High-dimensional feature-dependent unit capacity photovoltaic power prediction method and device

Abstract

The invention provides a high-dimensional feature-dependent unit capacity photovoltaic power prediction method and device, and relates to the technical field of renewable energy power generation. The method comprises the following steps: analyzing generated power data and startup capacity data in a historical period, and calculating photovoltaic power of unit capacity in the historical period; and a training sample is generated based on the photovoltaic power of the unit capacity and the meteorological characteristics, a photovoltaic prediction model is obtained through training, and photovoltaic power prediction for predicting the meteorological characteristics in a future time period is realized. According to the method, the noise influence caused by the state of the inverter is avoided, and the power prediction accuracy of the photovoltaic power station is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of renewable energy power generation, and particularly to a unit capacity photovoltaic power prediction method and device depending on high-dimensional feature dependencies. Background Art

[0002] The output power of photovoltaic power generation is directly related to solar irradiance. At the same time, irradiance is greatly affected by factors such as weather conditions, seasonal changes, and geographical locations. The output power of a photovoltaic power generation system has great volatility and uncertainty, which poses challenges to the stable operation of the power system.

[0003] Accurate prediction of photovoltaic power generation is crucial for optimizing power dispatching plans, improving the grid's acceptance capacity, and achieving effective energy management. However, traditional prediction methods usually rely on statistical analysis of historical data or simulation calculations based on physical models. There is a large amount of noise in historical data, resulting in low prediction accuracy of traditional prediction methods.

[0004] For example, traditional prediction methods generally monitor the real-time power generation of a power station as the training samples of the model to predict the power generation of the power station. Since the states of the inverters in the power station are not consistent, some are in normal working state, some are in standby state, and some are in fault state. The inverters in standby state and fault state do not generate electricity, resulting in a sudden drop in the power generation of the power station. The state of the inverter is not caused by significant changes in meteorological features such as irradiance and temperature, that is, there is a large amount of noise data in the training samples of traditional prediction methods, resulting in low accuracy of the prediction model. Summary of the Invention

[0005] The present invention provides a unit capacity photovoltaic power prediction method and device depending on high-dimensional feature dependencies, which can improve the accuracy of photovoltaic power station power prediction.

[0006] In a first aspect, the present invention provides a unit capacity photovoltaic power prediction method depending on high-dimensional feature dependencies. The method includes: obtaining the power generation data and the starting capacity data of a photovoltaic power station during a historical period, and the meteorological features of the area where the photovoltaic power station is located; calculating the unit capacity photovoltaic power of the historical period based on the power generation data and the starting capacity data of the photovoltaic power station; generating training samples based on the unit capacity photovoltaic power of the historical period and the meteorological features; performing neural network training based on the training samples to obtain a photovoltaic prediction model; obtaining the predicted meteorological features of a power station to be measured, and predicting the photovoltaic power of the power station to be measured based on the predicted meteorological features and the photovoltaic prediction model.

[0007] In a possible implementation, based on the unit capacity photovoltaic power and meteorological characteristics in historical periods, training samples are generated, including: calculating the correlation between each meteorological characteristic and the unit capacity power generation based on the unit capacity photovoltaic power and meteorological characteristics in historical periods; screening the meteorological characteristics based on the correlation between each meteorological characteristic and the unit capacity power generation to obtain relevant meteorological characteristics; performing high-dimensional feature extraction and feature transposition on the relevant meteorological characteristics to obtain high-dimensional dependent features; using the high-dimensional dependent features as inputs and the unit capacity photovoltaic power as outputs to generate multiple training samples.

[0008] In a possible implementation, calculating the correlation between each meteorological characteristic and the unit capacity power generation based on the unit capacity photovoltaic power and meteorological characteristics in historical periods includes: determining the correlation between each meteorological characteristic and the unit capacity power generation based on the following formula; where r xy is the correlation coefficient between the meteorological characteristic x and the unit capacity power generation; x i is the i-th feature of the meteorological characteristic x; the mean of each feature of the meteorological characteristic x; y i is the i-th unit capacity power generation; is the mean of each unit capacity power generation.

[0009] In a possible implementation, performing high-dimensional feature extraction and feature transposition on the relevant meteorological characteristics to obtain high-dimensional dependent features includes: performing a convolution operation on the relevant meteorological characteristics based on multiple preset convolution kernels to obtain multiple high-dimensional abstract features of each relevant meteorological characteristic; performing a pooling process on the multiple high-dimensional abstract features of each relevant meteorological characteristic to obtain the pooled high-dimensional abstract features; performing a transposition operation on the pooled high-dimensional abstract features to obtain high-dimensional dependent features.

[0010] In a possible implementation, performing neural network training based on the training samples to obtain a photovoltaic prediction model includes: Step 1, obtaining the predicted value of the unit capacity photovoltaic power corresponding to the current training sample based on the high-dimensional dependent features of the current training sample and the photovoltaic prediction model; Step 2, calculating the error between the predicted value of the unit capacity photovoltaic power corresponding to the current training sample and the true value; Step 3, if the error meets the set conditions, exit the iterative process and execute Step 4, if the error does not meet the set conditions, modify the model parameters of the photovoltaic prediction model and repeat Steps 1 to 3 until the error meets the set conditions; Step 4, if the training of the training samples is completed, output the current photovoltaic prediction model; if the training of the training samples is not completed, change the current training sample and repeat Steps 1 to 4 until the training of the training samples is completed.

