Method and system for predicting extremely short-term solar power generation for real-time control

Abstract

The method includes the steps of: performing normalization processing on meteorological data and feature correlation analysis; performing short-term solar power generation forecasting using a BP neural network, inputting meteorological data and historical output data, and outputting short-term forecast values ​​with a resolution of 15 minutes; and improving the temporal resolution of the forecast data by performing spline interpolation and outlier removal on the upper layer results of the BP neural network and inputting them as a long-short-term cyclic neural network, thereby obtaining extremely short-term solar power generation forecast data with a resolution of 1 minute. By comprehensively considering meteorological factors and employing advanced neural network models and data processing techniques, forecasting accuracy is guaranteed, and solar power generation forecasting on an extremely short-term scale can be achieved, making it applicable to real-time control and optimized operation of solar power plants.

Description

[Technical Field]

[0001] The present invention relates to the technology of ultra-short-term solar power generation forecasting, and more particularly to an ultra-short-term solar power generation forecasting method and system for real-time control. [Background technology]

[0002] Because solar power output is highly random, the operation of solar power plants becomes complex, easily leading to safety hazards and economic losses. Therefore, studying methods for predicting solar power output is of great importance for performance monitoring, fault detection, and operational control of solar power plants. There are many different methods for predicting solar power output, and depending on the prediction format, solar power output prediction can be mainly divided into point prediction, interval prediction, and probability prediction. Point prediction belongs to certainty prediction and provides a determined predicted output value at a certain future time; interval prediction provides upper and lower limit levels of output at a predetermined confidence level; and probability prediction provides expected output value and probability distribution information at a future time; these are collectively called uncertainty prediction. The method according to the present invention is mainly applied to the real-time control of solar power plants, and therefore, this document mainly examines the point prediction method. The artificial intelligence method does not require obtaining a specific expression between input and output, and obtains a prediction model of solar power output by training on historical data. Such methods require large amounts of historical data, are easy to operate and implement, and artificial intelligence prediction methods are currently one of the most commonly used point prediction methods for solar power output. However, current research on conventional artificial intelligence prediction methods focuses on short-term predictions of 1 to 10 days into the future and ultra-short-term predictions of 4 hours into the future, with a time resolution of 15 minutes. Generally, there is a problem with the prediction time scale not being sufficiently fine and the prediction accuracy being low, making it difficult to meet the flexible control needs of photohydrogen storage stations. [Overview of the project] [Problems that the invention aims to solve]

[0003] In view of the problems present in the prior art described above, the present invention is submitted.

[0004] Therefore, the present invention provides a method for real-time control of ultra-short-term solar power generation forecasting that can solve the problem of conventional forecasting scales not being sufficiently detailed and resulting in low forecasting accuracy. [Means for solving the problem]

[0005] To solve the above technical problems, the present invention provides the following technical solution, a method for predicting extremely short-term solar power generation for real-time control, which includes the steps of: normalizing radiation, temperature, humidity and solar power generation data using limit values ​​of year-round statistics, obtaining variables and solar power generation time series, performing feature correlation analysis, and constructing a two-layer solar power generation output prediction framework based on deep learning; performing short-term solar power generation prediction based on a BP neural network, taking weather data and historical output data as input, optimizing the loss function by dividing into time periods, and obtaining upper-layer short-term prediction values ​​with a resolution of 15 min; performing spline interpolation on the upper-layer results, extracting outliers using the box plot and violin plot methods, removing the outliers using the moving time window filtering method, improving the resolution of the prediction data based on a long-short-term cyclic neural network, achieving both prediction time scale and prediction accuracy, and obtaining lower-layer extremely short-term solar power generation prediction data with a resolution of 1 min.

[0006] As a preferred method for predicting extremely short-term solar power generation for real-time control described in the present invention, the normalization process is as follows: χ n =( χ n -χ min ) / (χ max -χ min ) Here, χ n is the nth data item to be processed, and χ max and χ min These are the local maximum and local minimum values ​​of the variable, respectively. A correlation analysis is performed on each meteorological factor and solar power output, and it is decided that numerical weather forecasts and historical measured weather data from the solar power plant, as well as effective power generation data from historical output data, will be used as input layer sample data for the training process of the BP neural network.

[0007] As a preferred solution for the ultra-short-term solar power generation prediction method for real-time control described in the present invention, the BP neural network includes an input layer, a hidden layer and an output layer, The input layer is used to receive input data for the BP neural network. Since the output power of the solar power plant correlates with solar radiation intensity and is also affected by temperature, relative humidity, and cloud cover, the correlation analysis is performed by combining the relationship between each meteorological factor and the solar power output. The predicted solar radiation intensity in the numerical weather forecast for the solar power plant, the actual solar radiation intensity in the historical weather measurements, and the effective power generation in the historical output data are determined and used as sample data for the input layer during the training process of the BP neural network. The output layer performs short-term power forecasting, and the output data of the output layer is the future 24-hour output power sequence of the solar power plant, with a time interval of 15 minutes. The selection of the number of neurons in the hidden layer is related to the size and accuracy of the neural network. Based on Kolmogorov's theorem, if the number of input variables is n, the number of neurons in the hidden layer is 2n+1. The number of neurons in the hidden layer is determined using an iterative test method. The training process of the aforementioned BP neural network is related to the prediction accuracy of the output power model of a photovoltaic power generation system, and is improved by combining a BP learning algorithm that increases the dynamic term and a BP learning algorithm that makes the learning rate variable, thereby improving the convergence speed and reducing the probability of falling into a local minimum.

[0008] A preferred approach to the ultra-short-term solar power generation prediction method for real-time control described in the present invention is to combine a BP learning algorithm that increases the dynamic term and a BP learning algorithm that makes the learning rate variable, which includes sequentially inputting P samples and calculating the output and backpropagation error of each layer; adjusting the connection weights of each layer according to a weight adjustment formula after all calculations for the P samples are completed; calculating the output, backpropagation error and total network error E(t) of each layer according to the new connection weights, terminating the training process if E(t) < ε or t > T; and if E(t) ≥ ε and t ≤ T, re-inputting the P samples and calculating the output and backpropagation error of each layer, and performing training for a new cycle. Here, P is the number of training samples, T is the maximum number of training iterations, and ε is the system tolerance.

