividing the training set and the validation set, and finally output the prediction value. However, the current methods are difficult to achieve the required effect under coarse-grained weather conditions, and the distributed photovoltaic power stations are severely punished by assessment.

[0006] The existing technologies for current-day photovoltaic output prediction can only solve the situation when the sampling granularity of weather conditions is the same as the required prediction granularity, and the prediction accuracy is not strong enough. Therefore, it is necessary to solve the problem of how to achieve accurate current-day 96-point photovoltaic prediction under low weather resolution. Summary of the Invention

[0007] To overcome the deficiencies of the above-mentioned existing technologies, the present invention provides a low weather resolution photovoltaic prediction method based on a similar day training set to achieve accurate 96-point photovoltaic prediction for the day ahead.

[0008] To achieve the above object, one or more embodiments of the present invention provide the following technical solutions:

[0009] In the first aspect, a low weather resolution photovoltaic prediction method based on a similar day training set is disclosed, including:

[0010] Obtain the weather forecast data of the day to be predicted;

[0011] Construct the nearest similar day similar to the weather forecast data of the day to be predicted, and obtain the meteorological data and photovoltaic power generation data of that day;

[0012] Calculate the similarity error between the photovoltaic power generation of the nearest similar day and the photovoltaic power generation of a set number of historical days, retain the historical day photovoltaic output data and weather data with errors lower than the specified threshold, and sort them from largest to smallest according to the error size to reconstruct a similar day training set;

[0013] Decompose the similar day training set into two modal components, establish a hybrid deep learning model for the two modal components for learning and training. When predicting, use the modal component photovoltaic output of the last set number of days of the similar day training set plus the weather forecast data of the day to be predicted to infer the modal component output corresponding to the two modalities, and add the two to obtain a trend photovoltaic output prediction curve;

[0014] Obtain a volatility prediction curve based on the similar day training set;

[0015] Perform weighted averaging on the obtained volatility prediction curve and trend prediction curve to obtain the final prediction result.

[0016] As a further technical solution, obtain N-day historical meteorological data before obtaining the weather forecast data of the day to be predicted, construct a historical time period database, the meteorological data of each day includes C meteorological variables, the time resolution of the meteorological data is 1 hour, and each day includes 24 meteorological segments;

[0017] For each meteorological hour segment, there are 4 photovoltaic output data points corresponding to it, with a resolution of 15 minutes.

[0018] As a further technical solution, the construction process of the historical time period database includes:

[0019] Among the N-day historical meteorological data, the weather data of any historical day n can be expressed as a T*C matrix ;

[0020] The photovoltaic output matrix corresponding to historical day n is expressed as a T*K matrix ;

[0021] Reconstruct and to obtain a two - period matrix and ;

[0022] For the weather data and photovoltaic output data of historical day n and perform a mirror flip in time to obtain a new weather data matrix and photovoltaic output matrix ;

[0023] Reconstruct and to obtain a two - period matrix and ;

[0024] Merge matrix and to obtain a historical - period weather data matrix ; Merge matrix and to obtain a historical - period output matrix ;

[0025] Merge all historical - period weather data matrices into a three - dimensional weather tensor, and merge all historical - period output matrices into a three - dimensional photovoltaic output tensor, thus completing the construction of the historical - period database.

[0026] As a further technical solution, construct the nearest similar day similar to the weather forecast data of the day to be predicted, specifically including:

[0027] The weather forecast data of the day to be predicted can also be expressed as a T * C matrix ;

[0028] Reconstruct the weather forecast data matrix to obtain a time - period reconstruction matrix of the weather forecast ;

[0029] Calculate the distance between each row vector in the time - period reconstruction matrix of the weather forecast and each historical row vector, sort them in descending order, and record the coordinate positions corresponding to the smallest distances corresponding to each row vector in the time - period reconstruction matrix of the weather forecast ;

[0030] From the three - dimensional tensors and of meteorology and photovoltaic power generation output, use the recorded coordinates to reconstruct approximately Meteorological data matrix of the day and photovoltaic output matrix ;

[0031] By performing weighted averaging on the obtained meteorological data matrix of the day and photovoltaic output matrix the weather period data matrix of the most similar day recently is finally obtained and ;

[0032] Rearranging the period data matrix and finally obtains the weather matrix of the most similar day recently and photovoltaic output matrix .

