Photovoltaic power prediction method and device with dynamic adjustment and feedback correction of fusion model

Abstract

The application provides a kind of fusion model dynamic adjustment and feedback correction photovoltaic power prediction method and device, it is related to photovoltaic power prediction technical field.The method comprises: obtaining first time sequence, input first photovoltaic power prediction model based on back propagation neural network, output first predicted photovoltaic power sequence, first time sequence includes the irradiance and temperature of 24 hours of day to be predicted;Second time sequence is obtained, and second photovoltaic power prediction model based on long short-term memory network is input to second time sequence, and second predicted photovoltaic power sequence is output, and second time sequence includes the historical photovoltaic power of 24 hours of day before N of day to be predicted;Determine the first category label of first predicted photovoltaic power sequence, determine the second category label of second predicted photovoltaic power sequence;Based on first category label and second category label, determine the target predicted photovoltaic power sequence of day to be predicted.The application can improve prediction accuracy.

Description

Technical Field

[0001] This application relates to the field of photovoltaic power prediction technology, and in particular to a photovoltaic power prediction method and apparatus that integrates dynamic adjustment and feedback correction of a fusion model. Background Technology

[0002] As a crucial component of clean energy, photovoltaic (PV) power generation exhibits significant intermittency, randomness, and volatility in its output. With the rapid growth of PV installed capacity, the high proportion of unstable PV power connected to the grid poses a severe challenge to the real-time balance of the power system, dispatch planning, and safe and stable operation. Against this backdrop, achieving high-precision short-term and ultra-short-term PV power forecasting has become a key technological support for ensuring grid absorption capacity, optimizing dispatch decisions, and supporting electricity market transactions, possessing significant engineering value and economic significance.

[0003] Currently, based on the characteristics and types of input and output data, the mainstream methods for predicting photovoltaic power using artificial intelligence can be mainly divided into two categories: one is to establish a complex nonlinear relationship between meteorological input characteristics (such as irradiance, temperature, humidity, etc.) and photovoltaic output power; the other is to fully explore the temporal evolution law of photovoltaic power output using historical power data. However, each category has its advantages and disadvantages, and neither can achieve completely accurate prediction.

[0004] Back Propagation Neural Network (BPNN) is a classic feedforward neural network model. Its core lies in iteratively correcting network weights through an error backpropagation algorithm. It possesses strong nonlinear mapping capabilities and good self-learning ability, and is widely used in the first type of problem. BPNN can effectively establish complex nonlinear relationships between meteorological input features (such as irradiance, temperature, and humidity) and photovoltaic output power. Its network structure is relatively simple, training complexity is low, and it is easy to implement in engineering. However, BPNN has inherent limitations in handling time-dependent problems. It assumes that the input and output are independent, making it difficult to effectively capture the temporal correlation features of photovoltaic power sequences. Furthermore, BPNN is sensitive to initial weights, easily getting trapped in local extrema, and its prediction accuracy significantly decreases when the input data contains noise or outliers. In photovoltaic power prediction scenarios, BPNN relies excessively on the accuracy of weather forecast data. If there are deviations in meteorological features such as irradiance and temperature on the prediction day, the output error of the BPNN will be further amplified.

[0005] Long Short-Term Memory (LSTM) networks are an improved Recurrent Neural Network (RNN) that effectively solves the gradient vanishing and gradient explosion problems inherent in traditional RNNs during long-sequence training by introducing three gating mechanisms: input gate, forget gate, and output gate. LSTM can selectively remember or forget historical information, possessing powerful time series modeling capabilities, and is therefore widely used in the second type of problem. In photovoltaic power prediction, LSTM can fully utilize historical power data to explore the temporal evolution patterns of photovoltaic output, exhibiting strong adaptability to the randomness and volatility of power. However, LSTM models also have significant limitations. First, LSTM is essentially a single-input, single-output time series prediction model, making it difficult to directly incorporate external meteorological features such as irradiance and temperature, which are precisely the key factors affecting photovoltaic output. Second, the predictive performance of LSTM is highly dependent on the completeness and continuity of historical data; when there are missing data or measurement anomalies in the historical power sequence, the prediction results will be significantly biased. Furthermore, LSTM can only learn the statistical patterns implicit in historical power data. When new situations that have not appeared in historical patterns occur on the day to be predicted (such as sunny turning cloudy or sudden cloud cover), LSTM cannot detect the weather changes in advance, and the prediction accuracy will drop sharply. Summary of the Invention

