Method and system for predicting irradiance of solar power plant

Abstract

The invention discloses a solar power plant irradiance prediction method and system, and relates to the technical field of new energy power generation, and the method comprises the steps: inputting collected meteorological data at k moments in continuous multiple days before a # imgabs0 # day into a long-time irradiance prediction model, and obtaining a long-time irradiance prediction result of the # imgabs1 # day; dividing the long-time irradiance prediction result of the # imgabs2 # day according to k moments to obtain irradiance prediction results of k moments in the # imgabs3 # day; selecting the irradiance prediction result at the ith moment from the obtained irradiance prediction results at the k moments in the # imgabs4 # day; and predicting the irradiance at the (i + 1) th moment by using a short-time irradiance prediction model according to the irradiance prediction result at the ith moment, the real meteorological data at the ith moment and the feature information of the real cloud picture data at the ith moment. According to the invention, meteorological information and cloud picture information can be combined to improve the prediction precision of the short-term irradiance.

Description

Technical Field

[0001] The present invention relates to the technical field of new energy power generation, and particularly to a method and system for predicting the irradiance of a solar power plant. Background Art

[0002] The change of irradiance is mainly affected by factors such as the solar altitude angle, weather conditions, seasonal changes, geographical location, atmospheric absorption and scattering. Although the solar radiation changes regularly with seasons and time, the change of atmospheric conditions will still affect the solar radiation. Cloud cover is the main reason for the sudden change of irradiance.

[0003] In the prior art, the method for predicting solar irradiance based on meteorological values insufficiently considers cloud information and does not take into account the generation, dissipation and movement of clouds on the irradiance of a heliostat field, and is only applicable to scenarios where the irradiance changes relatively smoothly. The method for predicting solar irradiance based on cloud images is affected by factors such as the performance of satellite sensors, atmospheric conditions and terrain, which may lead to inaccurate observation of clouds, and the influence of the atmosphere and terrain will also interfere with signal transmission and reception, thereby affecting the accuracy of data.

[0004] In summary, the current technical problem is how to combine meteorological information and cloud image information to improve the prediction accuracy of short-term irradiance. Summary of the Invention

[0005] Embodiments of the present invention provide a method and system for predicting the irradiance of a solar power plant, which can solve the problem in the prior art of how to combine meteorological information and cloud image information to improve the prediction accuracy of short-term irradiance.

[0006] Embodiments of the present invention provide a method for predicting the irradiance of a solar power plant, including the following steps: Collect irradiance data and meteorological data of a solar power plant at multiple consecutive moments within multiple days before the day before, and collect cloud image data a set number of times between two consecutive collection moments of meteorological data to obtain the characteristic information of the cloud image data; wherein, k 2; k≥ 2; Add an attention mechanism to a long short-term memory network (LSTM) to construct an optimized LSTM model and train it to obtain a long-term irradiance prediction model; according to the meteorological data of multiple consecutive moments collected, use the long-term irradiance prediction model to predict the long-term irradiance of the k day, and divide it according to moments to obtain the irradiance prediction results of k moments within the day; k 42; Construct a recurrent neural network (RNN) and train it to obtain a short-term irradiance prediction model; collect meteorological data at within the day k at the i th moment of the i day. Based on the irradiance prediction result at the i th moment, the real meteorological data at the i th moment, and the characteristic information of the real cloud map data at the i th moment, use the short-term irradiance prediction model to predict the irradiance at the +1th moment; wherein, the irradiance prediction results at k moments within the day include the irradiance prediction results at the i th moment, and 1 ≤ i ≤ k .

[0007] Further, after predicting the irradiance at the i +1th moment, it further includes: Obtain the real irradiance at the i +1th moment, the real meteorological data at the i +1th moment, and the characteristic information of the real cloud map data at the i +1th moment; According to the real irradiance at the i +1th moment, the real meteorological data at the i +1th moment, and the characteristic information of the real cloud map data at the i +1th moment, update the training set of the short-term irradiance prediction model, and train the short-term irradiance prediction model with the updated training set; Repeatedly use the updated short-term irradiance prediction model to sequentially predict the irradiance at subsequent moments of the i +1th moment.