[0033] As a further technical solution, the specific steps for reconstructing the similar-day training set are as follows:

[0034] Step 1: Flatten the historical daily photovoltaic power generation matrix and the photovoltaic power generation matrix of the most similar day recently by rows into normal time series and ;

[0035] Step 2: Use the area error method to calculate the similarity error between the photovoltaic power generation data of the most similar day recently and the historical day's photovoltaic power generation data day by day;

[0036] Step 3: Extract the part of the historical day that is less than the threshold and sort it from large to small according to the error to complete the construction of the similar-day training set.

[0037] As a further technical solution, the process of obtaining the trend photovoltaic output prediction curve is as follows:

[0038] Adopt the sliding window method, divide the similar-day training set into multiple subsets every four days, and adopt the sliding window method that advances daily to form a two-dimensional matrix;

[0039] Use the photovoltaic output data of the most similar day recently to replace the actual value of the prediction day and construct a complete matrix;

[0040] Use the MVMD algorithm to decompose this matrix. In this process, effectively convert the data covering four days into a one-dimensional time series, and convert the original long one-dimensional time series into a multi-dimensional short time series, so as to ensure that the training set and the test set have the same central frequency.

[0041] In the second aspect, a low weather resolution photovoltaic prediction system based on a similar-day training set is disclosed, including:

[0042] A data acquisition module, configured to: acquire weather forecast data for the day to be predicted;

[0043] A recent similar day construction module, configured to: construct the most recent similar day similar to the weather forecast data of the day to be predicted, and acquire the meteorological data and photovoltaic power generation data of that day;

[0044] A similar day training set reconstruction module, configured to: calculate the similarity error between the photovoltaic power generation of the most recent similar day and the photovoltaic power generation of a set number of historical days, retain the historical day photovoltaic output data and weather data with an error lower than the specified threshold, and sort them from largest to smallest according to the error size, and reconstruct a similar day training set;

[0045] A predicted curve acquisition module, configured to: decompose the similar day training set into two modal components, establish a hybrid deep learning model for the two modal components for learning and training. When predicting, use the modal component photovoltaic output of the last set number of days of the similar day training set plus the weather forecast data of the day to be predicted to infer the modal component output corresponding to the two modes, and add the two to obtain a trend-based photovoltaic output predicted curve; obtain a volatility prediction curve based on the similar day training set;

[0046] A prediction module, configured to: perform weighted averaging on the obtained volatility prediction curve and trend prediction curve to obtain a final prediction result.

[0047] The above one or more technical solutions have the following beneficial effects:

[0048] The technical solution of the present invention changes the traditional similar training set construction method of directly clustering multiple weather type training sets using weather information, but instead customizes a training set for the future day to be predicted and attempts to construct this training set directly based on the photovoltaic power generation curve. However, since the true photovoltaic power generation curve of the next day is still unknown, but the weather forecast information of the next day is known, therefore, according to the weather forecast information of the next day, the most similar day can be constructed from the historical database. Through similar matching of time periods, m historical weather information of each time period closest to the weather forecast information of the day to be predicted can be found, and the corresponding photovoltaic power generation time period information can be obtained accordingly. These time periods are weighted averaged and re-spliced into a complete day to obtain the data of the most recent similar day.

[0049] In order to further improve the prediction accuracy of the current-day photovoltaic output, the technical solution of the present invention proposes a new hybrid deep learning model that combines multi-stream data mining with a long short-term memory network (LSTM), a one-dimensional convolutional network (CNN), and a self-attention mechanism. Since the LSTM network and the one-dimensional convolutional layer have different methods and capabilities in mining time features, their cooperation can extract a large number of time implicit features from different perspectives for subsequent network learning.

