Intra-day photovoltaic power rolling prediction method considering residual correction

Abstract

The invention discloses an intra-day photovoltaic power rolling prediction method considering residual correction. The method comprises the following steps: firstly, acquiring numerical weather forecast of a target photovoltaic site in a prediction window and actual output of a photovoltaic site in a corresponding historical window as input; an intra-day photovoltaic power rolling prediction model based on the hybrid neural network is constructed, future numerical weather forecast in a prediction window and photovoltaic power station output in a historical window are used as input in each prediction, and future predicted photovoltaic power station power is output; and taking the residual error between the predicted value and the actual value as a prediction target, inputting the preliminarily predicted power of the photovoltaic power station into the residual error correction model, outputting a predicted residual error value, and combining and adding the predicted residual error value and the preliminarily predicted value to obtain a photovoltaic power prediction result considering residual error correction. According to the prediction model considering residual error correction, the accuracy of intra-day photovoltaic power rolling prediction can be improved, and the requirement of photovoltaic prediction is met.

Description

Technical Field

[0001] The present invention belongs to the field of new energy power generation prediction, and particularly relates to an intraday photovoltaic power rolling prediction method considering residual correction. Background Technique

[0002] With the transformation of the global energy structure, traditional fossil fuels are gradually being replaced by clean energy. As a green and low-carbon energy form, photovoltaic power generation is gradually becoming an important part of the global energy market. With the continuous development of photovoltaic technology and the significant reduction in costs, the construction and operation scale of photovoltaic power plants are constantly expanding. Photovoltaic power generation plays an increasingly important role in promoting the transformation of the energy structure and reducing greenhouse gas emissions.

[0003] However, the characteristics of photovoltaic power generation are volatility and intermittency, which means that its power generation is strongly affected by factors such as weather, season, sunlight intensity, and cloud cover changes, and there is often a certain degree of unpredictability and instability. For example, rapid changes in clouds, sudden meteorological conditions, etc. can all have a greater impact on the output of photovoltaic power generation, thereby affecting the stability of the power grid. Therefore, how to accurately predict the power of photovoltaic power generation has become one of the core issues in power system dispatching, load balancing, and energy storage management.

[0004] In this context, photovoltaic power prediction technology has been widely studied, and its prediction methods can be roughly divided into two categories: physical methods and statistical methods. Physical methods are based on the physical principles of photovoltaic power generation and mathematical models of meteorological conditions, and describe the working principle of photovoltaic cells and the influence of environmental conditions on power generation through the establishment of accurate physical models. Common physical methods include radiation transfer models, energy balance models, photovoltaic system electrical models, etc. Statistical methods mainly rely on historical data and statistical analysis techniques, and predict future power output by analyzing the statistical laws of historical photovoltaic power generation data and meteorological data. Common statistical methods include regression analysis, time series analysis, Bayesian networks, etc.

[0005] In recent years, with the rapid development of artificial intelligence technology, photovoltaic power prediction methods based on machine learning have gradually become a research hotspot. Machine learning methods can overcome the limitations of traditional physical models and statistical methods in terms of non-linearity and complexity by automatically learning potential patterns in a large amount of data. Summary of the Invention

[0006] The present invention proposes an intraday photovoltaic power rolling prediction method considering residual correction, aiming to construct a method that can perform rolling prediction step by step, and at the same time correct the residuals to further improve the prediction accuracy.

[0007] The specific technical solution adopted by the present invention is as follows:

[0008] An intraday photovoltaic power rolling prediction method considering residual correction, which comprises the following steps:

[0009] S1. For the target photovoltaic site, obtain the numerical weather forecast within the prediction window after the specified time as the meteorological feature, and obtain the output of the photovoltaic site within the historical window before the specified time as the output time series feature;

[0010] S2. Respectively input the meteorological feature and the output time series feature into the trained intraday rolling prediction model of photovoltaic power to obtain the preliminary prediction sequence of the power generation power of the target photovoltaic site within the prediction window;

[0011] S3. Input the preliminary prediction sequence of the power generation power of the target photovoltaic site within the prediction window into the trained photovoltaic power residual correction model to obtain the residual correction value sequence of the power generation power of the target photovoltaic site within the prediction window;

[0012] S4. Superimpose the preliminary prediction sequence of the power generation power of the target photovoltaic site within the prediction window and the residual correction value sequence of the power generation power, and finally obtain the photovoltaic power prediction result of the target photovoltaic site within the prediction window.