[0006] This application provides a photovoltaic power prediction method and apparatus that integrates dynamic model adjustment and feedback correction to solve the problems in existing photovoltaic power prediction technologies, such as the inherent defects of a single model, the large prediction deviation caused by noise and missing input data, the inability of a statically fixed model to adapt to the long-term time-varying characteristics of photovoltaic output, and the difficulty in making decisions when the prediction results of two models conflict.

[0007] Firstly, this application provides a photovoltaic power prediction method that integrates dynamic model adjustment and feedback correction, including: A first time series is obtained and input into a first photovoltaic power prediction model constructed based on a backpropagation neural network, and a first predicted photovoltaic power series is output. The first time series includes irradiance and temperature within 24 hours of the day to be predicted. A second time series is obtained and input into a second photovoltaic power prediction model constructed based on a long short-term memory network. The second predicted photovoltaic power series is output, and the second time series includes the historical photovoltaic power within 24 hours of N days before the date to be predicted. Determine a first category label for the first predicted photovoltaic power sequence, and determine a second category label for the second predicted photovoltaic power sequence; Based on the first category label and the second category label, the target predicted photovoltaic power sequence for the day to be predicted is determined.

[0008] Secondly, this application provides a photovoltaic power prediction device that integrates dynamic model adjustment and feedback correction, comprising: The first output module is used to acquire a first time series and input the first time series into a first photovoltaic power prediction model constructed based on a backpropagation neural network, and output a first predicted photovoltaic power series. The first time series includes irradiance and temperature within 24 hours of the day to be predicted. The second output module is used to acquire the second time series and input the second time series into the second photovoltaic power prediction model constructed based on the long short-term memory network, and output the second predicted photovoltaic power series. The second time series includes the historical photovoltaic power in the 24 hours of N days before the date to be predicted. A label determination module is used to determine a first category label for the first predicted photovoltaic power sequence and a second category label for the second predicted photovoltaic power sequence; A power determination module is used to determine the target predicted photovoltaic power sequence for the day to be predicted based on the first category label and the second category label.

[0009] This application provides a photovoltaic power prediction method and apparatus that integrates dynamic adjustment and feedback correction of the model. The method involves: acquiring a first time series and inputting it into a first photovoltaic power prediction model constructed based on a backpropagation neural network to output a first predicted photovoltaic power sequence. The first time series includes irradiance and temperature within 24 hours of the day to be predicted. Then, acquiring a second time series and inputting it into a second photovoltaic power prediction model constructed based on a long short-term memory network to output a second predicted photovoltaic power sequence. The second time series includes historical photovoltaic power within 24 hours of N days prior to the day to be predicted. Finally, the method determines a first category label for the first predicted photovoltaic power sequence and a second category label for the second predicted photovoltaic power sequence. Based on the first and second category labels, the method determines the target predicted photovoltaic power sequence for the day to be predicted. This application improves prediction accuracy by integrating the advantages of two different models and comprehensively utilizing various types of data features. Furthermore, by using meteorological and historical power data, it provides rich information for prediction from different perspectives. Compared to prediction methods using only a single data source, it can more comprehensively consider various factors affecting photovoltaic power, thereby improving prediction accuracy. Simultaneously, by determining the first category label of the first predicted photovoltaic power sequence and the second category label of the second predicted photovoltaic power sequence, it provides additional classification information for the prediction results. These category labels can reflect the characteristics and reliability of the prediction results from different models. Based on these category labels, the fusion strategy can be adjusted more flexibly to further optimize the target predicted photovoltaic power sequence, while providing more targeted support for photovoltaic system operation decisions. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart illustrating the implementation of the photovoltaic power prediction method based on dynamic adjustment and feedback correction of the fusion model provided in this application embodiment; Figure 2 This is a schematic diagram of photovoltaic prediction results with noise in the data provided in the embodiments of this application; Figure 3 This is a schematic diagram of photovoltaic prediction results under the condition that the data provided in the embodiments of this application is missing; Figure 4 This is a schematic diagram of the photovoltaic power prediction device with dynamic adjustment and feedback correction of the fusion model provided in the embodiments of this application. Detailed Implementation