[0050] The technical solution of the present invention can not only solve the situation where the weather granularity is the same as the output granularity to be predicted, but also has good prediction ability when the weather granularity is coarser than the output granularity to be predicted. Therefore, it can help a large number of distributed photovoltaic power generation stations that are difficult to obtain fine-grained weather data to improve the prediction accuracy or establish prediction ability. The method of the technical solution of the present invention has been used to predict the photovoltaic output of multiple distributed sites. Compared with the general LSTM conventional prediction model, the prediction accuracy of the present invention has increased by more than 15%.

[0051] Advantages of additional aspects of the present invention will be given in part in the following description, become apparent in part from the following description, or be learned through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The specification drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention.

[0053] Figure 1 It is a schematic diagram of the structure of the hybrid deep learning model according to an embodiment of the present invention;

[0054] Figure 2 It is a flowchart of the overall prediction method according to an embodiment of the present invention;

[0055] Figure 3 It is a flowchart of trend prediction using MVMD according to an embodiment of the present invention;

[0056] Figure 4 It shows a schematic diagram of the prediction effect on a sunny day;

[0057] Figure 5 It shows a schematic diagram of the prediction effect on a highly volatile weather with multiple sunny and cloudy transitions. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0058] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0059] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention.

[0060] In the case of no conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0061] Embodiment 1

[0062] This embodiment discloses a low weather resolution photovoltaic prediction method based on a similar day training set, including:

[0063] Step 1: Using the weather forecast data of the day to be predicted, a two-way time period window sliding similarity matching and multi-curve weighting method is adopted to construct a nearest similar day, and the meteorological and photovoltaic power generation data of this day are obtained.

[0064] Then, the regional error method is used to calculate the similarity error between the photovoltaic power generation of the nearest similar day and the photovoltaic power generation of each historical day. The historical day photovoltaic output data and weather data with errors lower than the specified threshold are retained, and they are sorted from large to small according to the error size, so as to reconstruct a similar day training set.

[0065] Through this step, the problem of unclear boundaries caused by traditional weather direct clustering and the loss of a large amount of boundary effective information is solved. At the same time, because a similar day training set is established specifically for the prediction day, the data utilization degree is stronger and it has the ability of real-time update. And this method fully considers the time period similarity, judges similarity from a more detailed perspective, rather than directly making a general judgment for the whole day. At the same time, by directly matching the output of the nearest similar day with the output of the historical day, the situation where the weather is similar from the weather perspective but the actual output is completely different can be avoided.

[0066] Step 2: It includes two parallel prediction branches - volatility and trend, to generate two different prediction curves. The volatility prediction is directly trained from the above-mentioned similar day training set; on the contrary, the trend prediction needs to use a method called multivariate variational mode decomposition (MVMD) to decompose and reconstruct the similar day training set before prediction to avoid information leakage.

[0067] To further improve the prediction accuracy, a multi-stream parallel hybrid deep learning architecture: MSLCnet is proposed, which combines long short-term memory network (LSTM), convolutional neural network (CNN), residual block network and self-attention mechanism.

[0068] The above multi-stream parallel structure allows the use of multiple neural networks for feature mining and extraction simultaneously, enabling more comprehensive mining and a greater variety of mining scales, and no longer being restricted to the perspective of a single neural network. The residual block network can effectively avoid problems such as gradient disappearance caused by an overly deep convolutional network. The self-attention mechanism automatically screens and assigns weights to a large number of mined features, highlighting important features.

[0069] Step 3: Perform weighted averaging on the two curves obtained in Step 2 to obtain the final prediction result. This step can fully combine the volatility prediction and trend prediction results, and the mutual correction of the two curves makes the prediction result more accurate from a statistical perspective.

[0070] Regarding the construction of the similar-day training set in Step 1: The technical solution of this embodiment changes the traditional similar training set construction method of directly clustering multiple weather type training sets using weather information. Instead, a training set is customized for the future day to be predicted, and an attempt is made to construct this training set directly based on the photovoltaic power generation curve.