[0009] A preferred approach to the ultra-short-term solar power generation forecasting method for real-time control described in the present invention involves employing a travel time window filtering method to remove outliers, the spline interpolation of the results obtained by the upper layer BP network, the display of quartiles, median, and data density distribution based on the box plot and violin plot, the initial determination of data concentration and dispersion by observing the shape of the figures and peak values, the identification of outliers by combining statistical methods, and the definition of outliers using the interquartile range. This includes selecting 4 hours as the size of the time shift window, calculating the mean and median statistics of the data within the time shift window, replacing outliers using the statistics within the time shift window until all data processing is complete, then shifting the time shift window forward by one time unit, performing data cleaning, and finally removing the outliers.

[0010] As a preferred solution of the very short-term solar power generation power prediction method for real-time control described in the present invention, the step of improving the resolution of prediction data based on a long short-term memory neural network means that the memory cell unit is the core assembly of the LSTM network, and the input of the model is the sequence input x at time t t , the hidden layer cell state h at time t-1 t-1 and the memory unit c t-1 , and the output includes the memory unit state c t and the hidden layer state h t , where each of c t and h t includes the long-term and short-term memory information of the model, and realizes the reading and modification of the memory cell unit by controlling the input gate, forget gate and output gate, and conducts the flow of information between the networks. Tanh represents the activation function of tanh. The input gate inputs parameters using the sigmoid activation function and controls the variable between [0,1] to realize the control of x t to c t . The forget gate selectively forgets the neuron state of the previous time, and the specific expression form is to utilize the control of the memory unit c t-1 to c t . The output gate is used to output and control the parameter variable, that is, to utilize the degree of influence on h t of c t . The calculation formulas are as follows respectively i t =σ(W ix x t +W ih h t-1 +b i ) f t =σ(W fx x t +W fh h t-1 +b f ) o t =σ(W ox x t +W oh h t-1 +b o ) Here, i t ,f t , o t These represent the state calculation results of the input gate, the forget gate, and the output gate, respectively, W ix and W ih , W fx and W fh , W ox and W oh and b i , b f , b o These represent the weight matrix and offset term of the corresponding gate, respectively, and σ represents the sigmoid activation function.

[0011] In the LSTM model, the output result of the memory module at time t is jointly determined by the output gate and the unit state, and the equation is as follows:

number

number

[0012] Another objective of the present invention is to provide an ultra-short-term solar power forecasting system for real-time control. The feature-related analysis module enhances the understanding of interactions between data through statistical normalization and time-series analysis of environmental and power data, providing a robust data foundation for accurate forecasting. The short-term power forecasting module utilizes weather and historical output data and BP neural networks to achieve optimized power resource allocation and cost control, significantly improving the efficiency and economics of operational decisions. The ultra-short-term high-resolution forecasting module improves the system's response speed to sudden climate changes through high-resolution forecasting and advanced data processing technology, thereby improving the reliability of the power grid and the maintenance efficiency of solar power generation equipment.

[0013] A preferred solution for the ultra-short-term solar power generation forecasting system for real-time control described in the present invention includes a feature analysis module, a short-term power forecasting module, and an ultra-short-term detailed forecasting module. The aforementioned feature analysis module normalizes radiation, temperature, humidity, and solar power generation data using the limits of year-round statistics, obtains the variables and solar power generation time series, and performs feature correlation analysis. The aforementioned short-term power forecasting module performs short-term solar power generation forecasting based on a BP neural network, takes weather data and historical output data as input, optimizes the loss function for each time period, and obtains upper-level short-term forecast values ​​with a resolution of 15 min. The ultra-short-term detailed prediction module includes the steps of performing spline interpolation on the upper layer results, extracting outliers using box plots and violin plots, removing the outliers using a moving time window filtering method, improving the resolution of the prediction data based on a long- and short-term cyclic neural network, achieving both prediction time scale and prediction accuracy, and obtaining lower-layer ultra-short-term solar power generation prediction data with a resolution of 1 min.

[0014] The computer equipment includes memory and a processor, the memory storing a computer program, and the processor, when executing the computer program, implements steps of the method of any one of the methods for predicting ultra-short-term solar power generation for real-time control.

[0015] A computer-readable storage medium stores a computer program, which, when executed by the processor, implements steps of any one of the methods for predicting extremely short-term solar power generation for real-time control. [Effects of the Invention]

[0016] The method of the present invention comprehensively considers meteorological factors and employs advanced neural network models and data processing technologies to guarantee prediction accuracy and enable solar power generation forecasting on an extremely short time scale, making it applicable to real-time control and optimized operation of solar power plants. [Brief explanation of the drawing]

[0017] To more clearly explain the technical concept of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly described below. Clearly, the drawings in the following description represent only a few embodiments of the present invention, and those skilled in the art can obtain other drawings based on these without any creative effort.

[0018] [Figure 1] This is a flowchart of the ultra-short-term solar power generation prediction method for real-time control according to the present invention. [Figure 2] This is a BP network structure and learning flowchart for a real-time control method for predicting extremely short-term solar power generation according to one embodiment of the present invention. [Figure 3] This is the basic structure of an LSTM network for a real-time control of extremely short-term solar power generation prediction method according to one embodiment of the present invention. [Figure 4]This shows the trend of the loss function change during the training process of a very short-term solar power generation prediction method for real-time control according to one embodiment of the present invention. [Figure 5] This is a schematic diagram comparing short-term predicted values ​​of a method for providing a training set of a real-time control method for predicting extremely short-term solar power generation according to one embodiment of the present invention, short-term predicted values ​​of an existing method, and measured values. [Figure 6] This is a schematic diagram comparing short-term predicted values ​​of a method for providing a test set of a real-time control method for predicting extremely short-term solar power generation according to one embodiment of the present invention, short-term predicted values ​​of an existing method, and measured values. [Figure 7] This is a schematic diagram of the mean squared error curve during the training process of a real-time control method for predicting extremely short-term solar power generation according to one embodiment of the present invention. [Figure 8] This is a schematic diagram of the loss function curve during the training process of a real-time control method for predicting extremely short-term solar power generation according to one embodiment of the present invention. [Figure 9] This is a schematic diagram comparing the prediction results of a training set for a real-time control method for predicting extremely short-term solar power generation according to one embodiment of the present invention. [Figure 10] This is a schematic diagram comparing the prediction results of a test set of a real-time, ultra-short-term solar power generation prediction method for real-time control according to one embodiment of the present invention. [Modes for carrying out the invention]