[0011] In a possible implementation, based on the high-dimensional dependent features of the current training sample and the photovoltaic prediction model, the predicted value of the photovoltaic power per unit capacity corresponding to the current training sample is obtained, including: based on the high-dimensional dependent features of the current training sample and the neural network module of the photovoltaic prediction model, the output power corresponding to each high-dimensional dependent feature is obtained; based on the output power corresponding to each high-dimensional dependent feature, the attention weight corresponding to each high-dimensional dependent feature is calculated; based on the output power and the attention weight corresponding to each high-dimensional dependent feature, the predicted value of the photovoltaic power per unit capacity of the current training sample is calculated.

[0012] In a possible implementation, based on the power generation data and the installed capacity data of the photovoltaic power station, the photovoltaic power per unit capacity in the historical period is calculated, including: based on the following formula, the photovoltaic power per unit capacity at each moment in the historical period is calculated; p i_unit = p i / c i ; where p i_unit is the photovoltaic power per unit capacity at the i-th moment, p i is the power generation at the i-th moment, and c i is the installed capacity at the i-th moment.

[0013] In a possible implementation, based on the predicted meteorological features and the photovoltaic prediction model, photovoltaic power prediction is performed on the power station to be measured, including: obtaining the working status of each inverter of the power station to be measured; based on the predicted meteorological features and the photovoltaic prediction model, obtaining the photovoltaic power per unit capacity of the power station to be measured; based on the working status of each inverter of the power station to be measured, determining the installed capacity of the power station to be measured; based on the predicted photovoltaic power per unit capacity of the power station to be measured and the installed capacity of the power station to be measured, determining the power generation of the power station to be measured during the prediction period.

[0014] In a second aspect, an embodiment of the present invention provides a device for predicting photovoltaic power per unit capacity based on high-dimensional feature dependence. The device includes: a communication module and a processing module. The communication module is used to obtain the power generation data and the installed capacity data of the photovoltaic power station in the historical period, and the meteorological features of the area where the photovoltaic power station is located. The processing module is used to calculate the photovoltaic power per unit capacity in the historical period based on the power generation data and the installed capacity data of the photovoltaic power station; generate a training sample based on the photovoltaic power per unit capacity in the historical period and the meteorological features; perform neural network training based on the training sample to obtain a photovoltaic prediction model. The communication module is further used to obtain the predicted meteorological features of the power station to be measured. The processing module is further used to perform photovoltaic power prediction on the power station to be measured based on the predicted meteorological features and the photovoltaic prediction model.

[0015] In a possible implementation, the processing module is specifically configured to calculate the correlation between each meteorological feature and the power generation power per unit capacity based on the power generation power per unit capacity of photovoltaic and meteorological characteristics in historical periods; screen the meteorological features based on the correlation between each meteorological feature and the power generation power per unit capacity to obtain relevant meteorological features; perform high-dimensional feature extraction and feature transposition based on the relevant meteorological features to obtain high-dimensional dependent features; and generate a plurality of training samples with the high-dimensional dependent features as the input and the power generation power per unit capacity as the output.

[0016] In a third aspect, an embodiment of the present invention provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor is configured to call and run the computer program stored in the memory to execute the steps of the method as described in the first aspect and any possible implementation manner in the first aspect above.

[0017] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium storing a computer program, characterized in that when the computer program is executed by a processor, the steps of the method as described in the first aspect and any possible implementation manner in the first aspect above are implemented.

[0018] The present invention provides a method and device for predicting the power generation power per unit capacity based on high-dimensional feature dependence. By analyzing the power generation power data and the startup capacity data in historical periods, the present invention calculates the power generation power per unit capacity of photovoltaic in historical periods; and generates training samples with the power generation power per unit capacity and meteorological characteristics, and trains a photovoltaic prediction model to realize the prediction of the power generation power of photovoltaic with the predicted meteorological characteristics in future time periods. The present invention avoids the noise influence caused by the inverter state and improves the accuracy of power prediction of photovoltaic power stations. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for use in the embodiments or the description of the prior art. Obviously, the following drawings are only some embodiments of the present invention, and those of ordinary skill in the art can also obtain other drawings without creative efforts based on these drawings.