[0071] However, since the true photovoltaic power generation curve for the next day is still unknown, but the weather forecast information for the next day is known, it is possible to construct the most similar day from the historical database according to the weather forecast information for the next day. Through similar matching of time periods, m historical weather information for each time period that is closest to the weather forecast information of the day to be predicted is found, and the corresponding photovoltaic power generation time period information is obtained accordingly. These time periods are weighted averaged and re-spliced into a complete day to obtain the nearest similar-day data.

[0072] Then, using the photovoltaic power generation curve of the nearest similar day as a benchmark, similar matching is performed in the historical database using the area error method to find the data of several days similar to the prediction day, and a similar-day training set corresponding to the prediction day is constructed.

[0073] The above steps include the following when specifically implemented:

[0074] Step 1-1) Mathematically represent the available data.

[0075] Suppose there is days of historical weather data, and the meteorological data for each day includes meteorological variables such as temperature, irradiance, wind speed, etc. The time resolution of the meteorological data is 1 hour, and each day contains meteorological segments. In addition, for each meteorological hour segment, there are photovoltaic power output data points corresponding to it, with a resolution of 15 minutes. Of course, there is also the weather forecast data for the day to be predicted, and the composition and format of this data are exactly the same as the historical weather data.

[0076] The weather data of historical day n can be expressed as a matrix, where C is the type of meteorological variable and T is the number of meteorological segments:

[0077]

[0078] Among them, represents the data of the th day, the th time period, and the th type of weather variable.

[0079] The photovoltaic output matrix corresponding to historical day n can be expressed as a matrix:

[0080]

[0081] Among them, represents the th day, the th time period, and the th 15-minute interval photovoltaic output value.

[0082] The weather forecast data of the day to be predicted can also be expressed as a matrix:

[0083]

[0084] Among them, represents the data of the th time period of the day to be predicted and the th type of weather variable.

[0085] To prevent serious errors in subsequent similar matching due to different dimensions of each meteorological variable, all the above meteorological variable data have been normalized to values between 0 and 1. The normalization formula is as follows:

[0086]

[0087] Among them, is the normalized data, is the minimum value in the data, is the maximum value in the data.

[0088] Step 1-2) Reconstruction of the historical time period database based on two-time period sliding and data augmentation.

[0089] After obtaining the matrices and , perform a reconstruction operation on these two matrices to obtain the two-time period matrices and , and the reconstruction method is as follows:

[0090] ;

[0091] ;

[0092] To perform data augmentation on the dataset, the weather and PV output data are mirror-flipped over time, so that for each historical day n, a new weather data matrix and PV output matrix .

[0093]

[0094]

[0095] For and perform the same reconstruction method as above to obtain two-period matrices and . Combine matrix and ; Combine matrix and to obtain the historical period weather data matrix and output matrix .

[0096]

[0097]

[0098] All historical period weather data matrices and output matrices can be combined into two three-dimensional tensors, i.e., the construction of the historical period database is completed. The historical period database includes a weather tensor and a PV output tensor.

[0099] Weather tensor: , and the weather tensor is specifically the historical meteorological information database, including the combination of historical weather matrices for N days of time periods;

[0100] PV output tensor: .

[0101] Step 1 - 3) Construction of the nearest similar day.

[0102] After obtaining the reconstructed historical meteorological information database , the weather forecast information is also reconstructed to facilitate similar matching of meteorological time periods. Since each meteorological time period is represented by the data of a point, the weather forecast data matrix is reconstructed in the following way to avoid duplicate matching of time periods:

[0103] .

[0104] The time period reconstruction matrix of weather forecast can be expressed as:

[0105]

[0106] Calculate each row vector and each row vector the distance between , and sort them in descending order, and record the corresponding to the smallest coordinate positions of the distances. The formula for calculating the distance is as follows:

[0107]

[0108] From the three-dimensional tensor database of meteorology and photovoltaic power generation output and , use the recorded coordinates to reconstruct the meteorological data matrix of about days and the photovoltaic output matrix . The construction method is as follows:

[0109]

[0110] .

[0111] By performing weighted averaging on the obtained days of meteorological and photovoltaic output data matrices, the weather time period data matrix of the nearest similar day and are finally obtained.