[0013] Preferably, the length of the prediction window is 24 hours, and the step intervals of the numerical weather forecast and the preliminary prediction sequence of the power generation power within the prediction window are both 15 minutes.

[0014] Preferably, the length of the historical window is 24 hours, and the step interval of the output of the photovoltaic site within the historical window is 15 minutes.

[0015] Preferably, the numerical weather forecast includes solar irradiance, temperature, relative humidity and wind speed.

[0016] Preferably, the intraday rolling prediction model of photovoltaic power is trained based on a hybrid deep neural network. The training of the hybrid deep neural network includes two input branches cascaded by one-dimensional convolutional layers and pooling layers, as well as a long short-term memory neural network layer and a fully connected layer; the meteorological feature and the output time series feature are respectively input through one input branch, and the convolutional layer and the pooling layer extract local features and patterns therefrom, and then the outputs of the two input branches are feature-stitched and input into the two-layer long short-term memory neural network layer to learn the long-term time features of the photovoltaic output, and finally, after two-layer dimensionality reduction by the fully connected layer, the preliminary prediction sequence of the power generation power of the target photovoltaic site within the prediction window is output.

[0017] Preferably, the photovoltaic power residual correction model is trained based on a multi-layer perceptron.

[0018] Preferably, both the intraday rolling prediction model of photovoltaic power and the residual correction model of photovoltaic power are pre-supervised trained on a sample data set. First, the intraday rolling prediction model of photovoltaic power is trained using the sample data set with true value labels. After the training is completed, the intraday rolling prediction model of photovoltaic power is used to predict each sample in the sample data set, and the residual correction value within the prediction window is calculated based on the predicted value and the true value label and used as the true value label during the training of the residual correction model of photovoltaic power, and the residual correction model of photovoltaic power is further trained.

[0019] Preferably, the pooling layer adopts a max pooling operation.

[0020] This method can achieve intraday photovoltaic power rolling prediction with a gradually increasing length, and at the same time correct the residuals of the prediction results during the rolling process, meeting the requirements for improving the accuracy of photovoltaic prediction. In some embodiments, the present invention can achieve intraday photovoltaic power prediction for the next 0-24h every 15 minutes. Therefore, this method fills the prediction gap for the 4-24h period that is not included in the existing ultra-short-term (0-4h) and short-term (1-3d) predictions. The results show that compared with the prediction model without considering residual correction, the prediction accuracy of the prediction model considering residual correction has increased by 0.4% - 1%, which can meet the requirements for improving the accuracy of photovoltaic prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is the flowchart of the steps of the present invention;

[0022] Figure 2 is the structure diagram of the hybrid neural network photovoltaic power rolling prediction model proposed by the present invention;

[0023] Figure 3 is the structure diagram of the photovoltaic power residual correction rolling prediction model proposed by the present invention;

[0024] Figure 4 is the change curve of the NRMSE error of the intraday 0-24h prediction in the embodiment of the present invention.

[0025] Figure 5 is the comparison chart of the prediction curves at the 4h and 24h prediction scales in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0026] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0027] Further, to enable the public to have a better understanding of the present invention, in the following detailed description of the present invention, some specific details are described in detail. Those skilled in the art can fully understand the present invention without the description of these details.

[0028] As Figure 1 shown, in a preferred implementation manner of the present invention, a photovoltaic population prediction method based on spatial correlation is provided, including the following steps S1 to S4. The specific implementation processes thereof will be described separately below.