[0020] Figure 1 is a schematic flowchart of a method for predicting the power generation power per unit capacity based on high-dimensional feature dependence provided by an embodiment of the present invention;

[0021] Figure 2 is a schematic flowchart of a high-dimensional feature extraction process provided by an embodiment of the present invention;

[0022] Figure 3 is another schematic flowchart of a high-dimensional feature extraction process provided by an embodiment of the present invention;

[0023] Figure 4 It is a schematic diagram of a photovoltaic power prediction process provided by an embodiment of the present invention;

[0024] Figure 5 It is a schematic structural diagram of a unit-capacity photovoltaic power prediction device dependent on high-dimensional features provided by an embodiment of the present invention;

[0025] Figure 6 It is a schematic structural diagram of an electronic device provided by an embodiment of the present invention. Detailed implementation manners

[0026] In the following description, for the purpose of illustration rather than limitation, specific details such as specific system structures and technologies are put forward in order to thoroughly understand the embodiments of the present invention. However, those skilled in the art should clearly understand that the present invention can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to impede the description of the present invention with unnecessary details.

[0027] In the description of the present invention, unless otherwise specified, " / " means "or". For example, A / B can represent A or B. The "and / or" herein is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, "at least one" and "a plurality of" refer to two or more. The terms such as "first" and "second" do not limit the quantity and execution order, and the terms such as "first" and "second" do not necessarily limit to be different.

[0028] In the embodiments of the present application, words such as "exemplary" or "for example" are used to represent examples, illustrations, or explanations. Any embodiment or design solution described as "exemplary" or "for example" in the embodiments of the present application should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Exactly speaking, using words such as "exemplary" or "for example" aims to present relevant concepts in a specific way for easy understanding.

[0029] In addition, the terms "including" and "having" mentioned in the description of the present application and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device including a series of steps or modules is not limited to the listed steps or modules, but may alternatively include other unlisted steps or modules, or may alternatively include other steps or modules inherent to these processes, methods, products, or devices.

[0030] To make the objectives, technical solutions and advantages of the present invention clearer, the following will be described through specific embodiments in conjunction with the drawings of the present invention.

[0031] As described in the background art, there is a problem of low prediction accuracy in the current power prediction method for photovoltaic power stations.

[0032] To solve the above technical problems, as Figure 1 shown, an embodiment of the present invention provides a unit-capacity photovoltaic power prediction method based on high-dimensional feature dependence. The method includes steps S101-S106.

[0033] S101. Obtain the power generation data and the installed capacity data of the photovoltaic power station during the historical period, as well as the meteorological characteristics of the area where the photovoltaic power station is located.

[0034] In some embodiments, the power generation data includes the power generation at each moment during the historical period. The installed capacity data includes the installed capacity at each moment during the historical period.

[0035] In some embodiments, the meteorological characteristics include irradiance, temperature, humidity, wind speed, wind force, visibility, air pressure, precipitation, etc.

[0036] Exemplarily, an embodiment of the present invention can monitor the working status of each inverter in the photovoltaic power station, determine whether the inverter is operating normally, screen out the normally operating inverters, and sum the capacities of the inverters to obtain the installed capacity of the photovoltaic power station.

[0037] S102. Calculate the unit-capacity photovoltaic power of the historical period based on the power generation data and the installed capacity data of the photovoltaic power station.

[0038] As a possible implementation manner, an embodiment of the present invention can calculate the unit-capacity photovoltaic power at each moment during the historical period based on the following formula.

[0039] p i_unit =p i / c i ;

[0040] where p i_unit is the unit-capacity photovoltaic power at the i-th moment, p i is the power generation at the i-th moment, and c i is the installed capacity at the i-th moment.

[0041] The present invention utilizes the unit-capacity power generation p i_unitAs the model output improves the quality of the data set, it can not only improve the power prediction accuracy, but also solve the problem that the power of newly built power stations cannot be predicted. Since newly built photovoltaic power stations lack sufficient historical data, it is difficult to build an effective power prediction model in the initial stage of their construction. The present invention proposes using the power generation power per unit capacity as the model output, which can quickly apply the model trained on existing power stations to newly built power stations. Just multiply the power generation power per unit capacity predicted by the model directly by the installed capacity of the newly built power station to obtain the power generation power of the newly built power station, thereby improving the power prediction accuracy in the early stage of newly built power stations.

[0042] S103. Generate training samples based on the power generation power per unit capacity of photovoltaic power and meteorological characteristics in historical periods.

[0043] As a possible implementation, step S103 can be specifically implemented as steps S1031 - S1034.

[0044] S1031. Calculate the correlation between each meteorological characteristic and the power generation power per unit capacity based on the power generation power per unit capacity of photovoltaic power and meteorological characteristics in historical periods.

[0045] In some embodiments, the embodiments of the present invention can calculate the correlation between each meteorological characteristic and the power generation power per unit capacity using the Pearson correlation coefficient method.

[0046] Exemplarily, the embodiments of the present invention can determine the correlation between each meteorological characteristic and the power generation power per unit capacity based on the following formula.

[0047]

[0048] where r xy is the correlation coefficient between the meteorological characteristic x and the power generation power per unit capacity; x i is the i-th feature of the meteorological characteristic x; the mean value of each feature of the meteorological characteristic x; y i is the i-th power generation power per unit capacity; is the mean value of each power generation power per unit capacity.

[0049] S1032. Screen the meteorological characteristics based on the correlation between each meteorological characteristic and the power generation power per unit capacity to obtain relevant meteorological characteristics.