[0112] Due to the relatively coarse weather granularity, considering the weighted average of the output fluctuations corresponding to multiple close weather time periods can make the final result as similar as possible to the true value and improve the similarity of the similar matching output.

[0113]

[0114] Rearrange the time period data matrices and , and finally obtain the weather matrix of the nearest similar day and the photovoltaic output matrix . The rearrangement method is as follows:

[0115]

[0116]

[0117] where: represents the matrix The element in the row and column of represents the matrix The element in the row and column of

[0118] Step 1-4): Construction of the similar-day training set.

[0119] Different from the traditional method of using all-day meteorological data for similarity clustering, the method for establishing the similar training set proposed by the present invention is based on the photovoltaic power generation data of the nearest similar days. The specific method is as follows:

[0120] Step 1-4-1): Flatten the historical daily photovoltaic power generation matrix and the nearest similar-day photovoltaic power generation matrix row by row into normal time series and .

[0121] Step 1-4-2): Calculate the similarity error between the photovoltaic power generation data of the nearest similar day and the historical day's photovoltaic power generation data day by day using the area error method. The calculation method of the area error method is as follows:

[0122]

[0123] represents the similarity area error between two curves, represents the start time of the photovoltaic output, represents the end time of the photovoltaic output, corresponds to the all-day photovoltaic output curve of the nearest similar day, while corresponds to the all-day photovoltaic output curve of a certain historical day.

[0124] Step 1-4-3): Extract the part of the historical day that is less than the threshold , and sort them from large to small according to the error to complete the construction of the similar-day training set. The threshold is determined according to the actual situation, but it is generally not recommended to be too small to avoid causing the training set to be too small and resulting in a decrease in prediction accuracy. Thus, the similar-day training set specifically includes the historical day weather and historical photovoltaic output that meet the above conditions.

[0125] Step Two: Construction of MSLCnet.

[0126] In order to further improve the prediction accuracy of photovoltaic power output for the day ahead, the technical solution of this embodiment proposes a new hybrid deep learning model named MSLCnet that combines multi-stream data mining with long short-term memory network (LSTM), one-dimensional convolutional network (CNN), and self-attention mechanism. The architecture of this model is as shown in Figure 1 shown.

[0127] After receiving various meteorological data and photovoltaic power output information for the past three days as input, MSLCnet uses a multi-stream parallel method to extract and process sequential features. Since the LSTM network and the one-dimensional convolutional layer have different methods and capabilities in mining time features, their collaboration can extract a large number of time implicit features from different perspectives for subsequent network learning. In addition, in order to mitigate the negative impact brought by the excessive depth of the network, MSLCnet introduces a residual block component to ensure that the depth of the convolutional layer does not affect the performance of the entire network. After parallel feature extraction, among the large number of obtained feature vectors, the importance levels of different feature quantities will inevitably vary, and there may even be misleading or fictional features. To solve this problem, a self-attention layer is introduced, which automatically filters and selects relevant important feature vectors through the self-attention mechanism, boosts their weights, and finally inputs them into the fully connected layer to establish a regression relationship. Therefore, the MSLCnet architecture can not only overcome the convergence challenges brought by linearly stacked networks but also flexibly utilize different scales and modalities of temporal features extracted from the input by different networks.

[0128] Step 2-1) Use the MVMD algorithm for trend prediction without information leakage.

[0129] The main application of the MVMD algorithm is that it can decompose multivariate time series into modal components sharing a common central frequency. Its constrained objective function is as follows:

[0130]

[0131] where: represents the number of decomposed modal components; represents the number of original signals input to the algorithm; represents the central frequency of each modal component after MVMD decomposition; represents the input multi-dimensional time series signal set; represents the multi-dimensional time series signal set after decomposition. The multi-dimensional time series signal set is the transformed photovoltaic power output matrix in this embodiment.

[0132] Then, use the Lagrange multiplier method to transform the constrained problem into an unconstrained problem. The form of its Lagrangian function is as follows:

[0133]

[0134] In this equation, is a penalty factor, which is a constant; while is a Lagrange multiplier that varies with time.