[0008] Further, the specific steps for obtaining the long-term irradiance prediction model include: According to the meteorological data at consecutive days before the k th moment and the irradiance data at the corresponding moments, use the LSTM model optimized by training with the Backpropagation Through Time (BPTT) algorithm to predict the long-term irradiance on the th day, and obtain the long-term irradiance prediction model.

[0009] Further, after predicting the irradiance at the The long-term irradiance for a certain day, and the specific steps include: inputting time, temperature, humidity, wind speed, wind direction, atmospheric pressure, irradiance, and solar altitude angle into the long-term irradiance prediction model; capturing the long-term dependencies in the time series data through LSTM, and paying attention to the information affecting irradiance through the attention mechanism to predict the long-term irradiance for a certain day.

[0010] Furthermore, the specific steps for obtaining the characteristic information of the cloud map data include: identifying N cloud layers in the cloud map data; taking the average brightness, pixel area, motion vector, relative distance, and cloud height of each cloud layer as the characteristic information.

[0011] Furthermore, the specific steps for the characteristic information of the real cloud map data at this moment include: selecting the last real cloud map data from the collected real cloud map data of a set number of times; obtaining the characteristic information of the last real cloud map data.

[0012] An embodiment of the present invention provides a prediction system for the irradiance of a solar power plant, including: A data acquisition module, used to collect the irradiance data and meteorological data of a solar power plant at successive multiple days before a certain day k at multiple moments, collecting the cloud map data of a set number of times between the acquisition moments of two successive meteorological data, and obtaining the characteristic information of the cloud map data; where k≥ 2; A long-term irradiance prediction module, used to add an attention mechanism to the long short-term memory network LSTM to construct and train an optimized LSTM model to obtain a long-term irradiance prediction model; according to the meteorological data at k multiple moments in successive multiple days collected, using the long-term irradiance prediction model to predict the long-term irradiance for a certain day, and dividing it according to multiple moments to obtain the irradiance prediction results at k multiple moments within a certain day; within a certain day k multiple moments; A short-term irradiance prediction module, used to construct and train a recurrent neural network RNN to obtain a short-term irradiance prediction model; collecting the meteorological data at multiple moments within a certain day, for the k multiple moments of a certain day, according to the irradiance prediction results at a certain moment of a certain day, the real meteorological data at i a certain moment, and the characteristic information of the real cloud map data at i a certain moment, using the short-term irradiance prediction model to predict the irradiance at i a certain moment, and the real cloud map data at i a certain moment, using the short-term irradiance prediction model to predict the irradiance at i the +1 moment; where, the within a day k The irradiance prediction results at i moments include the irradiance prediction results at the i ≤ k th moment, and 1 ≤

[0013] The embodiment of the present invention provides a method and system for predicting the irradiance of a solar power plant. Compared with the prior art, the beneficial effects are as follows: According to the meteorological data collected at k moments within consecutive multiple days, use the long-term irradiance prediction model to predict the long-term irradiance of the th day, and divide it according to k moments to obtain the irradiance prediction results at moments within the k th day; collect the meteorological data at moments within the k th day. For the th moment of the i th day, according to the irradiance prediction results at the i th moment, the real meteorological data at the i th moment, and the characteristic information of the real cloud map data at the i th moment, use the short-term irradiance prediction model to predict the irradiance at the i +1th moment.