[0050] Exemplarily, when r = +1, it indicates a perfect positive correlation; when r = 0, it indicates no linear correlation; when r = -1, it indicates a perfect negative correlation. Through the above calculations, the relevant meteorological characteristics with correlation in the embodiments of the present invention include five meteorological characteristics: irradiance, temperature, humidity, wind speed, and visibility.

[0051] S1033. Perform high-dimensional feature extraction and feature transposition based on the relevant meteorological characteristics to obtain high-dimensional dependent features.

[0052] In some embodiments, the embodiments of the present invention can use a Conv1d convolution module, a Relu activation function, and a MaxPool1d max pooling module to extract high-dimensional abstract features of five meteorological features including irradiance, temperature, humidity, wind speed, and visibility. And transpose the high-dimensional abstract features to capture the dependencies between the high-dimensional abstract features and construct high-dimensional dependent features.

[0053] Exemplarily, step S1033 can be specifically implemented as steps A1 - A3.

[0054] A1. Perform a convolution operation on relevant meteorological features based on multiple preset convolution kernels to obtain multiple high-dimensional abstract features of each relevant meteorological feature.

[0055] A2. Perform a pooling process on the multiple high-dimensional abstract features of each relevant meteorological feature to obtain the pooled high-dimensional abstract features.

[0056] In some embodiments, the formula for the convolution operation is as follows.

[0057] Y = (X * K) + b;

[0058] Where Y is the output feature, X is the input meteorological feature, K is the convolution kernel, * represents the convolution operation, and b is the bias term.

[0059] In some embodiments, after the convolution operation, the embodiments of the present invention can also introduce non-linearity using the Relu activation function. The formula for the activation operation is as follows.

[0060] Z = f(max(0, Y));

[0061] Where Z represents the activated feature map. f represents the activation function.

[0062] In some embodiments, after the activation operation, the embodiments of the present invention can use max pooling to reduce the spatial dimension of the feature map. The formula for the pooling operation is as follows.

[0063] P = MaxPooling(Z);

[0064] Where P is the high-dimensional abstract feature obtained after pooling. MaxPooling is the pooling function.

[0065] A3. Perform a transpose operation on the pooled high-dimensional abstract features to obtain high-dimensional dependent features.

[0066] Exemplarily, assume x g represents the high-dimensional dependent feature g, h g-1 is the hidden state of its previous high-dimensional dependent feature, c g-1If it is the cell state of its previous high-dimensional dependent feature, then the calculation process of LSTM for the high-dimensional dependent feature g is as follows:

[0067] The forget gate is output by the sigmoid layer and determines how much of the previous high-dimensional feature information to forget:

[0068] f g = σ(W f · [h g-1 , x g + b f );

[0069] The input gate determines which new information is updated to the cell state:

[0070] i g = σ(W i · [h g-1 , x g + b i );

[0071] Update the cell state:

[0072] c g = f g * c g-1 + i g * tanh(W c · [h g-1 , x g + b c );

[0073] Output gate:

[0074] h g = σ(W o · [h g-1 , x g + b o ) * tanh(c g );

[0075] Among them, W f , W i , W c , W o represent the weight matrices of the forget gate, input gate, candidate cell state, and output gate respectively, and b f , b i , b c , b o represent the bias terms corresponding to the forget gate, input gate, candidate cell state, and output gate respectively; σ represents the Sigmoid activation function, tanh represents the hyperbolic tangent activation function, * represents element-wise multiplication, and [h g-1 , x g represents concatenating the hidden state h of the previous high-dimensional feature g-1and the current high-dimensional feature x g A vector formed by connecting them together.

[0076] In this way, in the embodiments of the present invention, the LSTM can learn and capture the dependencies between meteorological high-dimensional features and jointly act on the photovoltaic power generation.

[0077] Exemplarily, in the embodiments of the present invention, taking the one-dimensional convolution operation and the one-dimensional max-pooling operation as examples, the main process of high-dimensional feature extraction is illustrated.

[0078] First, the one-dimensional convolution operation

[0079] To explain in detail how to use the one-dimensional convolution (Conv1D) to process five meteorological features (irradiance, temperature, humidity, wind speed, and visibility), the specific process is as follows.

[0080] Input data =

[0081] [[1, 2, 3, 4, 5, 6, 7], # Irradiance

[0082] [8, 9, 10, 11, 12, 13, 14], # Temperature

[0083] [15, 16, 17, 18, 19, 20, 21], # Humidity

[0084] [22, 23, 24, 25, 26, 27, 28], # Wind speed

[0085] [29, 30, 31, 32, 33, 34, 35]] # Visibility

[0087] Detailed steps

[0088] 1. Define the input data and the convolution kernel

[0089] The shape of the input data is (1, 5, 7), that is, one sample, 5 features, and each feature has 7 time steps.

[0090] Assume the convolution kernel is [0.1, 0.2, 0.7], and it will be applied to three consecutive time steps of each feature.

[0091] 2. Apply the convolution operation to each feature

[0092] Since it is a one-dimensional convolution operation, the convolution kernel will act on the time series of each feature respectively.