[0135] The ADMM method is used to solve this unconstrained optimization problem, which transforms the optimization problem into several simpler sub-optimization problems. The update equations for the modal components and the central frequency are as follows:

[0136]

[0137]

[0138] The following part provides guidance on using the MVMD algorithm while preventing the leakage of spatio-temporal information.

[0139] Step 2-2) In the traditional time series prediction and inference process, data from the previous few days is usually used to predict future data, and then the training set is divided into smaller training units.

[0140] In this embodiment, the technical solution selects the PV output of the past three days to predict the PV output of the following day. Therefore, the effective data span of the real single-input training set is four days. Therefore, in the first stage, a sliding window method is adopted, and the similar-day training set is divided into four subsets every four days using a daily-forward sliding window method, thus forming a two-dimensional matrix .

[0141]

[0142] However, it should be noted that the last row of this matrix lacks the data of the last day because that day is the prediction target of our test set. Incorporating the PV output data of the prediction day into it will cause the leakage of spatio-temporal information. To solve this problem, the actual value of the prediction day is replaced with the PV output data of the nearest similar day, so as to prevent the leakage of spatio-temporal information and construct a complete matrix for decomposition.

[0143] Next, simply decompose this matrix using the MVMD algorithm. In this process, the data covering four days is effectively converted into a one-dimensional time series, and the original long one-dimensional time series is converted into a multi-dimensional short time series, thereby ensuring that the training set and the test set have the same central frequency. When decomposing, it is generally decomposed into two modal components with different frequencies.

[0144]

[0145] During training, an MSLCnet is established for learning and training for each of the two modal components respectively. The inference method is to use the modal component photovoltaic output of the past three days plus the weather data of the current day to predict and infer the fourth day. During prediction, the modal component photovoltaic output of the last three days of the training set plus the weather forecast data of the day to be predicted are used to infer the modal component output corresponding to the two modes, and the two are added together to obtain the trend photovoltaic output prediction curve. See the appendix Figure 3 as shown.

[0146] Step 2-3) Directly use the similar-day training set for volatility prediction.

[0147] Similarly, the similar-day training set is divided into small units of four days in the above way. The output data of the first three days plus the weather data of the fourth day are used to infer the output data of the fourth day. However, no MVMD decomposition is performed, but training is directly carried out on the training set. During prediction and inference, the photovoltaic output of the last three days of the training set and the weather forecast data of the day to be predicted are input to obtain the volatility prediction curve. Only one MSLCnet model needs to be established during training.

[0148] Step 3-1) Final prediction curve generation.

[0149] After obtaining the trend and volatility prediction curves, the final prediction result can be obtained by performing weighted averaging on the two curves. The formula is as follows:

[0150]

[0151] It is usually recommended to set and both to 0.5, but the specific parameters can be adjusted through experiments according to a specific power station.

[0152] Deployment algorithm:

[0153] The algorithm of this embodiment exists in the form of a python library service and is integrated in an intelligent computing all-in-one machine. Whenever prediction is required, the relevant historical output and weather data and weather forecast data are input, and the intelligent computing all-in-one machine will automatically start the training and prediction services and output the required day-ahead output curve.

[0154] The method described in this embodiment has been used to predict the photovoltaic output of multiple distributed sites. Compared with the general LSTM conventional prediction model, the prediction accuracy of the present invention has increased by more than 15%. See the appendix Figure 4 、 5 as shown.

[0155] Embodiment 2

[0156] The purpose of this embodiment is to provide a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps of the above method are implemented.

[0157] Embodiment III

[0158] The purpose of this embodiment is to provide a computer-readable storage medium.

[0159] A computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, the steps of the above method are executed.