[0012] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0013] To make the objectives, technical solutions, and advantages of this application clearer, the following description will be provided in conjunction with the accompanying drawings and specific embodiments.

[0014] Figure 1 The implementation flowchart of the photovoltaic power prediction method with dynamic adjustment and feedback correction of the fusion model provided in the embodiments of this application is described in detail below: In step 101, a first time series is obtained and input into a first photovoltaic power prediction model constructed based on a backpropagation neural network, and a first predicted photovoltaic power series is output. The first time series includes irradiance and temperature within 24 hours of the day to be predicted.

[0015] The first predicted photovoltaic power sequence is composed of the photovoltaic power prediction results at each hour within a 24-hour period.

[0016] In this embodiment of the application, a first time series of irradiance and temperature at each time point within 24 hours of the day to be predicted is obtained for a photovoltaic power station, photovoltaic array, or distributed photovoltaic power station cluster. ,in, For the first Irradiance at a given time point For the first Temperature at each point in time, The first time series is then input into the first photovoltaic power prediction model built on a back propagation neural network (BPNN), which outputs the first predicted photovoltaic power series. ,in, To be predicted In the The first predicted photovoltaic power at each time point The date is yet to be predicted.

[0017] Among them, the backpropagation neural network has a strong ability to handle nonlinear mapping relationships and can capture the complex nonlinear relationship between irradiance and temperature and photovoltaic power.

[0018] In one possible implementation, the method may further include, before obtaining the first time series: Obtain historical photovoltaic power data, which includes historical photovoltaic power curves for multiple days; Historical photovoltaic power data is clustered to obtain a power category labeling system, which includes labels for sunny days, cloudy days, and rainy days.

[0019] Optionally, historical photovoltaic power data including historical photovoltaic power curves for multiple days can be obtained, and then a clustering algorithm can be used to divide the daily historical photovoltaic power curves into typical category labels such as sunny day label, cloudy day label, and rainy day label. The sunny day label, cloudy day label, and rainy day label together constitute a power category label system.

[0020] For example, historical photovoltaic power data containing the historical photovoltaic power curve of the previous year is obtained, and then K-means is used to divide the historical photovoltaic power curve of the previous year into three typical category labels: sunny day label, cloudy day label, and rainy day label.

[0021] In step 102, a second time series is obtained and input into a second photovoltaic power prediction model constructed based on a long short-term memory network, and a second predicted photovoltaic power series is output. The second time series includes the historical photovoltaic power within 24 hours of N days before the date to be predicted.

[0022] The second predicted photovoltaic power sequence is composed of the photovoltaic power prediction results at each hour within a 24-hour period.

[0023] In this embodiment of the application, a second time series of historical photovoltaic power within 24 hours of N days prior to the prediction date is also obtained for the photovoltaic power plant, photovoltaic array, or distributed photovoltaic power plant cluster. When N is 3, the result is: ,in, Forecast date Heavenly The historical photovoltaic power at each time point is calculated. Then, the second time series is input into a second photovoltaic power prediction model built on a Long Short-Term Memory (LSTM) network, outputting a second predicted photovoltaic power series. ,in, To be predicted In the The second predicted photovoltaic power at a given time point.

[0024] Among them, Long Short-Term Memory (LSTM) networks excel at processing sequential data, and can effectively uncover long-term and short-term patterns in time-dependent data such as historical photovoltaic power sequences.

[0025] In steps 101 and 102 of this application embodiment, the advantages of two different models are integrated and multiple types of data features are comprehensively utilized, thereby improving the accuracy of prediction.