[0093] Taking irradiance as an example (the first feature):

[0094] The first sliding (covering the positions with indices 0 to 2):

[0095] ​Calculation: 1*0.1 + 2*0.2 + 3*0.7 = 0.1 + 0.4 + 2.1 = 2.6;

[0096] Second sliding (covering positions with indices from 1 to 3):

[0097] Calculation: 2*0.1 + 3*0.2 + 4*0.7 = 0.2 + 0.6 + 2.8 = 3.6;

[0098] Similarly calculate until the last window... 5*0.1 + 6*0.2 + 7*0.7 = 6.6

[0099] Repeat the above steps for all 5 features (irradiance, temperature, humidity, wind speed, visibility) to perform the convolution operation, obtaining the output sequence after convolution for each feature.

[0100] 3. Output feature map

[0101] After obtaining the convolution results for each feature, combine these convolution results to form a new feature representation. For each feature, an output sequence with a length of time_steps (the length of the original data) - filter_size (the size of the convolution kernel) + 1 is obtained. In this example, each feature will generate a new sequence with a length of 5 (because the length of the original sequence is 7 and the size of the convolution kernel is 3).

[0102] Finally, we will obtain a new feature map with a data shape of (1, 5, 5), representing the high-dimensional abstract features extracted from the original input through the one-dimensional convolution operation.

[0103] Finally, the new feature representation is obtained:

[0105] [New feature sequence of irradiance],

[0106] [New feature sequence of temperature],

[0107] [New feature sequence of humidity],

[0108] [New feature sequence of wind speed],

[0109] [New feature sequence of visibility]

[0111] Combining the above content, it can be seen that defining how many convolution kernels will result in obtaining that many high-dimensional abstract features through the convolution operation. Each convolution operation with a convolution kernel will obtain a high-dimensional abstract feature. Combining the descriptions in the above part, it can be known that the convolution operation is to perform a dot product operation between the feature and the corresponding positions of the convolution kernel to obtain a new feature value. The convolution process is as Figure 2 shown: The data in the first row participates in the convolution. Among them, a represents the sample data. W represents the convolution kernel, and f represents the result. As​​Figure 3 As shown, the data in the nth row participates in the convolution.

[0112] Second and first-dimensional max pooling

[0113] After the one-dimensional convolution operation, one-dimensional pooling (such as MaxPooling1D) is usually applied to further process the data. The main purpose of the pooling operation is to reduce the dimension of the feature map while retaining the most important information.

[0114] Assume that the output feature after the previous convolution =

[0115] [[2.6, 3.6, 4.6, 5.6, 6.6], # Irradiance

[0116] [9.8, 10.8, 11.8, 12.8, 13.8], # Temperature

[0117] [17.0, 18.0, 19.0, 20.0, 21.0], # Humidity

[0118] [24.2, 25.2, 26.2, 27.2, 28.2], # Wind speed

[0119] [31.4, 32.4, 33.4, 34.4, 35.4]] # Visibility

[0121] Detailed steps

[0122] 1. Define the pooling window size and stride

[0123] The pooling window size (pool size) is 2.

[0124] The stride is 2.

[0125] 2. Apply the max pooling operation to each feature

[0126] Take irradiance as an example (the first feature):

[0127] The first pooling (covering positions from index 0 to 1):

[0128] Take the maximum value: max(2.6, 3.6) = 3.6

[0129] The second pooling (covering positions from index 2 to 3):

[0130] Take the maximum value: max(4.6, 5.6) = 5.6

[0131] And so on for the pooling of the 2nd, 3rd.....nth features.

[0132] 3. Output the feature map​

[0133] Finally, a new feature map with a data shape of (1, 5, 2) is obtained, representing the higher-level abstract features extracted from the original input through one-dimensional convolution and one-dimensional max pooling operations.

[0134] The output two-dimensional list values after pooling are as follows:

[0136] [3.6, 5.6], / / Irradiance

[0137] [10.8, 12.8], / / Temperature

[0138] [18.0, 20.0], / / Humidity

[0139] [25.2, 27.2], / / Wind speed

[0140] [32.4, 34.4] / / Visibility

[0142] This result represents the high-dimensional abstract features extracted from the original input data after one-dimensional convolution and one-dimensional max pooling.

[0143] After that, the high-level abstract features are transposed to obtain high-dimensional dependent features.

[0144] Generally, the data input shape of the LSTM model in time series prediction is (number of samples, number of time steps, number of features). By introducing a "memory" mechanism, the core is to capture the temporal dependence relationships between time series, focusing on the changes of the same feature at different time points and the dependence relationships of such changes. Since traditional prediction methods use meteorological forecast data as input and photovoltaic power as output, there is no temporal relationship between the input and output data, so it is impossible to fit a better time series prediction effect.