[0029] S1. For the target photovoltaic site, obtain the numerical weather prediction (NWP) within the prediction window after the specified time as the meteorological feature, and obtain the output of the photovoltaic site within the historical window before the specified time as the output time series feature.

[0030] It should be noted that the specific lengths of the above prediction window and historical window, as well as the interval step lengths of the time series data therein, can be reasonably adjusted according to actual needs.

[0031] In the embodiment of the present invention, the length of the prediction window is set to 24 hours, and the step intervals of the numerical weather prediction and the preliminary prediction sequence of the power generation power within the prediction window are both set to 15 minutes. The length of the historical window is 24 hours, and the step interval of the output of the photovoltaic site within the historical window is also set to 15 minutes. Based on this window length and the corresponding interval, in the present invention, it is necessary to obtain the numerical weather prediction of the target photovoltaic site within the prediction window and the actual output of the photovoltaic site in the previous 24 hours corresponding to the history, and the numerical weather prediction is the latest data available at the current T moment. Specifically, when predicting at the T moment, it is necessary to obtain the numerical weather prediction for the time period from T to T + 24h, and the historical photovoltaic output for the time period from T - 24h to T. At a 15-minute resolution, there will be 96 data points in 24 hours. If there are n types of meteorological parameter types in the numerical weather prediction, the input dimension of the numerical weather prediction as the meteorological feature is n * 96, and the input dimension of the historical photovoltaic output as the output time series feature is 1 * 96. The meteorological feature and the output time series feature are respectively input into the model, and the prediction dimension is 1 * 96.

[0032] In addition, it should be noted that for the numerical weather prediction, the meteorological parameters mainly include solar irradiance, temperature, zenith angle, relative humidity, wind speed, etc. According to the obtained meteorological factors, it is necessary to analyze the correlation between various meteorological factors and photovoltaic output to select the final input features. In the embodiment of the present invention, the finally used meteorological parameters are solar irradiance, temperature, relative humidity, and wind speed, so n is 4.

[0033] S2. Input the meteorological features and the output time series features into the trained intra-day rolling prediction model of photovoltaic power respectively to obtain the preliminary prediction sequence of the power generation power of the target photovoltaic site within the prediction window.

[0034] In the present invention, it is necessary to construct a hybrid deep neural network according to the meteorological features of numerical weather prediction and the output time series features of historical photovoltaic output, and obtain the intra-day rolling prediction model of photovoltaic power by training the hybrid deep neural network, so as to perform rolling prediction on the power generation power sequence within the prediction window.

[0035] In the embodiment of the present invention, since the interval step within the prediction window is 15 minutes, the intra-day rolling prediction model of photovoltaic power can realize the rolling prediction of the power of the photovoltaic power station in the future 0-24 hours every 15 minutes. If it is necessary to predict the power of the future photovoltaic power station at time T, the predicted future numerical weather prediction of T-T+24 hours and the output of the photovoltaic power station in the historical T-24h-T period can be used as inputs each time for prediction, and the predicted power of the future photovoltaic power station of T-T+24h can be output.

[0036] It should be noted that the specific form of the above hybrid deep neural network needs to be reasonably designed according to the specific input data and prediction object of the present invention. In the embodiment of the present invention, according to the meteorological features in numerical weather prediction and the time series features of historical photovoltaic output, an intra-day rolling prediction model of photovoltaic power based on a hybrid deep neural network is proposed. As Figure 2 shown, the training of the hybrid deep neural network includes two input branches cascaded by one-dimensional convolutional layers and pooling layers (using max pooling operations), as well as long short-term memory neural network layers (LSTM) and fully connected layers (FC layers); the meteorological features and the output time series features are respectively input through one input branch, and local features and patterns are extracted from them by the convolutional layers and pooling layers, and then the outputs of the two input branches are feature-stitched (Combination) and input into two layers of long short-term memory neural network layers to learn the long-term time features of photovoltaic output, and finally, after dimensionality reduction through two layers of fully connected layers, the preliminary prediction sequence of the power generation power of the target photovoltaic site within the prediction window is output.