[0026] In step 103, a first category label for the first predicted photovoltaic power sequence is determined, and a second category label for the second predicted photovoltaic power sequence is determined.

[0027] In this embodiment, the first predicted photovoltaic power sequence is matched with the typical power curves of each category in the pre-constructed power category labeling system based on morphological similarity, and the category label with the highest matching degree is selected as the first category label of the first predicted photovoltaic power sequence. Similarly, the second predicted photovoltaic power sequence is matched with the typical power curves of each category in the pre-constructed power category labeling system based on morphological similarity, and the category label with the highest matching degree is selected as the second category label of the second predicted photovoltaic power sequence.

[0028] This application provides additional classification information for the prediction results by determining a first category label for the first predicted photovoltaic power sequence and a second category label for the second predicted photovoltaic power sequence. Based on these category labels, the fusion strategy can be adjusted more flexibly, further optimizing the target predicted photovoltaic power sequence, while providing more targeted support for the operation decisions of the photovoltaic system.

[0029] In one possible implementation, determining a first category label for a first predicted photovoltaic power sequence and determining a second category label for a second predicted photovoltaic power sequence may include: The first predicted photovoltaic power sequence is clustered, and the clustering results are mapped to a power category label system to obtain the first category label; The second predicted photovoltaic power sequence is clustered, and the clustering results are mapped to the power category label system to obtain the second category label.

[0030] Optionally, the temporal amplitude, fluctuation characteristics, peak time, and daily power output pattern characteristics of the first predicted photovoltaic power sequence are extracted, clustered, and the clustering results are mapped to a pre-constructed power category labeling system according to a preset matching result. The sunny day label, cloudy day label, or rainy day label with the highest matching degree with the first predicted photovoltaic power sequence is obtained as the first category label. Simultaneously, the temporal amplitude, fluctuation characteristics, peak time, and daily power output pattern characteristics of the second predicted photovoltaic power sequence are extracted, clustered, and the clustering results are mapped to a pre-constructed power category labeling system according to a preset matching result. The sunny day label, cloudy day label, or rainy day label with the highest matching degree with the second predicted photovoltaic power sequence is obtained as the second category label.

[0031] This application's embodiments rapidly assign category labels to two prediction sequences through clustering mapping. This allows for a direct determination of whether the prediction results of the first and second photovoltaic power prediction models belong to the same weather type, providing a clear trigger condition for subsequent bidirectional feedback correction. Furthermore, this application's embodiments are based on power sequence morphology clustering rather than a single threshold judgment, ensuring stable classification even in scenarios with large power output fluctuations such as cloudy or abrupt weather changes, improving label accuracy under complex operating conditions. Simultaneously, the clustering mapping logic is simple and computationally inexpensive, enabling rapid online category determination without increasing the system's real-time prediction latency, making it suitable for engineering deployment needs of ultra-short-term / short-term photovoltaic power prediction.

[0032] In step 104, the target predicted photovoltaic power sequence for the day to be predicted is determined based on the first category label and the second category label.

[0033] In this embodiment of the application, based on the comparison result of whether the labels of the first category label and the second category label determined in step 103 are consistent, the corresponding fusion strategy or feedback correction process is executed to finally determine the target predicted photovoltaic power sequence for the date to be predicted.

[0034] Accurate photovoltaic (PV) power forecasting is crucial for grid dispatching and operation management. The high-precision target-predicted PV power sequence provided in this application can help grid dispatching departments understand the expected output of PV power generation in advance, rationally arrange the dispatching of other power generation resources, optimize grid operation, and improve grid stability and reliability. Simultaneously, accurate forecasting results can also reduce grid reserve capacity requirements, reduce energy waste, and improve energy utilization efficiency.