[0014] Among them, use the meteorological data collected at k moments within consecutive multiple days to predict the long-term irradiance of the th day through the long-term irradiance prediction model; collect the meteorological data at moments within the k th day. For the th moment of the i th day, since the cloud map data cannot be predicted, it is necessary to collect the characteristic information of the real cloud map data. Take the characteristic information of the real cloud map data at the i th moment and the real meteorological data at the i th moment as the input of the short-term irradiance prediction model to obtain the irradiance prediction results at the i +1th moment, ensuring that the prediction results of the short-term irradiance prediction model are more accurate with the participation of the real cloud map data, and finally realizing the combination of meteorological information and cloud map information to improve the prediction accuracy of short-term irradiance. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is the long-term irradiance prediction flowchart of a method for predicting the irradiance of a solar power plant provided by an embodiment of the present invention; Figure 2Composition diagram of the dataset of the long-term irradiance prediction model for a solar power plant irradiance prediction method provided by an embodiment of the present invention; Figure 3 Structural diagram of the LSTM prediction model for a solar power plant irradiance prediction method provided by an embodiment of the present invention; Figure 4 Structural diagram of the LSTM for a solar power plant irradiance prediction method provided by an embodiment of the present invention; Figure 5 Flow chart of the rolling prediction at time T and time T+1 based on the RNN prediction model for a solar power plant irradiance prediction method provided by an embodiment of the present invention; Figure 6 Composition diagram of the training dataset for the rolling prediction of the RNN network for a solar power plant irradiance prediction method provided by an embodiment of the present invention; Figure 7 Structural diagram of the rolling prediction model based on the RNN for a solar power plant irradiance prediction method provided by an embodiment of the present invention; Figure 8 Structural diagram of the recurrent neural network for a solar power plant irradiance prediction method provided by an embodiment of the present invention; Figure 9 Input composition of the cloud feature information at the first moment and the second moment for a solar power plant irradiance prediction method provided by an embodiment of the present invention, where (a) is the cloud feature information at the first moment and (b) is the cloud feature information at the second moment. Detailed implementation manners

[0016] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will describe the detailed implementation manners of the present invention with reference to the accompanying drawings. Many specific details are set forth in the following description to fully understand the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0017] An embodiment of the present invention provides a solar power plant irradiance prediction method, including the following steps: Step 1: Collect the irradiance data and meteorological data of the solar power plant at successive multiple days before k a certain day, and collect the cloud map data a set number of times between two successive meteorological data collection times to obtain the feature information of the cloud map data; where k≥ 2.

[0018] Step 2: Add an attention mechanism to the long short-term memory network LSTM to build an optimized LSTM model and train it to obtain a long-term irradiance prediction model; k The meteorological data at the moment is used to predict the The long-term irradiance of the day is calculated according to k time, and obtain the Within days k The irradiance prediction results at each moment.

[0019] Step 3: Construct a recurrent neural network (RNN) and train it to obtain a short-time irradiance prediction model; collect Within days k The meteorological data at the moment Day i At this moment, according to i The irradiance prediction results at the moment i The real weather data at the moment and i The characteristic information of the real cloud image data at the moment is used to predict the i +1 moment of irradiance; among them, Within days k The irradiance prediction results at each moment include i The irradiance prediction result at each moment, and 1≤ i ≤ k .

[0020] Step 4: Get the i +1 moment of real irradiance, i +1 real weather data and i +1 characteristic information of the real cloud image data at a certain moment; i +1 moment of real irradiance, i +1 real weather data and i +1 moment, the training set of the short-time irradiance prediction model is updated, and the short-time irradiance prediction model is trained by the updated training set; the updated short-time irradiance prediction model is used cyclically to predict the first i +1 time later the irradiance.

[0021] The present invention is mainly divided into two parts: 1. Long-term irradiance prediction.

[0022] A model based on Long Short-Term Memory (LSTM) is established to predict long-term irradiance. Through the multi-value input composed of various meteorological information such as irradiance, temperature, humidity, wind speed, wind direction, atmospheric pressure, and solar altitude angle, as well as time information, a reasonable long-term irradiance prediction model is constructed to predict the change trend of irradiance in the next day based on historical meteorological data. The LST model can be used for irradiance prediction in different seasons, and this prediction model can accurately predict the change trend of irradiance on sunny days.