[0160] Embodiment IV

[0161] The purpose of this embodiment is to provide a second aspect, which discloses a low-weather-resolution photovoltaic prediction system based on a similar-day training set, including:

[0162] A data acquisition module, configured to: acquire weather forecast data of the day to be predicted;

[0163] A nearest similar-day construction module, configured to: construct the nearest similar day similar to the weather forecast data of the day to be predicted, and acquire the meteorological data and photovoltaic power generation data of that day;

[0164] A similar-day training set reconstruction module, configured to: calculate the similarity error between the photovoltaic power generation of the nearest similar day and the photovoltaic power generation of a set number of historical days, retain the historical-day photovoltaic output data and weather data with errors lower than a specified threshold, and sort them in descending order according to the error size to reconstruct a similar-day training set;

[0165] A prediction curve acquisition module, configured to: decompose the similar-day training set into two modal components, establish a hybrid deep learning model for the two modal components for learning and training. During prediction, use the modal component photovoltaic output of the last set number of days of the similar-day training set plus the weather forecast data of the day to be predicted to infer the modal component output corresponding to the two modes, and add the two to obtain a trend photovoltaic output prediction curve; obtain a volatility prediction curve based on the similar-day training set;

[0166] A prediction module, configured to: perform weighted averaging on the obtained volatility prediction curve and trend prediction curve to obtain a final prediction result.

[0167] Embodiment V

[0168] The purpose of this embodiment is to provide a computer program product containing instructions, which, when running on a computer, enables the computer to execute the methods and functions involved in any one of the above embodiments.

[0169] In the device of the above embodiments, each step involved corresponds to the first method embodiment. For the specific implementation, please refer to the relevant description part of the first embodiment. The term "computer-readable storage medium" should be understood to include a single medium or multiple media containing one or more instruction sets; it should also be understood to include any medium that can store, encode, or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0170] Those skilled in the art should understand that the above-mentioned modules or steps of the present invention can be implemented by a general-purpose computer device. Optionally, they can be implemented by program codes executable by a computing device, so that they can be stored in a storage device for execution by the computing device, or they can be separately fabricated into individual integrated circuit modules, or multiple modules or steps among them can be fabricated into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0171] Although the specific implementation of the present invention has been described above in conjunction with the accompanying drawings, it is not a limitation on the protection scope of the present invention. Those skilled in the art should understand that, based on the technical solution of the present invention, various modifications or deformations that can be made without creative efforts by those skilled in the art are still within the protection scope of the present invention.

Claims

1. A low-weather-resolution photovoltaic prediction method based on a similar day training set, characterized by: include: Obtain weather forecast data for the day to be predicted; before obtaining the weather forecast data for the day to be predicted, obtain N days of historical meteorological data and build a historical period database. The meteorological data for each day contains C meteorological variables; For each meteorological hour period, there is a corresponding photovoltaic output data point; The construction process of the historical period database includes: reconstructing the matrix of weather data of any historical day and its corresponding photovoltaic output matrix respectively to obtain corresponding two-stage time matrix; mirroring the matrix of weather data of any historical day and its corresponding photovoltaic processing matrix according to time to obtain the flipped weather data matrix and photovoltaic output matrix, and performing the same reconstruction operation on the flipped weather data matrix and photovoltaic output matrix to obtain the two-stage time matrix of the flipped matrix; vertically splicing the two-stage time matrix corresponding to the weather data matrix before and after the flipping, and rewriting them into tensor form, recorded as weather tensor; vertically splicing the two-stage time matrix corresponding to the photovoltaic output matrix before and after the flipping, and rewriting them into tensor form, recorded as photovoltaic output tensor; the weather tensor and the photovoltaic output tensor jointly construct the historical period database; Construct the nearest similar day with weather forecast data similar to the day to be predicted, and obtain the meteorological data and photovoltaic power generation data of that day; Calculate the similarity error between the photovoltaic power generation of the most recent similar day and the photovoltaic power generation of the set historical day, retain the photovoltaic output data and weather data of the historical day with errors below the specified threshold, and sort them from large to small according to the error size to reconstruct a similar day training set; The similar day training set is decomposed into two modal components, and a hybrid deep learning model is established for the two modal components for learning and training. When predicting, the modal component photovoltaic output of the last set number of days in the similar day training set is added to the weather forecast data of the predicted day to infer the modal component output corresponding to the two modes, and the two are added together to obtain the trend photovoltaic output prediction curve; Obtain volatility prediction curve based on similar day training set; Perform weighted averaging on the obtained volatility prediction curve and trend prediction curve to obtain the final prediction result; The hybrid deep learning model is MSLCnet, which consists of a multi-stream parallel LSTM network, a one-dimensional convolutional network CNN and a residual block network, and integrates a self-attention mechanism for feature filtering.