[0019] To make the above-mentioned objectives, features, and advantages of the present invention easier to understand, specific embodiments of the present invention will be described in detail below with reference to the drawings of the specification, and it is clear that the embodiments described are some, but not all, embodiments of the present invention. All other embodiments that can be obtained by those skilled in the art without creative work based on the embodiments of the present invention are all within the scope of protection of the present invention.

[0020] (Example 1) Referring to Figures 1 to 3, the first embodiment of the present invention provides a method for predicting extremely short-term solar power generation for real-time control. First, based on meteorological data and historical output data, a two-layer deep learning-based solar power output prediction framework, as shown in Figure 1, is constructed. By utilizing the complex nonlinear mapping approximation capabilities of BP neural networks and long-term-short-term cyclic neural networks (LSTMs), long-term dependencies in power sequence data are captured and modeled to obtain solar power output prediction data that can meet the real-time control needs of solar power plants, thereby solving the problem of insufficient prediction scale and low prediction accuracy. Next, the upper layer performs short-term solar power generation forecasting based on a BP neural network, replaces the limit values ​​of the year-round statistics, performs normalization preprocessing and feature correlation analysis, optimizes the loss function for each time period, obtains short-term forecast values ​​with a resolution of 15 minutes, and improves forecast accuracy. Finally, the lower layer performs spline interpolation fitting on the upper layer results, extracts outliers using methods such as box plots and violin plots, removes outliers using a moving time window filtering method, and improves the resolution of the prediction data by obtaining extremely short-term solar power generation prediction data with a resolution of 1 min, while balancing the prediction time scale and prediction accuracy based on a long-short-term cyclic neural network (LSTM).

[0021] Compared to conventional technologies, the prediction method according to the present invention has a good prediction accuracy rate and high prediction resolution, and can meet the needs of real-time control.

[0022] S1: Using the limit values ​​of year-round statistics, normalization is performed on radiation, temperature, humidity, and solar power generation data. Variables and solar power generation time series are obtained, and feature correlation analysis is performed.

[0023] Furthermore, regarding the designed normalization and feature analysis method, when predicting solar power generation using multivariable time series, the dimensions and numerical values ​​differ significantly between different input variables. Considering the input / output range of the nonlinear activation function in the model, it is necessary to perform normalization on the variables and the solar power generation time series in order to avoid neuron saturation and to equally consider the effect of each variable on solar power generation. The prediction model of the present invention employs a sigmoid activation function, and the output is limited to between 0 and 1. Inputs such as radiation, temperature, humidity, and solar power generation are normalized using the limit values ​​of year-round statistics, and the numerical values ​​are normalized within the interval [0,1]. The expression is as follows: χ n =( χ n -χ min ) / (χ max -χ min ) (1) Here, χ n is the nth data item to be processed, and χ max and χ min These are the local maximum and local minimum values ​​of the variable, respectively. A correlation analysis is performed on each meteorological factor and solar power output, and it is decided that numerical weather forecasts and historical measured weather data from the solar power plant, as well as effective power generation data from historical output data, will be used as input layer sample data for the training process of the BP neural network.

[0024] S2: Short-term solar power generation forecasts are performed based on a BP neural network. Numerical weather forecasts, historical measured weather data, and active power generation are used as inputs, and the loss function is optimized for each time period to obtain upper-level short-term forecast values ​​with a resolution of 15 min.

[0025] It includes three parts: the input layer, the hidden layer, and the output layer.

[0026] 1) The input layer is used to receive input data for the BP neural network. The output power of a solar power plant is mainly correlated with solar radiation intensity and is also affected by temperature, relative humidity, and cloud cover. Correlation analysis is performed by combining the relationship between each meteorological factor and solar power output to determine the predicted solar radiation intensity in numerical weather forecasts for solar power plants, the actual solar radiation intensity in historical weather measurements, and the effective power generation. These are then used as input layer sample data for the training process of the BP neural network.

[0027] 2) In the output layer, the output layer performs short-term power forecasting, and the output data of the output layer is the future 24-hour output power sequence of the solar power plant, with a time interval of 15 minutes.

[0028] 3) In the hidden layer, the accuracy of the artificial neural network can be improved by adopting a single hidden layer and increasing its number of neurons. This is simpler in terms of structural implementation than increasing the number of hidden layers, and therefore, the solar power output prediction model adopts a single hidden layer structure. The choice of the number of neurons in the hidden layer is directly related to the size and accuracy of the neural network. Based on Kolmogorov's theorem, if the number of input variables is n, the number of neurons in the hidden layer can generally be 2n+1, and the number of neurons in the hidden layer can also be determined using the iterative test method.

[0029] The training process of the BP neural network is directly related to the prediction accuracy of the output power model of the solar power generation system. The conventional BP neural network training adopts the error backpropagation learning algorithm, which has problems such as being prone to falling into local minima and having a slow convergence speed. Scholars at home and abroad in China have improved the conventional BP learning algorithm from different angles. Generally speaking, it can be divided into two categories: (1) improvements based on the standard gradient descent method, such as the BP learning algorithm with an increased momentum term, the BP learning algorithm with a variable learning rate, etc.; (2) improvements based on numerical optimization algorithms, such as the quasi-Newton method, the Levenberg-Marquardt algorithm, the conjugate gradient method, etc.