[0023] II. Short-term irradiance prediction.

[0024] Based on the multi-value input composed of the long-term irradiance prediction result, various meteorological information, time information, and the characteristic information of clouds, a rolling prediction model based on the Recurrent Neural Network (RNN) is constructed to achieve short-term irradiance prediction, making up for the deficiency of long-term irradiance prediction.

[0025] The technical solution adopted by the present invention is to combine short-term irradiance prediction with long-term irradiance prediction.

[0026] The present invention overcomes the problems that the method for predicting solar irradiance based on meteorological values cannot accurately predict the irradiance when blocked by clouds, and the method for predicting solar irradiance based on cloud images is subject to factors such as the performance of satellite sensors and atmospheric conditions, which affect the prediction accuracy.

[0027] This irradiance prediction method is divided into two parts: long-term irradiance prediction and short-term irradiance prediction. The specific steps are as follows:

[0028] First, long-term irradiance prediction is carried out. The flow chart of long-term irradiance prediction is as Figure 1 shown.

[0029] Step 1: Data collection and data processing. Six meteorological data, namely temperature, humidity, wind speed, wind direction, atmospheric pressure, and irradiance, are collected through a meteorological instrument and an irradiance meter. At the same time, the solar altitude angle at the collection moment is calculated through an astronomical algorithm. Since the change of irradiance follows a certain pattern with the daily time, time is also regarded as one of the factors affecting irradiance. However, the time variable cannot be directly input into the network. Therefore, the sampling time is divided into levels, that is, 8:00, 8:05, 8:10,..., 17:55 are recorded as 1, 2, 3,..., 120. After obtaining the sample data, the sample data is normalized. Then the two data sets are respectively divided into a training data set and a test data set.

[0030] Use the data of multiple historical days to train the model. The input of this model has 9 variables. For example: the d -3rd, the and The irradiance, as well as time, temperature, humidity, wind speed, wind direction, atmospheric pressure, and solar altitude angle; the output is the irradiance on the th day. The d -3rd, and are consecutive days before the th day.

[0031] To reduce the data acquisition frequency and minimize acquisition errors, the originally collected data at 5-minute intervals is fused. By calculating the average of every 2 five-minute data points, they can be combined into a data point with a 10-minute resolution, resulting in 60 solar irradiance recording points in a day. The meteorological input data and irradiance data are expressed as:

[0032] (1).

[0033] (2).

[0034] In the formula: —— The th input variable, ; —— The value of the th input variable at the t-th moment; —— The solar irradiance value at the th moment on the

[0035] The composition of the dataset for the long-term irradiance prediction model is as Figure 2 shown.

[0036] Step 2: Train the LSTM model. The LSTM network structure is used for long-term irradiance prediction. The LSTM network structure is as Figure 3 shown.

[0037] The long short-term memory network (LSTM) is a variant of the recurrent neural network (RNN). It uses a gating mechanism and memory units to manage the flow of information, is suitable for processing time series data, and can effectively capture long-term dependencies in time series data. In irradiance prediction, historical meteorological data and time information are key features, and the LSTM is good at processing this kind of time series data. The LSTM can effectively solve the problem of gradient vanishing or gradient explosion that traditional RNNs encounter when processing long sequence data through the gating mechanism, and can better handle long-term dependencies, making it suitable for the long-term influencing factors involved in irradiance prediction.

[0038] AsFigure 4 As shown, in the LSTM structure, three "gates" are introduced to control and relate the information flow, namely the forget gate, the input gate, and the output gate.

[0039] The forget gate is used to control which information in the memory state at the previous moment should be forgotten. The inputs include the hidden state at the previous moment and the input data at the current moment , and these two parts of data simultaneously pass through the sigmoid activation function to obtain the output of the forget gate at the current moment . Since the output range of the sigmoid activation function is [0, 1], each element in the output is a value between 0 and 1, indicating the degree of information retention in the memory state at the corresponding position. Multiplying with the memory state at the previous moment indicates which parts of the memory state at each position need to be forgotten. The output of the forget gate has the mathematical expression as shown in formula (3).