[0035] In one possible implementation, determining the target predicted photovoltaic power sequence for the date to be predicted, based on a first category label and a second category label, may include: Determine if the first category label and the second category label are consistent; If the first category label and the second category label are the same, then the first predicted photovoltaic power sequence and the second predicted photovoltaic power sequence are weighted and fused to obtain the target predicted photovoltaic power sequence; If the first category label and the second category label are inconsistent, a bidirectional feedback correction mechanism is triggered to correct the first time series and the second time series. After the bidirectional feedback mechanism reaches the preset convergence condition, a weighted fusion calculation is performed on the first predicted photovoltaic power series corresponding to the corrected first time series and the second predicted photovoltaic power series corresponding to the corrected second time series to obtain the target predicted photovoltaic power series.

[0036] Optionally, for the first category label of the first predicted photovoltaic power sequence and the second category label of the second predicted photovoltaic power sequence determined in step 103, it is determined whether the first category label and the second category label belong to the same category in the power category labeling system.

[0037] If the first category label and the second category label are consistent, it indicates that the prediction trends of the first photovoltaic power prediction model and the second photovoltaic power prediction model are consistent. Then, using the preset weight coefficient, the first predicted photovoltaic power sequence and the second predicted photovoltaic power sequence are weighted and fused, and the weighted fusion result is used as the target predicted photovoltaic power sequence.

[0038] The specific weighted fusion calculation formula is as follows:

[0039] in, To be predicted No. Target predicted photovoltaic power at each time point These are the preset weighting coefficients for the first predicted photovoltaic power sequence. The preset weighting coefficients for the second predicted photovoltaic power sequence are, in this embodiment of the application, determined by comparing the accuracy of the prediction results with that of historical data.

[0040] If the first category label and the second category label are inconsistent, it indicates that there is a conflict between the prediction results of the first photovoltaic power prediction model and the second photovoltaic power prediction model. In this case, the bidirectional feedback correction mechanism is automatically triggered to correct the first time series and the second time series. After the bidirectional feedback correction mechanism reaches the preset convergence condition, the first predicted photovoltaic power series corresponding to the corrected first time series and the second predicted photovoltaic power series corresponding to the corrected second time series are weighted and fused to obtain the target predicted photovoltaic power series.

[0041] The preset convergence condition is that the first predicted photovoltaic power sequence corresponding to the corrected first time series and the second predicted photovoltaic power sequence corresponding to the corrected second time series converge, that is, the category labels of the two are consistent, or the number of iterations of the first photovoltaic power prediction model and the second photovoltaic power prediction model reaches the preset maximum number of iterations.

[0042] In one possible implementation, triggering the bidirectional feedback correction mechanism may include: The second time series is corrected using the first predicted photovoltaic power series. The first time series is corrected using the second predicted photovoltaic power series. Using the corrected first time series and the second time series, the first photovoltaic power prediction model and the second photovoltaic power prediction model are iteratively solved until the first predicted photovoltaic power sequence corresponding to the corrected first time series and the second predicted photovoltaic power sequence corresponding to the corrected second time series converge or the preset maximum number of iterations is reached.

[0043] Optionally, using the first predicted photovoltaic power series as a matching benchmark, the second time series of input data for the second photovoltaic power prediction model is corrected, specifically as follows: Using the power curve of the first predicted photovoltaic power sequence as the first matching template, select the first historical day corresponding to the power curve that matches the first matching template, and extract the historical actual power sequence within 24 hours of N days before the first historical day.

[0044] By using the historical actual power at each time point and the historical photovoltaic power in the second time series, the corrected second time series is determined. That is, for each time point N days before the first historical date, the historical actual power at that time point is averaged with the historical photovoltaic power at that time point in the second time series to obtain the historical photovoltaic power corresponding to that time point in the corrected second time series.

[0045] For example, assuming N is 3, the full-day time-series power curve of the first photovoltaic power prediction model is used as the first matching template. From the historical database, the Kth historical day corresponding to the power curve with the highest similarity is selected as the first historical day through morphological similarity matching. The historical actual power sequence of the three days prior to the first historical day is then extracted. It is used to correct the second time series of input data for the second photovoltaic power prediction model.

[0046] For each time point within 3 days prior to the first historical date, the corrected historical photovoltaic power corresponding to that time point is calculated using the following formula:

[0047] in, First Historical Day forward The first day Historical photovoltaic power at each point in time. First Historical Day forward The first day Historical actual power at each point in time, To be predicted forward The first day Historical photovoltaic power at a specific point in time.