[0030] The present invention combines the BP learning algorithm with an increased momentum term and the BP learning algorithm with a variable learning rate to improve the conventional BP learning algorithm, thereby improving the convergence speed and reducing the probability of falling into local minima. There are P training samples, T is the maximum number of training times, w is the connection weight, w(t) is the weight at the t-th iteration, Δw BP (t) is the weight change amount at the t-th iteration of the conventional BP learning algorithm, E(t) is the total network error at the t-th iteration, ε is the system allowable error, and η is set as the learning rate. The flowchart of the BP learning algorithm is as shown in Figure 2, and the algorithm steps are as follows.

[0031] (1) Initialize the weight w, set the current as the t-th iteration, (2) Input P samples in sequence, set the currently input sample as the p-th sample, calculate the output of each layer and the backpropagation error, (3) If p < P, then p = p + 1, go to step (2); if p ≥ P, go to step (4), (4) Adjust the connection weights of each layer according to the weight adjustment formula, (5) Based on the new connection weights, calculate the output of each layer, the backpropagation error, and the total network error E(t). If E(t) < ε or t > T, terminate the training process. If the opposite is true, t = t + 1, proceed to step (2) and train a new cycle. Here, P is the number of training samples, and T is the maximum number of training iterations.

[0032] Since solar power generation only outputs during the daytime, the prediction accuracy during daytime hours is of crucial importance for flexible control. Therefore, the prediction accuracy is calculated by dividing it into multiple time periods, day and night, the accuracy calculation formula is integrated into the loss function, and the model's penalty coefficient is increased based on parameters of data that do not meet the standard. Limited computational resources are applied to models with poor accuracy for priority processing, and the results of the failing models are repeatedly validated.

[0033] S3: Spline interpolation is performed on the upper layer results, outliers are extracted using box plots and violin plots, and outliers are removed using a moving time window filtering method. The resolution of the prediction data is improved based on a long- and short-term cyclic neural network, achieving a balance between prediction time scale and prediction accuracy, and obtaining lower layer ultra-short-term solar power generation prediction data with a resolution of 1 min.

[0034] The structure of the designed LSTM cyclic neural network is shown in Figure 3. Before training the LSTM with data, it is necessary to preprocess the data collected by the SCADA system. The results obtained by the upper layer BP network are spline interpolated, and the quartiles, median, and data density distribution are displayed based on the box plot and violin plot. An initial determination of the concentration and dispersion of the data is made by observing the shape of the figures and peak values, and outliers are identified by combining statistical methods, and these outliers are defined using the interquartile range.

[0035] Four hours is selected as the size of the time-travel window. Statistics such as the mean and median of the data within the time-travel window are calculated. Outliers are replaced using these statistics until all data has been processed. Then, the time-travel window is moved forward by one time unit. After data cleaning, outliers are removed, which makes the statistical characteristics of the data clearer and is advantageous for training the neural network. The LSTM network structure consists of input gates, output gates, and forget gates, and differs from conventional cyclic neural networks in that it has multiple hidden layers, and the neurons in the hidden layers are replaced by memory units with gate control mechanisms. Its topology incorporates gate control units to control the degree to which the information at the current time influences the information at the previous time, giving the model a relatively long-term "memory function" and making it suitable for long-term nonlinear sequence problems.

[0036] The memory cell unit is the core assembly of the LSTM network, and the input to the model is the time sequence input x t , hidden layer cell state h at time t-1 t-1 and memory unit c t-1 The output includes the memory unit state c. t and hidden layer state h t This includes, where c t and h t Each of these contains the long-term and short-term memory information of the model, and reading and modifying the memory cell units is achieved by controlling the input gate, forget gate, and output gate, and information flows between the network, tanh represents the activation function of tanh, the input gate uses the sigmoid activation function to input parameters and controls the variable between [0,1], x t c t Control is achieved to the forget gate, which selectively forgets the neuron state from the previous time, and the specific representation is memory unit c t-1 c t This involves utilizing control over the output gate, which is used to output and control the parameter variable, i.e., ct h of t Using the degree of influence on..., the calculation formulas are as follows respectively: i t = σ(W ix x t + W ih h t-1 + b i ) f t = σ(W fx x t + W fh h t-1 + b<​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​The symbols represent the state weight matrix and offset term of the input layer, respectively, and * indicates that each element is multiplied by its position.

[0038] (Example 2) Figures 4 to 10 show a second embodiment of the present invention, which provides a method for predicting extremely short-term solar power generation for real-time control. Scientific verification was conducted experimentally to verify the beneficial effects of the present invention.

[0039] To verify the effectiveness of the method provided in this invention, a solar power plant located in Delingha City, Haixi Mongolian and Tibetan Autonomous Prefecture, Qinghai Province (97.37°E, 37.37°N) was selected. The plant had a capacity of 100 MW and a grid connection rate of 85%. Historical power output data and weather history information for April 1, 2023 and May 4, 2023 were selected and predicted using the above model. Days 1-20 were used as the training set, and days 21-34 were used as the test set.

[0040] First, the upper layer BP neural network was trained and its predictive effect evaluated. The number of hidden layers was 2, both using the Sigmoid activation function, with a learning rate of 0.1, and optimized using random gradient descent (SGD). The model was trained 10,000 times, and the training loss was recorded every 100 iterations. Since solar power generation is not zero only when there is sunlight, the error calculation was divided into two cases: one that considers nighttime and one that does not.

[0041] To accurately evaluate a predictive model, it is necessary to adopt certain evaluation metrics. The most commonly used metric for evaluation and prediction is the Mean Absolute Percentage Error (MAPE), as shown in equation (4).

number

[0042] There is a zero value in solar power generation, |P f i -P a i | / P a i As MAPE approaches infinity, it becomes meaningless. Therefore, the root mean square error (RMSE) is used to evaluate the prediction results. RMSE represents the model's ability to control absolute error, and its calculation formula is as follows.

number

number

[0043] Figure 4 shows the trend of change in the loss function during the training process. The loss function of the neural network in the method according to the present invention converges after approximately 1000 training iterations, in less than 1 minute. This indicates that the training time is short, convergence is fast, and the method has high practical value.