[0040] (3).

[0041] In the formula: , ——The parameter matrix of the forget gate; ——The bias.

[0042] The input gate is used to control whether the input at the corresponding position is memorized. The inputs include and two parts. The input gate includes two parts of data. The first part of data is to pass the two parts of inputs through the sigmoid activation function simultaneously to obtain the output , and the second part of data is to pass the two parts of inputs through the hyperbolic tangent (tanh) activation function simultaneously to obtain the output , and then multiply these two parts of data to update the cell state. The mathematical expressions are shown in formulas (4) and (5).

[0043] (4).

[0044] (5).

[0045] From the above description, it can be seen that the outputs of the forget gate and the input gate will both affect the update of the cell state. The cell state is updated by two parts. The first part is the product of the cell state from the previous moment and the output of the forget gate at the current moment; the second part is the output of the input gate at the current moment and Perform the multiplication, and then add the result of the multiplication to the result of the first part to obtain the cell state at the current moment. , and its mathematical expression is shown in Equation (6).

[0046] (6).

[0047] The output gate mainly determines the output at the current moment through the information at the current moment and the information passed from the previous moment. The output includes two parts. One part is the hidden state at the previous moment and the input data at the current moment which pass through the Sigmoid activation function to obtain the output at the current moment ; the second part is the cell state at the current moment which is multiplied by the output at the current moment through the tanh activation function to obtain the hidden layer output at the current moment . Its mathematical expressions are shown in Equations (7) and (8).

[0048] (7).

[0049] (8).

[0050] The training of LSTM uses Backpropagation Through Time (BPTT). The BPTT algorithm is similar to the Backpropagation (BP) algorithm, that is, first calculate the error term of each LSTM neuron in the forward process, then calculate the gradient of each weight according to the error term, and update the weights using the gradient optimization algorithm.

[0051] Although LSTM can capture long-term dependencies in time series data, as the number of time steps increases, information may be lost or diluted during transmission in the network. By introducing the attention mechanism, the modeling ability of the model for long-distance dependencies can be further enhanced, helping the model to better understand and utilize different parts of information in the sequence, focusing on the learning of key information without being interfered by other irrelevant information, thereby improving the generalization ability of the model. At the same time, by introducing the attention mechanism, the flow of information can be more finely controlled, preventing information loss and improving the performance of the model.

[0052] The use of the LSTM mechanism can make the prediction model more flexible and accurate in processing time series data, improving the prediction performance and generalization ability of the irradiance prediction model. LSTM captures the correlations and long-term dependencies existing between front and back data points through training, so as to better predict future values and trends.

[0053] Step 3: After the model training is completed, use the data of sunny and cloudy days of a domestic power station as the test set for long-term irradiance prediction. In order to reduce the data acquisition frequency to reduce the acquisition error, the original data is fused.

[0054] The results show that the long-term irradiance prediction model can accurately predict the irradiance on sunny days. For the special case of sudden changes in irradiance caused by clouds, since this model predicts the irradiance of the next day through historical meteorological information and does not consider the occlusion of the sun by clouds, it is unable to predict the sudden changes in irradiance caused by cloud factors. Therefore, the prediction results of the long-term irradiance prediction model LSTM can be used as a reference value for short-term irradiance prediction.

[0055] Step 4: After the long-term irradiance prediction is completed, short-term irradiance prediction is carried out. The short-term irradiance prediction is carried out using a rolling prediction model based on RNN. The flow chart is as Figure 5 shown. First, the data is preprocessed, and the meteorological data and cloud map data of the experimental site are selected. The meteorological data is collected at intervals of 5 minutes, and the cloud map data is collected at intervals of 20 seconds. The sampling times of the meteorological data are divided into levels, that is, 8:00, 8:05, 8:10,..., 17:55 are recorded as . The sampling times of the cloud map data are recorded as . Then and The corresponding relationship is that when increases to , increases 15 times to .