[0145] In the present invention, through high-dimensional feature dependence modeling, after extracting high-dimensional abstract features, the data is transposed and then input into the LSTM model. The shape of the transposed data is (number of samples, number of features, number of time steps). By inputting the transposed data into the LSTM model and using the "memory" mechanism, the core is to capture the feature dependence relationships, discover the potential associations between features, and pay more attention to the mutual influence and dependence relationships between different features at the same time point. For example, meteorological features such as irradiance or temperature at a future moment affect each other, and finally jointly act on the photovoltaic power at that moment, improving the photovoltaic power prediction effect.

[0146] Such as Figure 4 ​​As shown in the figure, an embodiment of the present invention provides a schematic diagram of a photovoltaic power prediction process. First, the present invention extracts 36 high-dimensional abstract features from 5 meteorological features through convolutional pooling, transposes the feature data and inputs it into an LSTM to construct a feature dependence model. Under the action of the forgetting gate and cell update state of the LSTM, the high-dimensional feature 1 is passed down through information memory or information discard to act on each subsequent high-dimensional feature, and at the same time each high-dimensional feature outputs a power value. However, the output power is not only determined by its corresponding high-dimensional feature, but also includes the combined action of the previous high-dimensional features through information transmission to predict the power. Finally, a weight attention mechanism is added. As shown in the figure, the output powers corresponding to the 36 high-dimensional features are dynamically adjusted according to their importance to highlight the influence of key features.

[0147] S1034: Using the high-dimensional dependence features as input and the photovoltaic power per unit capacity as output, generate multiple training samples.

[0148] S104: Based on the training samples, perform neural network training to obtain a photovoltaic prediction model.

[0149] As a possible implementation, step S104 can be specifically implemented as steps one to four.

[0150] Step one: Based on the high-dimensional dependence features of the current training sample and the photovoltaic prediction model, obtain the predicted value of the photovoltaic power per unit capacity corresponding to the current training sample.

[0151] Exemplarily, step one can be specifically implemented as steps B1 - B4.

[0152] B1: Based on the high-dimensional dependence features of the current training sample and the neural network module of the photovoltaic prediction model, obtain the output power corresponding to each high-dimensional dependence feature.

[0153] B2: Based on the output powers corresponding to each high-dimensional dependence feature, calculate the attention weight corresponding to each high-dimensional dependence feature.

[0154] B3: Based on the output power and attention weight corresponding to each high-dimensional dependence feature, calculate the predicted value of the photovoltaic power per unit capacity of the current training sample.

[0155] In some embodiments, the embodiment of the present invention introduces a weight attention mechanism to dynamically adjust the contribution degree according to the importance of each feature and highlight the influence of key features; convert all output powers into a probability distribution through the Softmax function to obtain the attention weight corresponding to the output power of each high-dimensional dependence feature. The calculation formula of the attention weight is as follows.

[0156]

[0157] Among them, p i represents the output power of the i-th high-dimensional dependent feature, and α i represents the weight of the output power of the i-th high-dimensional dependent feature. n represents the total number of high-dimensional dependent features.

[0158] After that, the embodiment of the present invention can obtain the final predicted value of the photovoltaic power generation per unit capacity through weighted summation. The calculation formula is as follows.

[0159]

[0160] Among them, output represents the predicted value of the photovoltaic power generation per unit capacity, and α i represents the weight of the output power of the i-th high-dimensional dependent feature, and p i represents the output power of the i-th high-dimensional dependent feature. T represents the number of high-dimensional dependent features.

[0161] Step 2: Calculate the error between the predicted value and the true value of the photovoltaic power per unit capacity corresponding to the current training sample.

[0162] Step 3: If the error meets the set conditions, exit the iterative process and execute Step 4. If the error does not meet the set conditions, modify the model parameters of the photovoltaic prediction model and repeat Steps 1 to 3 until the error meets the set conditions.

[0163] Step 4: If the training of the training sample is completed, output the current photovoltaic prediction model. If the training of the training sample is not completed, change the current training sample and repeat Steps 1 to 4 until the training of the training sample is completed.

[0164] S105: Obtain the predicted meteorological features of the power station to be measured.

[0165] S106: Based on the predicted meteorological features and the photovoltaic prediction model, perform photovoltaic power prediction on the power station to be measured.

[0166] As a possible implementation manner, Step S106 can be specifically implemented as S1061 - S1064.

[0167] S1061: Obtain the working states of the inverters of the power station to be measured.

[0168] Exemplarily, the power station to be measured can be the photovoltaic power station corresponding to the training data of the photovoltaic prediction model, or it can also be a newly built photovoltaic power station in the area where the photovoltaic power station corresponding to the training data of the photovoltaic prediction model is located.

[0169] S1062: Based on the predicted meteorological features and the photovoltaic prediction model, obtain the photovoltaic power per unit capacity of the power station to be measured.

[0170] S1063. Determine the starting capacity of the power station to be measured based on the operating states of the inverters in the power station to be measured.

[0171] S1064. Determine the power generation power of the power station to be measured during the prediction period based on the future per-unit capacity photovoltaic power of the power station to be measured and the starting capacity of the power station to be measured.