[0044] Figure 5 compares the predicted output of the training set with the measured value and the conventional short-term prediction value. The three curves in the figure represent the measured value of solar power generation output, the short-term prediction value of the provided method, and the short-term prediction value of the conventional method, respectively. Figure 6 compares the short-term prediction results with the measured value and the conventional short-term prediction value.

[0045] To verify the prediction accuracy of the provided method, the RMSE of the short-term prediction results and the measured results are shown in Table 1, where the prediction accuracy was measured by (1-RMSE). In addition, the following situations (1) to (3) were considered when calculating the prediction accuracy: (1) Daytime and nighttime were considered. Since solar power generation outputs only during the daytime, the prediction accuracy for daytime hours in the prediction results is of great importance for flexible control. (2) Only the prediction error of the test set was considered. Although all sets include training set data, generally the accuracy of the prediction results of the training set is higher than that of the test set and may interfere with the decision. Therefore, the test set was used as the reference when calculating the prediction accuracy. (3) Spline interpolation was performed on the test set results after the prediction was completed, and the error after interpolation was calculated. There was no clear improvement in prediction accuracy before and after interpolation. As can be seen from the table, the provided short-term solar power generation prediction method is more accurate than conventional short-term and ultra-short-term prediction methods, and the prediction accuracy reached 85%.

[0046] Table 1 Prediction accuracy analysis [Table 1]

[0047] Subsequently, the predictive effect of the lower-layer LSTM network was evaluated. Here, the prediction results of the upper-layer BP neural network were used as input to the LSTM network. MATLAB® was used as the deep learning training platform, and the Adam optimizer with adaptive learning capabilities was attached. The LSTM model was debugged, including methods such as modifying the network structure, modifying the ratio of training-validation-test sets, increasing or decreasing the number of datasets, and adjusting the training step size. The final network parameters and calculation results are shown in Table 2.

[0048] Table 2 LSTM Neural Network Parameters [Table 2]

[0049] The future time zone predictions and historical time zone measurements of weather data, along with the short-term predictions of solar power generation from the upper layer BP neural network, were used as `train_inputs`, and the actual solar power generation values ​​were used as `train_targets` for training. 70% of the existing dataset was used as the training set, and the remaining 30% as the test set, with interpolation performed.

[0050] During the training process, the maximum number of training iterations (Max Epochs) was set to 50, and all training data was fed into the model for one training run in each epoch. To improve computational efficiency during training, the data was usually grouped according to the Mini Batch Size before being fed into the model for training, with a batch size of 128. The initial learn rate was set to 0.001, and this parameter determined the step size for parameter updates during the optimization process. The learn rate descent strategy (Learn Rate Schedule) was set to piecewise, and during training, the learn rate usually decreased as training progressed, which helps the model to make finer adjustments in the later stages of training. The learn rate drop factor was set to 0.1, and the learn rate descent period was set to 40, and after each learning rate descent period ended, the learn rate was updated by multiplying it by the drop factor. In addition, the data was disrupted after each epoch was set. By disrupting the data, the generalization performance of the model can be improved.

[0051] The structure of the LSTM neural network is shown in Table 3, where the number of neurons in the hidden layer was 20 and the number of linear layers was 1. The activation parameters of the Long Short-Term Memory (LSTM) neural network are usually expressed as (C) × (B) × (S), where (C) was the feature dimension or the number of hidden units; (B) was the rod size, i.e., the number of samples processed at one time; and (S) was the time step size, i.e., the sequence length.

[0052] Table 3 LSTM Neural Network Structure Parameters [Table 3]

[0053] The mean squared error curve and loss function curve during the training process are shown in Figures 7 and 8, respectively. It was found that a good training effect was achieved after approximately 10 repetitions. The training consisted of a total of 50 cycles, with 267 repetitions per cycle and a maximum of 13,350 repetitions. The total duration of the training was 2 minutes and 25 seconds.

[0054] The LSTM neural network prediction results are shown in Table 4. The coefficient of determination R 2 The reliability of the model was evaluated by this, and the value represents the degree of correlation between the model's predicted value and the actual value. R in the training set 2 The value was 0.9995, which means the model can interpret nearly 99.95% of the power changes in the training set, indicating a very good fitting effect of the model in the training set. R in the test set 2The value was 0.99906, which indicates that the model's generalization performance on the test set was also very high. Model prediction accuracy was measured by the mean absolute error (MAE), and the MAE for the training set and test set were 0.27026 and 0.34514, respectively. The relatively small MAE values ​​demonstrate excellent representation of the model's prediction accuracy. In short, this invention demonstrated satisfactory performance in the solar power generation prediction task based on an LSTM model. While both the coefficient of determination and the mean absolute error indicated the model's excellent performance, future research should explore other evaluation metrics such as maximum error and prediction accuracy at different light intensity levels to gain a more comprehensive understanding of the model's performance.

[0055] Table 4 LSTM Neural Network Calculation Results [Table 4]

[0056] Figures 9 and 10 show the prediction results corresponding to the training set and test set, respectively. The horizontal axis represents the number of predicted samples, and the time interval between adjacent points is 1 minute. This indicates that the underlying neural network achieves high-resolution solar power generation prediction, and this time resolution is suitable for the real-time control needs of solar power plants, giving it good application value in construction.

[0057] The solar power generation forecasting method designed in this invention achieves both a favorable forecasting time scale and accuracy compared to conventional forecasting methods. Furthermore, it acquires extremely short-term solar power generation forecasting data with a resolution of 1 minute, meeting the real-time control needs of solar power plants and possessing significant value for construction applications.

[0058] (Example 3) A third embodiment of the present invention differs from the two embodiments described above in the following respects.