[0056] For the N clouds identified in the cloud map, taking as an example, they are successively represented by area size as . Let the average brightness, pixel area, motion vector, relative distance, and cloud height of a cloud be represented by variables respectively. Then the nodes of the graph structure are represented by the set , and the relationship between the nodes, that is, the information of the edges, is represented by the adjacency matrix , as shown in formula (14).

[0057] (9).

[0058] This rolling prediction model jointly predicts the irradiance of the next moment based on the characteristic information, meteorological information, and time data of 4 clouds. The composition of the training data set for the rolling prediction of the RNN network is as Figure 6 shown.

[0059] Among them The output data at the moment is represented as The input data is represented as shown in Equation (15).

[0060] (10).

[0061] Wherein, represents the input at time represents the th data at time represents the result of the long-term irradiance prediction at the corresponding time.

[0062] Step 5: Predict the irradiance information at the next time according to the historical cloud feature information and in combination with the long-term irradiance, meteorological data, and time data at the current time. Finally, add the actual input data at the next time to the training samples and update the RNN network training model online until the prediction ends. The structure diagram of the rolling prediction model based on RNN is as Figure 7 shown.

[0063] The input of the recurrent neural network (RNN) model is the integration result of the long-term irradiance prediction result, meteorological data, time data, and the feature information of the clouds at the 15th time at the current time. The output of the RNN model is the irradiance data at the next time. Compared with the traditional feedforward neural network, RNN has a memory function and introduces the concept of time series, that is, the state at the previous time will affect the state at the next time, and it can better process data with time series. Its network structure is as Figure 8 shown.

[0064] Although RNN has the problem of long-term dependence and it is difficult for the network to effectively capture information in the relatively distant future. However, for short-term irradiance prediction, more attention is paid to the information in the previous few times. Therefore, this section is based on RNN to perform short-term irradiance prediction. RNN also includes an input layer, a hidden layer, and an output layer. Among them, the nodes between the hidden layers are connected. During the forward propagation of the network, the output of the hidden layer at the current time is determined by both the input at the current time and the output of the hidden layer at the previous time, that is:

[0065] (11).

[0066] (12).

[0067] (13).

[0068] In the formula: , —— the hidden layer states at time and time ——activation function; , , —— The parameter matrices between the input layer and the hidden layer, from the hidden layer at the previous moment to the hidden layer at the current moment, and from the hidden layer to the output layer, respectively; —— The input at time , —— The bias; , —— is The output of the hidden layer at time

[0069] Step 6: After the model training is completed, predict the irradiance of the next day. Since all data is collected starting from time 1, and real data cannot be collected, and the input of the meteorological information at this time is 0, the predicted value of the irradiance at time 1 is equal to the result of the long-term irradiance prediction. After the data at time 1 is collected, the model starts rolling prediction. At this time, the input composition of the cloud feature information predicted at time 2 is Figure 9 as shown in (a), and there is only one set of cloud feature information at this time. After the data at time 2 is collected, the model rolls and updates the collected data. At this time, 16 sets of cloud feature information have been collected. At this time, the input composition of the cloud feature information predicted at time 3 is Figure 9 as shown in (b). Perform rolling prediction in this way until the prediction ends.

[0070] An embodiment of the present invention provides a prediction system for the irradiance of a solar power plant, including: A data acquisition module, configured to acquire the irradiance data and meteorological data of the solar power plant at consecutive multiple days before the k th day, acquire the cloud map data a set number of times between two consecutive meteorological data acquisition times, and obtain the feature information of the cloud map data; wherein, k≥ 2.

[0071] A long-term irradiance prediction module, configured to add an attention mechanism to the long short-term memory network LSTM to construct an optimized LSTM model and perform training to obtain a long-term irradiance prediction model; according to the meteorological data at k consecutive multiple times collected, use the long-term irradiance prediction model to predict the long-term irradiance of the th day, and divide it according to k times to obtain the irradiance prediction results at consecutive multiple times within the k th day.