2. The low-weather-resolution photovoltaic prediction method based on similar day training sets as claimed in claim 1, characterized in that: Construct the nearest similar day with weather forecast data similar to the day to be predicted, including: The weather forecast data for the day to be predicted can also be expressed as a matrix ; Reconstruct the weather forecast data matrix to obtain the weather forecast period reconstruction matrix ; Calculate the distance between each row vector in the weather forecast period reconstruction matrix and each historical row vector, sort them in order from large to small, and record the smallest corresponding to each row vector in the weather forecast period reconstruction matrix. The coordinate position corresponding to the distance ; Three-dimensional tensor from meteorology and photovoltaic power output and In the example, the recorded coordinates are used. To reconstruct approximately Weather data matrix for the day and photovoltaic output matrix ; By obtaining Weather data matrix for the day and photovoltaic output matrix Perform weighted averaging to finally obtain the data matrix of the weather period of the nearest similar day and ; The time period data matrix and Rearrange and finally get the weather matrix of the nearest similar day and photovoltaic output matrix .

3. The low-weather-resolution photovoltaic prediction method based on similar day training sets as claimed in claim 1, characterized in that The specific steps of the daily training set are: Step 1: Matrix the historical daily photovoltaic power generation and the photovoltaic power generation matrix of the most recent similar day Flatten rows into normal time series and ; Step 2: Use the area error method to calculate the similarity error between the photovoltaic power generation data of the most recent similar day and the photovoltaic power generation data of the historical day on a daily basis; Step 3: Extract historical daily values ​​less than the threshold , and sort them from large to small according to the error to complete the construction of the similar day training set.

4. The low-weather-resolution photovoltaic prediction method based on similar day training sets as claimed in claim 1, characterized in that: The process of obtaining the trend photovoltaic output prediction curve is as follows: Using the sliding window method, the similar day training set is divided into four subsets every four days, and the sliding window method is used to advance daily, thus forming a two-dimensional matrix; Use the PV output data of the nearest similar day to replace the actual value of the forecast day and construct a complete matrix; The matrix is ​​decomposed using the MVMD algorithm, during which the data covering four days is effectively converted into a one-dimensional time series and the original long one-dimensional time series is converted into a multi-dimensional short time series, thereby ensuring that the training set and the test set have the same central frequency.

5. A low-weather-resolution photovoltaic prediction system based on a similar-day training set, using a low-weather-resolution photovoltaic prediction method based on a similar-day training set as claimed in any one of claims 1 to 4, characterized in that: include: The data acquisition module is configured to: acquire weather forecast data for the day to be predicted; The nearest similar day construction module is configured to: construct the nearest similar day with weather forecast data similar to the day to be predicted, and obtain meteorological data and photovoltaic power generation data of the day; The similar day training set reconstruction module is configured to: calculate the similarity error between the photovoltaic power generation of the most recent similar day and the photovoltaic power generation of a set number of historical days, retain the photovoltaic output data and weather data of the historical days with errors below the specified threshold, and sort them from large to small according to the error size, and reconstruct a similar day training set; The prediction curve acquisition module is configured as follows: decomposing the two modal components of the similar day training set, establishing a hybrid deep learning model for the two modal components for learning and training, and when predicting, using the modal component photovoltaic output of the last set number of days in the similar day training set plus the weather forecast data of the day to be predicted to infer the modal component outputs corresponding to the two modes, adding the two together to obtain a trend photovoltaic output prediction curve; obtaining a volatility prediction curve based on the similar day training set; The prediction module is configured to: perform weighted averaging on the obtained volatility prediction curve and trend prediction curve to obtain a final prediction result.

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

7. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the method described in any one of claims 1 to 4 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method described in any one of claims 1 to 4 are performed.

Images