[0059] The above functions may be implemented in the form of a software function unit and, if sold or used as an independent product, may be stored on a computer-readable storage medium. Based on this understanding, the proposed technical ideas of the present invention may be embodied in the form of a software product, which is essentially or partially contributes to the prior art, or a part thereof, which is stored on a storage medium and contains several instructions for causing a computer device (which may be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage mediums include a variety of media capable of storing program code, such as U disks, removable hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0060] The logic and / or steps shown in the flowchart or otherwise described herein can be considered, for example, a sequence of executable instructions for realizing a logical function and can be concretely realized on any computer-readable medium for use in or in combination with instruction execution systems, devices or equipment (e.g., computer-based systems, systems including processors, or other systems capable of reading instructions from instruction execution systems, devices or equipment and executing those instructions). In this specification, “computer-readable medium” may be an instruction execution system, device or equipment, or any device for containing, storing, communicating, propagating or transmitting programs for use in combination with such instruction execution systems, devices or equipment.

[0061] More specific examples of computer-readable media (a non-exclusive list) include electrical connections with one or more wires (electronic devices), portable computer cartridges (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber devices, and portable compact disk read-only memory (CDROM). Computer-readable media may also be paper or other suitable media on which the program can be printed, for example, by optically scanning the paper or other media, then editing, interpreting, or processing it in any other suitable manner to obtain the program electronically, and then storing it in computer memory.

[0062] It should be understood that each part of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, a plurality of steps or methods may be implemented by software or firmware stored in memory and executed by an appropriate instruction execution system. For example, if implemented in hardware, as in other embodiments, it may be implemented by any or a combination thereof of technologies known in the art, such as discrete logic circuits having logic gate circuits for implementing logic functions for data signals, dedicated integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0063] (Example 4) A fourth embodiment of the present invention provides a very short-term solar power forecasting system for real-time control, comprising a feature analysis module, a short-term power forecasting module, and a very short-term detailed forecasting module. The above feature analysis module normalizes radiation, temperature, humidity, and solar power generation data using the limits of year-round statistics, obtains the variables and solar power generation time series, and performs feature correlation analysis. The above short-term power forecasting module performs short-term solar power generation forecasting based on a BP neural network, takes weather data and historical output data as input, optimizes the loss function for each time period, and obtains upper-level short-term forecast values ​​with a resolution of 15 min. The above-described ultra-short-term detailed prediction module is characterized by the steps of performing spline interpolation on the upper layer results, extracting outliers using the methods of box plots and violin plots, removing the outliers using the method of moving time window filtering, improving the resolution of the prediction data based on a long-short-term cyclic neural network, achieving both prediction time scale and prediction accuracy, and obtaining lower-layer ultra-short-term solar power generation prediction data with a resolution of 1 min.

[0064] The above embodiments are merely for illustrating the technical concept of the present invention and are not limiting. While the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical concept of the present invention without departing from the spirit and scope of the invention, and that any such modifications or substitutions should fall within the scope of the claims of the present invention.

[0065] (Note) (Note 1) A method for predicting extremely short-term solar power generation for real-time control, The process involves normalizing radiation, temperature, humidity, and solar power generation data using the limits of year-round statistics, obtaining variable and solar power generation time series, performing feature correlation analysis, and constructing a two-layer solar power generation output prediction framework based on deep learning. The process involves performing short-term solar power generation forecasts based on a BP neural network, using weather data and historical output data as input, optimizing the loss function for each time period, and obtaining upper-level short-term forecast values ​​with a resolution of 15 minutes. A method for predicting extremely short-term solar power generation for real-time control, comprising the steps of: performing spline interpolation on the upper layer results; extracting outliers using the methods of box plots and violin plots; removing the outliers using the method of moving time window filtering; improving the resolution of the prediction data based on a long- and short-term cyclic neural network; achieving both prediction time scale and prediction accuracy; and obtaining lower layer extremely short-term solar power generation prediction data with a resolution of 1 min.

[0066] (Note 2) The normalization process described above is shown below: χ n =( χ n -χ min ) / (χ max -χ min ) Here, χ n is the nth data item to be processed, and χ max and χ min These are the local maximum and local minimum values ​​of the variable, respectively. A method for predicting extremely short-term solar power generation for real-time control, as described in Appendix 1, characterized by performing a correlation analysis on each meteorological factor and solar power generation output, and deciding to use numerical weather forecasts and historical measured weather data from the solar power plant's meteorological data, and effective power generation from historical output data, as input layer sample data for the training process of the BP neural network.

[0067] (Note 3) The aforementioned BP neural network includes an input layer, a hidden layer, and an output layer. The input layer is used to receive input data for the BP neural network. Since the output power of the solar power plant correlates with solar radiation intensity and is also affected by temperature, relative humidity, and cloud cover, the correlation analysis is performed by combining the relationship between each meteorological factor and the solar power output. The predicted solar radiation intensity in the numerical weather forecast for the solar power plant, the actual solar radiation intensity in the historical weather measurements, and the effective power generation in the historical output data are determined and used as sample data for the input layer during the training process of the BP neural network. The output layer performs short-term power forecasting, and the output data of the output layer is the future 24-hour output power sequence of the solar power plant, with a time interval of 15 minutes. The selection of the number of neurons in the hidden layer is related to the size and accuracy of the neural network. Based on Kolmogorov's theorem, if the number of input variables is n, the number of neurons in the hidden layer is 2n+1. The number of neurons in the hidden layer is determined using an iterative test method. The method for predicting extremely short-term solar power generation power for real-time control, as described in Appendix 2, is characterized in that the training process of the BP neural network is related to the prediction accuracy of the output power model of the solar power generation system, and is improved by combining a BP learning algorithm that increases the dynamic term and a BP learning algorithm that makes the learning rate variable, thereby improving the convergence speed and reducing the probability of falling into a local minimum.

[0068] (Note 4) Combining the aforementioned BP learning algorithms that increase the dynamic term and those that make the learning rate variable involves initializing the weight w and setting the current as the tth iteration, The process includes sequentially inputting P samples and calculating the output and backpropagation error of each layer; after all calculations for the P samples are complete, adjusting the connection weights of each layer according to the weight adjustment formula; calculating the output, backpropagation error, and total network error E(t) of each layer according to the new connection weights; terminating the training process if E(t) < ε or t > T; and re-inputting the P samples, calculating the output and backpropagation error of each layer, and performing training for a new cycle if E(t) ≥ ε and t ≤ T. The ultra-short-term solar power generation prediction method for real-time control described in Appendix 3, characterized in that, here, the number of training samples is P, T is the maximum number of training iterations, and ε is the system tolerance.