[0048] Based on the above formula, the second time series after the correction of the second photovoltaic power prediction model can be obtained, namely: .

[0049] Then, using the second predicted photovoltaic power sequence as a matching benchmark, the first time series of input data for the first photovoltaic power prediction model is corrected, specifically as follows: Using the power curve of the second predicted photovoltaic power sequence as the second matching template, the second historical day corresponding to the power curve that matches the second matching template is selected, and the historical actual irradiance and historical actual temperature within 24 hours of the second historical day are extracted. The second historical day is the day before the N consecutive days before the existence of a power curve that matches the second matching template.

[0050] By using the historical actual irradiance and historical actual temperature at each time point and the irradiance and temperature in the first time series, the corrected first time series is determined. That is, for each time point of the second historical day, the average value of the historical actual irradiance at that time point and the irradiance at that time point in the first time series is calculated to obtain the average irradiance at that time point, and the average value of the historical actual temperature at that time point and the temperature at that time point in the first time series is calculated to obtain the average temperature at that time point. The average irradiance and average temperature at that time point are used as the corresponding irradiance and temperature in the corrected first time series at that time point.

[0051] For example, assuming N is 3, the full-day time-series power curve of the second predicted photovoltaic power sequence is used as the second matching template. Then, the three consecutive days most similar in shape to the second matching template are retrieved from the historical database, namely K-3, K-2, and K-1 historical days. Will The day preceding each of the three consecutive days is designated as the second historical day (i.e., the Kth historical day). Then, the historical actual irradiance and historical actual temperature sequences corresponding to the second historical day are extracted for the 24 hours. This is used to correct the input data of the first photovoltaic power prediction model.

[0052] For each point in time on the second historical day, the following formula is used to calculate... Should The corrected irradiance and temperature sequence corresponding to time, i.e.:

[0053] in, For the first Corrected irradiance and temperature at each time point The Second Historical Day In the Historical actual irradiance and historical actual temperature series at each time point For the first Irradiance and temperature sequences at various time points.

[0054] Then, using the corrected first and second time series, the first photovoltaic power prediction model and the second photovoltaic power prediction model are rerun for iterative updates until the output results of the first and second photovoltaic power prediction models converge, that is, the category labels of the first predicted photovoltaic power sequence corresponding to the corrected first time series and the second predicted photovoltaic power sequence corresponding to the corrected second time series are consistent, or the preset maximum number of iterations is reached.

[0055] In one possible implementation, after determining the target predicted photovoltaic power sequence for the date to be predicted based on the first category label and the second category label, the method may further include: The network parameters of the first and second photovoltaic power prediction models are optimized using a resource-allocating network (RAN).

[0056] Optionally, embodiments of this application employ a RAN to adapt to data distribution shifts caused by environmental changes, dynamically adjusting the structures of the BPNN and LSTM to adapt to the data distribution in real time. If the outputs of the above two networks... The deviation from the actual photovoltaic power is greater than a preset critical value, and the average distance between the new sample and the historical sample is greater than the critical distance threshold. When the sample is in use, it is stored in the RAN buffer. sample number quantity When the preset limit is reached, add a new node to the network. The current buffer data is normalized to train new hidden layer nodes, thereby adapting to the dynamic changes in data distribution.

[0057] The embodiments of this application are based on RAN to realize the dynamic adjustment of BPNN and LSTM network parameters. New samples can be stored and retrained, enabling the model to adapt online to the time-varying characteristics of photovoltaic power with seasonal and weather changes, effectively alleviating the performance degradation problem after long-term deployment.

[0058] Among them, RAN can be through Preset The standard is used to dynamically modify hidden nodes to meet the requirements of network structure adaptation and to help with online adjustment of non-stationary processes. To avoid interference from abnormal noise, a buffer can be established in the dynamic adjustment strategy to record samples with large errors, and then the frequency of their occurrence can be recorded. Samples are selected based on the frequency to dynamically adjust the model.