[0037] This hybrid deep neural network combines the advantages of convolutional neural networks (CNNs) and long short-term memory neural networks (LSTMs) in dealing with spatial and temporal problems. First, numerical weather prediction data and historical photovoltaic power output data are preliminarily processed through convolutional layers and pooling layers respectively, and the processed information is fused. Then, through further processing by two LSTM layers and two fully connected layers, the photovoltaic power prediction results are finally output. The convolutional layer slides a convolutional kernel over the input data to extract local features, which helps to automatically identify local patterns in numerical weather prediction (NWP) and historical photovoltaic power output, enabling the model to effectively capture the relationship between meteorological factors and photovoltaic power generation. After the max-pooling operation, the output of the convolutional layer is dimensionally reduced, which not only reduces the sensitivity to the input data but also lowers the computational complexity, thus improving the robustness and generalization ability of the model. Next, the long short-term memory network (LSTM) layer can capture the long-term time dependencies in the photovoltaic power data. After two fully connected layers perform non-linear and linear transformations on the data, the model finally outputs the predicted values of the intra-day photovoltaic power. Through this series of processes, the model effectively integrates feature information at different levels, thus showing high accuracy in photovoltaic power prediction.

[0038] S3. Input the preliminary prediction sequence of the power generation power of the target photovoltaic site within the prediction window into the trained photovoltaic power residual correction model to obtain the sequence of the photovoltaic power residual correction values of the target photovoltaic site within the prediction window.

[0039] In an embodiment of the present invention, the above photovoltaic power residual correction model can be trained based on a multi-layer perceptron (MLP). The input of the multi-layer perceptron is the preliminary prediction sequence of the power generation power within the prediction window output by the hybrid deep neural network, and the output is the sequence of the photovoltaic power residual correction values of the target photovoltaic site within the prediction window. Therefore, it is necessary to construct training samples corresponding to this input and output.

[0040] Therefore, in an embodiment of the present invention, the model structure of the photovoltaic power residual correction model and the overall photovoltaic power residual correction model is as Figure 3 shown. The above intra-day rolling prediction model of photovoltaic power and the photovoltaic power residual correction model can be supervised and trained on the sample data set successively. The specific training method is as follows:

[0041] First, use the sample data set with true value labels (i.e., the true values of the power generation power of the target photovoltaic site within the prediction window) to train the intra-day rolling prediction model of photovoltaic power;

[0042] Then, after the training is completed, use the intraday rolling prediction model of photovoltaic power to predict each sample in the sample dataset, calculate the residual correction value within the prediction window based on the predicted value and the true value label, and use it as the true value label during the training of the photovoltaic power residual correction model. Specifically, the trained intraday rolling prediction model of photovoltaic power can be used to predict the preliminary prediction sequence of power generation for each sample in the sample dataset, calculate the residual sequence between the preliminary prediction sequence of power generation and the actual value sequence of power generation, and use it as the new label value class of the residual to train the MLP, so that the MLP can further calculate the residual sequence between the predicted value and the true value of the power generation from T to T+24h based on the output of the intraday rolling prediction model of photovoltaic power. This residual sequence is also a 1*96-dimensional sequence.

[0043] Finally, use the sample dataset with the true value label with the residual correction value to retrain the photovoltaic power residual correction model. The sample input for this step of training is the preliminary prediction sequence of power generation within the prediction window, and the true value label is the residual correction value sequence of power generation within the prediction window.

[0044] Therefore, in the embodiments of the present invention, based on constructing the photovoltaic power residual correction model, the preliminarily predicted power of the photovoltaic power station is input into the residual correction model, and the predicted residual value is output. By learning the mapping relationship between the predicted value and the residual value, it is possible to realize the rolling prediction of the power of the photovoltaic power station for the next 0-24h with a 15-minute interval.