[0172] The present invention provides a method for predicting the per-unit capacity photovoltaic power dependent on high-dimensional features. By analyzing the power generation power data and starting capacity data in the historical period, the per-unit capacity photovoltaic power in the historical period is calculated; and using the per-unit capacity photovoltaic power and meteorological features, training samples are generated, and a photovoltaic prediction model is trained to realize the prediction of the photovoltaic power for the predicted meteorological features in the future period. The present invention avoids the noise influence caused by the inverter state and improves the accuracy of the power prediction of the photovoltaic power station.

[0173] It should be understood that the magnitudes of the sequence numbers of the steps in the above embodiments do not mean the order of execution. The execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present invention.

[0174] The following is an apparatus embodiment of the present invention. For the details not described in detail, reference may be made to the corresponding method embodiment above.

[0175] Figure 5 The structural schematic diagram of a per-unit capacity photovoltaic power prediction apparatus provided by an embodiment of the present invention is shown. The prediction apparatus 200 includes a communication module 201 and a processing module 202.

[0176] The communication module 201 is used to obtain the power generation power data and starting capacity data of the photovoltaic power station in the historical period, and the meteorological features of the area where the photovoltaic power station is located.

[0177] The processing module 202 is used to calculate the per-unit capacity photovoltaic power in the historical period based on the power generation power data and starting capacity data of the photovoltaic power station; generate training samples based on the per-unit capacity photovoltaic power and meteorological features in the historical period; and perform neural network training based on the training samples to obtain a photovoltaic prediction model.

[0178] The communication module 201 is further used to obtain the predicted meteorological features of the power station to be measured.

[0179] The processing module 202 is further used to perform photovoltaic power prediction on the power station to be measured based on the predicted meteorological features and the photovoltaic prediction model.

[0180] Figure 6 The structural schematic diagram of an electronic device provided by an embodiment of the present invention is shown. As Figure 6As shown, the electronic device 300 includes: a processor 301, a memory 302, and a computer program 303 stored in the memory 302 and executable on the processor 301. When the processor 301 executes the computer program 303, the steps in the above method embodiments are implemented, such as Figure 1 the steps S101 - S106 shown. Alternatively, when the processor 301 executes the computer program 303, the functions of each module / unit in the above device embodiments are implemented. For example, Figure 5 the functions of the communication module 201 and the processing module 202 shown.

[0181] Exemplarily, the computer program 303 can be divided into one or more modules / units. The one or more modules / units are stored in the memory 302 and executed by the processor 301 to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing specific Figure 3 functions, and these instruction segments are used to describe the execution process of the computer program 303 in the electronic device 300. For example, the computer program 303 can be divided into Figure 5 the communication module 201 and the processing module 202 shown.

[0182] The so-called processor 301 may be a central processing unit (CPU), or may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0183] The memory 302 may be an internal storage unit of the electronic device 300, such as a hard disk or memory of the electronic device 300. The memory 302 may also be an external storage device of the electronic device 300, such as a plug-in hard disk equipped on the electronic device 300, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, etc. Further, the memory 302 may also include both an internal storage unit of the electronic device 300 and an external storage device. The memory 302 is used to store the computer program and other programs and data required by the terminal. The memory 302 may also be used to temporarily store data that has been output or is to be output.

[0184] The above-described embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A high-dimensional feature-dependent unit capacity photovoltaic power prediction method, characterized in that: include: Obtain the power generation data and startup capacity data of photovoltaic power stations in the historical period, as well as the meteorological characteristics of the areas where the photovoltaic power stations are located; Calculate the photovoltaic power per unit capacity in the historical period based on the power generation data and the startup capacity data of the photovoltaic power station; generating training samples based on the photovoltaic power per unit capacity in the historical period and the meteorological characteristics; Based on the training samples, neural network training is performed to obtain a photovoltaic prediction model; The forecast meteorological characteristics of the power station to be tested are obtained, and based on the forecast meteorological characteristics and the photovoltaic prediction model, the photovoltaic power of the power station to be tested is predicted.

2. The high-dimensional feature-dependent unit capacity photovoltaic power prediction method according to claim 1 is characterized in that: The generating of training samples based on the photovoltaic power per unit capacity in the historical period and the meteorological characteristics comprises: Based on the photovoltaic power per unit capacity in the historical period and the meteorological characteristics, calculating the correlation between each meteorological characteristic and the power generation per unit capacity; Based on the correlation between the meteorological characteristics and the power generation per unit capacity, the meteorological characteristics are screened to obtain relevant meteorological characteristics; Based on the relevant meteorological features, high-dimensional feature extraction and feature transposition are performed to obtain high-dimensional dependent features; A plurality of training samples are generated by taking the high-dimensional dependent features as input and the unit capacity photovoltaic power as output.

3. The high-dimensional feature-dependent unit capacity photovoltaic power prediction method according to claim 2 is characterized in that: The calculating, based on the photovoltaic power per unit capacity in the historical period and the meteorological characteristics, the correlation between each meteorological characteristic and the power generation per unit capacity includes: Based on the following formula, the correlation between each meteorological characteristic and the power generation per unit capacity is determined; Among them, r xy is the correlation coefficient between meteorological feature x and unit capacity power generation; i is the i-th feature of meteorological feature x; The mean value of each characteristic of meteorological characteristic x; y i is the power generation of the i-th unit capacity; It is the mean value of power generation per unit capacity.