[0069] (Note 5) The step of removing outliers by employing a moving time window filtering method involves spline interpolation of the results obtained by the upper layer BP network, displaying the quartiles, median, and data density distribution based on the box plot and violin plot, making an initial judgment on the concentration and dispersion of data by observing the shape of the figures and peak values, identifying the outliers by combining statistical methods, and defining the outliers using the interquartile range. A method for predicting extremely short-term solar power generation for real-time control, as described in Appendix 4, characterized by: selecting 4 hours as the size of the travel time window; calculating the mean and median statistics of the data within the travel time window; replacing outliers using the statistics within the travel time window until all data processing is complete; then moving the travel time window forward by one time unit; and after data cleaning, removing the outliers.

[0070] (Note 6) The step of improving the resolution of the prediction data based on a long-short-term cyclic neural network represents that the memory cell unit is the core assembly of the LSTM network, and the input to the model is the time sequence input x t , hidden layer cell state h at time t-1 t-1 and memory unit c t-1 The output includes the memory unit state c. t and hidden layer state h t This includes, where c t and h t Each of these contains the long-term and short-term memory information of the model, and reading and modification of the memory cell unit are realized by controlling the input gate, forget gate and output gate, and information flows between the network, tanh represents the activation function of tanh, the input gate uses the sigmoid activation function to input parameters and controls the variable between [0,1], x t c t Control is achieved to the forget gate, which selectively forgets the neuron state from the previous time, and the specific representation is the memory unit c t-1 c tThis involves utilizing control to the output gate, which is used to output and control the parameter variable, i.e., c t h t Using the degree of impact, the calculation formulas are as follows: i t =σ(W ix x t +W ih h t-1 +b i ) f t =σ(W fx x t +W fh h t-1 +b f ) o t =σ(W ox x t +W oh h t-1 +b o ) Here, i t ,f t , o t These represent the state calculation results of the input gate, the forget gate, and the output gate, respectively, W ix and W ih , W fx and W fh , W ox and W oh and b i , b f , b o The ultra-short-term solar power generation power prediction method for real-time control according to Appendix 5, characterized in that represents the weight matrix and offset term of the corresponding gate, and σ represents the sigmoid activation function.

[0071] (Note 7) The step of improving the resolution of predictive data based on a long- and short-term cyclic neural network is: In the LSTM model, the output result of the memory module at time t is jointly determined by the output gate and the unit state, and the equation is as follows:

number

number

[0072] (Note 8) A system based on a real-time control ultra-short-term solar power generation forecasting method described in any one of Appendix 1 to 7, comprising a feature analysis module, a short-term power forecasting module, and an ultra-short-term detailed forecasting module, The aforementioned feature analysis module normalizes radiation, temperature, humidity, and solar power generation data using the limits of year-round statistics, obtains the variables and solar power generation time series, and performs feature correlation analysis. The aforementioned short-term power forecasting module performs short-term solar power generation forecasting based on a BP neural network, takes weather data and historical output data as input, optimizes the loss function for each time period, and obtains upper-level short-term forecast values ​​with a resolution of 15 min. The system is characterized by comprising the steps of: performing spline interpolation on the upper layer results, extracting outliers using box plots and violin plots, removing the outliers using a moving time window filtering method, improving the resolution of the prediction data based on a long- and short-term cyclic neural network, achieving both prediction time scale and prediction accuracy, and obtaining lower layer ultra-short-term solar power generation prediction data with a resolution of 1 min.

[0073] (Note 9) A computer device comprising a memory in which a computer program is stored and a processor, wherein the processor, when executing the computer program, performs a step according to any one of the methods described in Appendix 1 to 7.

[0074] (Note 10) A computer-readable storage medium in which a computer program is stored, characterized in that when the computer program is executed by a processor, it realizes a step according to any one of the methods described in Appendix 1 to 7.

Claims

1. A method for predicting extremely short-term solar power generation for real-time control, The process involves normalizing radiation, temperature, humidity, and solar power generation data using the limits of year-round statistics, obtaining variable and solar power generation time series, performing feature correlation analysis, and constructing a two-layer solar power generation output prediction framework based on deep learning. The process involves performing short-term solar power generation forecasts based on a BP neural network, using weather data and historical output data as input, optimizing the loss function for each time period, and obtaining upper-level short-term forecast values ​​with a resolution of 15 min. A method for predicting extremely short-term solar power generation for real-time control, characterized by comprising the steps of: performing spline interpolation on the upper layer results; extracting outliers using the methods of box plots and violin plots; removing the outliers using the method of moving time window filtering; improving the resolution of the prediction data based on a long- and short-term cyclic neural network; achieving both prediction time scale and prediction accuracy; and obtaining lower layer extremely short-term solar power generation prediction data with a resolution of 1 min.

2. The normalization process described above is shown below: x ｎ =(x) ｎ -x ｍｉｎ ) / (x ｍａｘ -x ｍｉｎ ) Here, χ ｎ is the nth data item to be processed, and χ ｍａｘ and χ ｍｉｎ These are the local maximum and local minimum values ​​of the variable, respectively. A method for predicting extremely short-term solar power generation power for real-time control, as described in claim 1, characterized in that a correlation analysis is performed on each meteorological factor and solar power generation output, and it is determined that numerical weather forecasts and historical measured weather data of the solar power plant, and effective power generation in historical output data will be used as input layer sample data for the training process of the BP neural network.