[0045] S4. Superimpose the preliminary prediction sequence of power generation and the residual correction value sequence of power generation of the target photovoltaic site within the prediction window to finally obtain the photovoltaic power prediction result of the target photovoltaic site within the prediction window.

[0046] In the embodiments of the present invention, the 0-24h power of the photovoltaic power station preliminarily predicted by the intraday rolling prediction model of photovoltaic power is input into the residual correction model, and then the predicted 1*96-dimensional residual value sequence Et_f is output. Combine and add it to the 1*96-dimensional preliminary prediction value sequence Yt_f directly output by the intraday rolling prediction model of photovoltaic power to obtain the photovoltaic power prediction result considering residual correction. The final result is also a 1*96-dimensional sequence.

[0047] To prove the advantages of the above technical solutions in the present invention, the above method will be applied to specific examples below to demonstrate its technical effects. The specific process is as described above and will not be repeated. Below, the data and technical effects will be mainly shown.

[0048] Embodiment

[0049] This embodiment constructs and trains the model according to the methods shown in S1 to S4 above, and analyzes the data set of a photovoltaic site in Gansu Province, China based on the final trained model. These data come from the National Key R&D Program-Large-Scale Wind Power / Photovoltaic Multi-Time Scale Power Supply Capacity Prediction Technology (2022YFB2403000). The data covers the period from March 26 to December 31, 2022, with a time resolution of 15 minutes. The numerical weather forecast data comes from the China Electric Power Research Institute and is selected according to the latitude and longitude of the site, with the same data length and resolution. The latitude, longitude and installed capacity of the site are shown in Table 1. The data of the first 21 days of every 28 days are used for the training set, and the data of the last 7 days are used for the test set.

[0050] Table 1 Latitude, longitude and installed capacity of the sites

[0051]

[0052] For the convenience of description, the training obtained in the method of the present invention is Figure 3 The photovoltaic power residual correction model and the photovoltaic power residual correction model cascade model shown are collectively referred to as CLSTM_C (the one-dimensional convolutional layer is denoted as Conv1D, two layers of LSTM are designed and denoted as LSTM_1 and LSTM_2, and the two fully connected layers are denoted as MLP_1 and MLP_2, respectively). In order to verify the prediction accuracy of the method proposed in the present invention, two auxiliary prediction models without residual correction are introduced for comparison. The first model is called CNN, which includes a convolutional layer and two fully connected layers. The second model is called CLSTM, which includes a convolutional layer, two long short-term memory networks and two fully connected layers. These two models perform single-site prediction for this photovoltaic site respectively. The hyperparameter design of each model is shown in Table 2.

[0053] Table 2 Hyperparameter design of prediction model

[0054]

[0055] In order to compare the accuracy difference between the algorithm proposed in the present invention and the auxiliary prediction model, the accuracy evaluation standard adopts the normalized root mean square error (NRMSE) and the normalized mean absolute error (NMAE) evaluation:

[0056]

[0057] Where n is the number of prediction points, P cap is the installed capacity of photovoltaic power station equipment, P ri is the actual power value of the i-th point, P pi is the predicted power value of the i-th point.

[0058] The present invention performs a rolling prediction of the next 24 hours every 15 minutes. Therefore, there are 96 prediction scales in the range of 0 - 24 hours, namely the 15th minute, 30th minute, 45th minute, 1 hour, …, 24 hours. Taking the NRMSE and NMAE at the 4th hour, 8th hour, 16th hour, and 24th hour as the accuracy evaluation criteria, the results of the proposed model and the comparison model are shown in Table 3.

[0059] Table 3 Comparison of Model Accuracy

[0060]

[0061] To more clearly show the accuracy improvement of the proposed method compared with the benchmark model, Figure 4 a curve graph showing the change of the NRMSE error of various methods in the intraday 0 - 24 hour prediction is plotted. Taking the 4th hour and 24th hour prediction scales, the prediction curves of typical days are selected and shown together with the actual power curves, as Figure 5 shown.