[0072] A short-term irradiance prediction module, configured to construct a recurrent neural network RNN and perform training to obtain a short-term irradiance prediction model; acquire the within a day k meteorological data at each moment, for the day's i th moment, according to the irradiance prediction result at the i th moment, the real meteorological data at the i th moment and the characteristic information of the real cloud map data at the i th moment, use the short - term irradiance prediction model to predict the irradiance at the i +1 th moment; wherein, the irradiance prediction results within the day k include the irradiance prediction results at the i th moment, and 1≤ i ≤ k .

[0073] A specific embodiment is as follows: This embodiment discloses a method for predicting the irradiance of a solar power plant, and the specific steps are as follows: S1. Long - term irradiance prediction: The first stage is data collection and data processing. Six kinds of meteorological data, namely temperature, humidity, wind speed, wind direction, atmospheric pressure and irradiance, are collected through a meteorological instrument and an irradiance meter. At the same time, the solar altitude angle at the collection moment is calculated through an astronomical algorithm. Since the change of irradiance follows a certain pattern with the daily time, time is also regarded as one of the factors affecting irradiance. However, the time variable cannot be directly input into the network, so the sampling time is divided into levels, that is, 8:00, 8:05, 8:10,..., 17:55 are recorded as 1, 2, 3,..., 120. After obtaining the sample data, the sample data is normalized. Then the two data sets are respectively divided into a training data set and a test data set.

[0074] The second stage is to establish an LSTM prediction model and train the model. LSTM captures the correlation and long - term dependence relationship existing between the front and back data points through training, so as to better predict future values and trends. After establishing the model, the model is trained using two different data sets respectively.

[0075] After training the model with the data set, the model is verified using the validation set. The results prove that the long - term irradiance prediction model LSTM can accurately predict the irradiance value and change trend on sunny days. However, in cloudy weather, due to the lack of consideration of the characteristic information of clouds, this prediction model cannot accurately predict the sudden change of irradiance before and after the sun is blocked by clouds.

[0076] S2. Short - term irradiance prediction: The first stage is data collection and data processing. The processing methods for meteorological data and time data are the same as those for long-term irradiance prediction. The cloud image data is collected by a tracking cloud image measurement system. The sampling time of the cloud image data is once every 20 s. The non-cloud image data is collected once every 5 min. Therefore, it is necessary to align the data, that is, one set of non-cloud image data corresponds to 15 sets of cloud image data. Then the data set is divided into a training data set and a test data set that are continuous on the time scale. The second stage is the establishment and training of a rolling prediction model based on RNN. The input of the RNN model is the long-term irradiance prediction result at the current moment, meteorological data, time data, and the characteristic information of the cloud layer at the 15th moment. The output of the RNN model is the irradiance data at the next moment. The RNN model is used to predict the irradiance information at the next moment, and the irradiance information at the next moment is predicted by combining the long-term irradiance, meteorological data, and time data at the current moment. Finally, the actual input data at the next moment is added to the training sample, and the RNN network training model is updated online until the prediction ends.

[0077] After training the model with the data set, the validation set is used to validate the model. The rolling prediction model based on RNN corrects the long-term irradiance prediction result according to the characteristic information of the cloud layer and the meteorological information at the previous moment, improving the accuracy of the irradiance prediction. When the irradiance changes suddenly, its change trend can be accurately predicted.

[0078] The above-described embodiments merely represent several implementation manners of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the invention patent shall be subject to the appended claims.