3. The BP neural network includes an input layer, a hidden layer, and an output layer. The input layer is used to receive input data for the BP neural network. Since the output power of the solar power plant is correlated with solar radiation intensity and is also affected by temperature, relative humidity, and cloud layer, the correlation analysis is performed by combining the relationship between each meteorological factor and the solar power output. The predicted solar radiation intensity in the numerical weather forecast for the solar power plant, the actual solar radiation intensity in the historical weather measurements, and the effective power generation in the historical output data are determined and used as sample data for the input layer during the training process of the BP neural network. The output layer performs short-term power forecasting, and the output data of the output layer is the future 24-hour output power sequence of the solar power plant, with a time interval of 15 minutes. The selection of the number of neurons in the hidden layer is related to the size and accuracy of the neural network. Based on the Kolmogolov theorem, if the number of input variables is n, the number of neurons in the hidden layer is 2n+1. The number of neurons in the hidden layer is determined using an iterative test method. The method for predicting extremely short-term solar power generation power for real-time control, as described in claim 2, is characterized in that the training process of the BP neural network is related to the prediction accuracy of the output power model of the solar power generation system, and is improved by combining a BP learning algorithm that increases the dynamic term and a BP learning algorithm that makes the learning rate variable, thereby improving the convergence speed and reducing the probability of falling into a local minimum.

4. Combining the aforementioned BP learning algorithm with an increasing dynamic term and a BP learning algorithm with a variable learning rate involves initializing the weight w and setting the current state as the tth iteration. The process includes: sequentially inputting P samples and calculating the output and backpropagation error of each layer; adjusting the connection weights of each layer according to a weight adjustment formula after all calculations for the P samples are complete; calculating the output, backpropagation error, and total network error E(t) of each layer according to the new connection weights; terminating the training process if E(t) < ε or t > T; and re-inputting P samples, calculating the output and backpropagation error of each layer, and performing training for a new cycle if E(t) ≥ ε and t ≤ T. The method for predicting extremely short-term solar power generation for real-time control according to claim 3, characterized in that, here, the number of training samples is P, T is the maximum number of training iterations, and ε is the system tolerance.

5. The step of removing outliers by employing a moving time window filtering method involves spline interpolation of the results obtained by the upper layer BP network, displaying the quartiles, median, and data density distribution based on the box plot and violin plot, making an initial judgment on the concentration and dispersion of data by observing the shape of the figures and peak values, identifying the outliers by combining statistical methods, and defining the outliers using the interquartile range. A method for predicting extremely short-term solar power generation for real-time control, as described in claim 4, comprising: selecting 4 hours as the size of the travel time window; calculating the mean and median statistics of the data within the travel time window; replacing outliers using the statistics within the travel time window until all data processing is complete; then moving the travel time window forward by one time unit; and after data cleaning, removing the outliers.

6. The step of improving the resolution of prediction data based on a long short-term recurrent neural network indicates that a memory cell unit is the core assembly of the LSTM network, and the input of the model is the sequence input x at time t ｔ , the hidden layer cell state h at time t-1 ｔ－１ and the memory unit c ｔ－１ , and the output includes the memory unit state c ｔ and the hidden layer state h ｔ , where c ｔ and h ｔ each contain the long-term and short-term memory information of the model, and realize the reading and modification of the memory cell unit by controlling the input gate, forget gate and output gate, perform the information flow between networks, tanh represents the activation function of tanh, the input gate inputs parameters using the sigmoid activation function, controls the variable between [0, 1], and realizes the control of x ｔ to c ｔ , the forget gate is to selectively forget the neuron state of the previous time, and the specific expression form is to utilize the control of the memory unit c ｔ－１ to c ｔ , the output gate is used to output and control parameter variables, that is, to utilize the degree of influence on h ｔ of c ｔ , and the calculation formulas are as follows respectively i ｔ =σ(W ｉｘ x ｔ +W ｉｈ h ｔ－１ +b ｉ ) f ｔ =σ(W ｆｘ x ｔ +W ｆｈ h ｔ－１ +b ｆ ) o ｔ =σ(W ｏｘ x ｔ +W ｏｈ h ｔ－１ +b ｏ ) Here, i ｔ , f ｔ , o ｔ These represent the state calculation results of the input gate, the forget gate, and the output gate, respectively, W ｉｘ and W ｉｈ , W ｆｘ and W ｆｈ , W ｏｘ and W ｏｈ and b ｉ , b ｆ , b ｏ The ultra-short-term solar power generation power prediction method for real-time control according to claim 5, wherein each of the following represents the weight matrix and offset term of the corresponding gate, and σ represents the sigmoid activation function.

7. The step of improving the resolution of predictive data based on a long- and short-term cyclic neural network is: In the LSTM model, the output result of the memory module at time t is jointly determined by the output gate and the unit state, and the formula is as follows: [Math 1] Here, [Math 2] represents the unit state input at time t, where tanh is the hyperbolic tangent activation function, and W ｃ , b ｃ The ultra-short-term solar power generation power prediction method for real-time control according to claim 6, wherein * represents the state weight matrix and offset term of the input layer, respectively, and * indicates that each element is multiplied by position.

8. A system based on a real-time control method for predicting extremely short-term solar power generation according to any one of claims 1 to 7, comprising a feature analysis module, a short-term power prediction module, and an extremely short-term detailed prediction module, The aforementioned feature analysis module normalizes radiation, temperature, humidity, and solar power generation data using the limits of year-round statistics, obtains the variables and solar power generation time series, and performs feature correlation analysis. The aforementioned short-term power forecasting module performs short-term solar power generation forecasting based on a BP neural network, takes weather data and historical output data as input, optimizes the loss function for each time period, and obtains upper-level short-term forecast values ​​with a resolution of 15 min. The system is characterized by comprising the steps of: performing spline interpolation on the upper layer results, extracting outliers using the methods of box plots and violin plots, removing the outliers using the method of moving time window filtering, improving the resolution of the prediction data based on a long- and short-term cyclic neural network, achieving both prediction time scale and prediction accuracy, and obtaining lower layer ultra-short-term solar power generation prediction data with a resolution of 1 min.

9. A computer device comprising a memory in which a computer program is stored and a processor, wherein the processor, upon executing the computer program, realizes the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium in which a computer program is stored, characterized in that when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are realized.

Images