4. The high-dimensional feature-dependent unit capacity photovoltaic power prediction method according to claim 2 is characterized in that: The high-dimensional feature extraction and feature transposition based on the relevant meteorological features to obtain high-dimensional dependent features include: Based on multiple preset convolution kernels, convolution operations are performed on the relevant meteorological features to obtain multiple high-dimensional abstract features of each relevant meteorological feature; Pooling is performed on multiple high-dimensional abstract features of each relevant meteorological feature to obtain pooled high-dimensional abstract features; A transposition operation is performed on the pooled high-dimensional abstract features to obtain the high-dimensional dependency features.

5. The high-dimensional feature-dependent unit capacity photovoltaic power prediction method according to claim 1, characterized in that: The step of performing neural network training based on the training samples to obtain a photovoltaic prediction model includes: Step 1: Based on the high-dimensional dependency features of the current training sample and the photovoltaic prediction model, the predicted value of the photovoltaic power per unit capacity corresponding to the current training sample is obtained; Step 2: Calculate the error between the predicted value and the true value of the unit capacity photovoltaic power corresponding to the current training sample; Step 3: If the error meets the set conditions, exit the iteration process and execute step 4; if the error does not meet the set conditions, modify the model parameters of the photovoltaic prediction model and repeat steps 1 to 3 until the error meets the set conditions; Step 4: if the training sample is trained, the current photovoltaic prediction model is output; if the training sample is not trained, the current training sample is changed, and steps 1 to 4 are repeated until the training sample is trained.

6. The high-dimensional feature-dependent unit capacity photovoltaic power prediction method according to claim 5, characterized in that: The method of obtaining the predicted value of the photovoltaic power per unit capacity corresponding to the current training sample based on the high-dimensional dependency features of the current training sample and the photovoltaic prediction model includes: Based on the high-dimensional dependency features of the current training samples and the neural network module of the photovoltaic prediction model, the output power corresponding to each high-dimensional dependency feature is obtained; Based on the output power corresponding to each high-dimensional dependent feature, the attention weight corresponding to each high-dimensional dependent feature is calculated; Based on the output power and attention weight corresponding to each high-dimensional dependent feature, the predicted value of the photovoltaic power per unit capacity of the current training sample is calculated.

7. The high-dimensional feature-dependent unit capacity photovoltaic power prediction method according to claim 1, characterized in that: The calculating of the photovoltaic power per unit capacity in the historical period based on the power generation data and the startup capacity data of the photovoltaic power station includes: Based on the following formula, the photovoltaic power per unit capacity at each moment in the historical period is calculated; p i_unit =p i / c i ; Among them, p i_unit is the photovoltaic power per unit capacity at the i-th moment, p i is the power generation at the i-th moment, c i is the startup capacity at the i-th moment.

8. The high-dimensional feature-dependent unit capacity photovoltaic power prediction method according to claim 1, characterized in that: The photovoltaic power prediction of the power station to be tested based on the predicted meteorological characteristics and the photovoltaic prediction model includes: Obtaining the working status of each inverter of the power station to be tested; Based on the forecast meteorological characteristics and the photovoltaic prediction model, the photovoltaic power per unit capacity of the power station to be tested is obtained; Determining the startup capacity of the power station to be tested based on the working status of each inverter of the power station to be tested; Based on the photovoltaic power per unit capacity of the power station to be tested and the startup capacity of the power station to be tested, the power generation power of the power station to be tested in the predicted period is determined.

9. A high-dimensional feature-dependent unit capacity photovoltaic power prediction device, characterized in that: include: The communication module is used to obtain the power generation data and startup capacity data of the photovoltaic power station in the historical period, as well as the meteorological characteristics of the area where the photovoltaic power station is located; A processing module, configured to calculate the photovoltaic power per unit capacity in a historical period based on the power generation data and the startup capacity data of the photovoltaic power station; and to generate training samples based on the photovoltaic power per unit capacity in the historical period and the meteorological characteristics; Based on the training samples, neural network training is performed to obtain a photovoltaic prediction model; The communication module is also used to obtain the forecast meteorological characteristics of the power station to be tested; The processing module is also used to predict the photovoltaic power of the power station to be tested based on the predicted meteorological characteristics and the photovoltaic prediction model.

10. The high-dimensional feature-dependent unit capacity photovoltaic power prediction device according to claim 9, characterized in that: The processing module is specifically used to calculate the correlation between each meteorological feature and the unit capacity power generation based on the unit capacity photovoltaic power of the historical period and the meteorological features; based on the correlation between each meteorological feature and the unit capacity power generation, the meteorological features are screened to obtain relevant meteorological features; based on the relevant meteorological features, high-dimensional feature extraction and feature transposition are performed to obtain high-dimensional dependency features; A plurality of training samples are generated by taking the high-dimensional dependent features as input and the unit capacity photovoltaic power as output.

Images