[0062] In summary, as can be seen from the above table and the power curve graph, the rolling prediction model considering residual correction has higher accuracy and can meet the requirements of photovoltaic prediction for prediction accuracy.

[0063] The above-described embodiments are only a preferred solution of the present invention, but they are not intended to limit the present invention. Those of ordinary skill in the relevant technical fields can still make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all technical solutions obtained by means of equivalent replacement or equivalent transformation fall within the protection scope of the present invention.

Claims

1. A method for intraday photovoltaic power rolling forecasting considering residual correction, characterized in that: The steps include: S1. For the target photovoltaic site, obtain the numerical weather forecast in the prediction window after the specified time as the meteorological feature, and obtain the photovoltaic site output in the historical window before the specified time as the output time series feature; S2, inputting the meteorological characteristics and the output time series characteristics into the trained photovoltaic power intraday rolling prediction model respectively, to obtain a preliminary prediction sequence of the power generation power of the target photovoltaic site within the prediction window; S3, inputting the preliminary prediction sequence of the power generation of the target photovoltaic site within the prediction window into the trained photovoltaic power residual correction model to obtain the power generation residual correction value sequence of the target photovoltaic site within the prediction window; S4. Superimposing the preliminary prediction sequence of the power generation of the target photovoltaic site within the prediction window and the power generation residual correction value sequence, and finally obtaining the photovoltaic power prediction result of the target photovoltaic site within the prediction window.

2. The intraday photovoltaic power rolling forecasting method considering residual correction according to claim 1 is characterized in that: The length of the forecast window is 24 hours, and the step intervals of the numerical weather forecast and the preliminary power generation forecast sequences within the forecast window are both 15 minutes.

3. The intraday photovoltaic power rolling forecasting method considering residual correction according to claim 1 is characterized in that: The length of the historical window is 24 hours, and the step interval of the photovoltaic site output within the historical window is 15 minutes.

4. The intraday photovoltaic power rolling forecasting method considering residual correction according to claim 1 is characterized in that: The numerical weather forecast includes solar irradiance, temperature, relative humidity and wind speed.

5. The intraday photovoltaic power rolling forecasting method considering residual correction according to claim 1 is characterized in that: The photovoltaic power intraday rolling prediction model is obtained based on hybrid deep neural network training, which includes two input branches formed by cascading one-dimensional convolutional layers and pooling layers, as well as long short-term memory neural network layers and fully connected layers; the meteorological characteristics and the output time series characteristics are respectively input through an input branch, and local features and patterns are extracted therefrom by the convolutional layer and the pooling layer, and then the outputs of the two input branches are feature spliced ​​and input into two layers of long short-term memory neural network layers to learn the long-term time characteristics of photovoltaic output, and finally, after dimensionality reduction through two layers of fully connected layers, a preliminary prediction sequence of the power generation power of the target photovoltaic site within the prediction window is output.

6. The intraday photovoltaic power rolling forecasting method considering residual correction according to claim 1 is characterized in that: The photovoltaic power residual correction model is based on multi-layer perceptron training.

7. The intraday photovoltaic power rolling forecasting method considering residual correction according to claim 1 is characterized in that: The photovoltaic power intraday rolling prediction model and the photovoltaic power residual correction model are both supervisedly trained on a sample data set in advance, wherein the photovoltaic power intraday rolling prediction model is first trained using a sample data set with true value labels, and after the training, the photovoltaic power intraday rolling prediction model is used to predict each sample in the sample data set, and the residual correction value in the prediction window is calculated based on the predicted value and the true value label and used as the true value label during the training of the photovoltaic power residual correction model, and the photovoltaic power residual correction model is further trained to obtain the photovoltaic power residual correction model.

8. The intraday photovoltaic power rolling forecasting method considering residual correction according to claim 5 is characterized in that: The pooling layer adopts the maximum pooling operation.

Images