Claims

1. A method for predicting irradiance of a solar power plant, characterized in that: The following steps are involved: Solar power plant In the past few consecutive days k The irradiance data and meteorological data at each moment are collected, and the cloud image data of a set number of times is collected between two consecutive meteorological data collection moments to obtain the characteristic information of the cloud image data; wherein, k≥ 2; Add an attention mechanism to the long short-term memory network LSTM to build an optimized LSTM model and train it to obtain a long-term irradiance prediction model; k The meteorological data at the moment is used to predict the The long-term irradiance of the day is calculated according to k time, and obtain the Within days k Irradiance prediction results at each moment; Construct a recurrent neural network (RNN) and train it to obtain a short-time irradiance prediction model; collect Within days k The meteorological data at the moment Day i At this moment, according to i The irradiance prediction results at the moment i The real weather data at the moment and i The characteristic information of the real cloud image data at the moment is used to predict the i +1 moment of irradiance; wherein, the Within days k The irradiance prediction results at each moment include i The irradiance prediction result at each moment, and 1≤ i≤k .

2. A method for predicting irradiance of a solar power plant according to claim 1, characterized in that: The prediction i After +1 moment of irradiance, it also includes: Get the i +1 moment of real irradiance, i +1 real weather data and i +Feature information of real cloud image data at 1 moment; According to i +1 moment of real irradiance, i +1 real weather data and i + The characteristic information of the real cloud image data at 1 moment is used to update the training set of the short-time irradiance prediction model, and the short-time irradiance prediction model is trained by the updated training set; Cyclic use of the updated short-time irradiance prediction model to predict the i +1 time later irradiance.

3. A method for predicting irradiance of a solar power plant according to claim 1, characterized in that: The specific steps of obtaining the long-term irradiance prediction model include: According to In the past few consecutive days k The meteorological data at the moment and the irradiance data at the corresponding moment are used to train the optimized LSTM model using the time-based back propagation algorithm BPTT to predict the The long-term irradiance of the day is used to obtain the long-term irradiance prediction model.

4. A method for predicting irradiance of a solar power plant according to claim 1, characterized in that: The prediction The long-term irradiance of the day, the specific steps include: Input time, temperature, humidity, wind speed, wind direction, atmospheric pressure, irradiance and solar altitude angle into the long-term irradiance prediction model; The long-term dependencies in time series data are captured by LSTM, and the information affecting irradiance is focused on through the attention mechanism to predict the Long-term irradiance of the day.

5. A method for predicting irradiance of a solar power plant according to claim 1, characterized in that: The specific steps of obtaining the characteristic information of the cloud image data include: Identify N cloud layers in cloud image data; The average brightness, pixel area, motion vector, relative distance and cloud height of each cloud layer are taken as feature information.

6. A method for predicting irradiance of a solar power plant according to claim 1, characterized in that: The characteristic information of the real cloud image data at the moment specifically comprises the following steps: From the real cloud image data collected for a set number of times, select the last real cloud image data; Get the feature information of the last real cloud image data.

7. A solar power plant irradiance prediction system, characterized in that: include: The data acquisition module is used to collect the data of the solar power plant in the In the past few consecutive days k The irradiance data and meteorological data at each moment are collected, and the cloud image data of a set number of times is collected between two consecutive meteorological data collection moments to obtain the characteristic information of the cloud image data; wherein, k≥ 2; The long-term irradiance prediction module is used to add an attention mechanism to the long short-term memory network LSTM to build an optimized LSTM model and train it to obtain a long-term irradiance prediction model; based on the collected data over multiple consecutive days, k The meteorological data at the moment is used to predict the The long-term irradiance of the day is calculated according to k time, and obtain the Within days k Irradiance prediction results at each moment; The short-time irradiance prediction module is used to construct and train the recurrent neural network RNN ​​to obtain the short-time irradiance prediction model; collect the first Within days k The meteorological data at the moment Day i At this moment, according to i The irradiance prediction results at the moment i The real weather data at the moment and i The characteristic information of the real cloud image data at the moment is used to predict the i +1 moment of irradiance; wherein, the Within days k The irradiance prediction results at each moment include i The irradiance prediction result at each moment, and 1≤ i≤k .

Images