[0059] The output of the RAN network is:

[0060] in, This is the power prediction value from the previous moment. For the first The output weights of the hidden nodes for The input feature vector at time t, For the first The center vector of each hidden node For the first The radial basis function width parameter of each hidden node. For the network bias term, The total number of existing hidden nodes. This is the index of the hidden node.

[0061] The constraints are:

[0062]

[0063] in, This is the error threshold.

[0064] The new sample is compared with the samples in the buffer. If no similar sample exists in the current buffer, the candidate data is recorded as a new sample in the buffer, and the support of the sample is set to 1. If a similar sample exists in the current buffer, the data is grouped into the class labeled by that similar sample. Should The support score of the sample class is increased by 1.

[0065] When all conditions are met, additional hidden nodes are dynamically added to learn the selected features and associated with the parameters:

[0066]

[0067]

[0068] in, For the newly added The center vector of each hidden node For the selected class in the buffer One labeled data, For the threshold, For the newly added The output weights of the hidden nodes For the first in the buffer The true power value of each sample For the first in the buffer The predicted power value for each sample. For the newly added The radial basis function width parameter of each hidden node. This serves as the center for newly added nodes.

[0069] To verify the feasibility of this application, it was compared with different models. Apply The predictions were validated using varying degrees of data noise and data gaps. For detailed prediction results, please refer to [link to specific data]. Figure 2 and Figure 3 As shown, when the deviation between the BPNN and LSTM results is large, triggering the bidirectional feedback mechanism, it can effectively correct damaged data within a certain range, reduce data distribution offset, and improve prediction accuracy.

[0070] To more comprehensively evaluate the predictive capability of this application, 5% noise was added to the input data during testing, and the prediction errors (NRMSE) of this application and different models under varying degrees of data corruption were calculated. Table 1 shows the errors of various models in six scenarios. It can be seen that without a feedback mechanism, model performance decreases as the degree of data corruption increases. The selected model's results are slightly better than those of Support Vector Machine (SVM) and Convolutional Neural Networks (CNN). Although choosing a suitable algorithm can slightly improve prediction accuracy, it cannot avoid the serious impact of data corruption. By using a feedback learning mechanism to achieve model fusion, the input data error is reduced in essence, and the model's prediction accuracy is improved. Table 1 clearly shows that the feedback mechanism significantly improves prediction accuracy, with the average error under the same scenario... reduce Approximately 10%.

[0071] Table 1. Error Comparison under Different Scenarios

[0072] This application provides a photovoltaic power prediction method that integrates dynamic adjustment and feedback correction of the model. The method involves: acquiring a first time series and inputting it into a first photovoltaic power prediction model constructed based on a backpropagation neural network to output a first predicted photovoltaic power sequence. The first time series includes irradiance and temperature over 24 hours before the prediction date. Then, a second time series is acquired and input into a second photovoltaic power prediction model constructed based on a long short-term memory network to output a second predicted photovoltaic power sequence. The second time series includes historical photovoltaic power over 24 hours N days before the prediction date. A first category label is determined for the first predicted photovoltaic power sequence, and a second category label is determined for the second predicted photovoltaic power sequence. Based on the first and second category labels, a target predicted photovoltaic power sequence for the prediction date is determined. This application improves prediction accuracy by integrating the advantages of two different models and comprehensively utilizing various types of data features. Furthermore, by using meteorological and historical power data, it provides rich information for prediction from different perspectives. Compared to prediction methods using only a single data source, it can more comprehensively consider various factors affecting photovoltaic power, thereby improving prediction accuracy. Simultaneously, by determining the first category label of the first predicted photovoltaic power sequence and the second category label of the second predicted photovoltaic power sequence, it provides additional classification information for the prediction results. These category labels can reflect the characteristics and reliability of the prediction results from different models. Based on these category labels, the fusion strategy can be adjusted more flexibly to further optimize the target predicted photovoltaic power sequence, while providing more targeted support for photovoltaic system operation decisions.

[0073] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0074] The following are device embodiments of this application. For details not described in detail, please refer to the corresponding method embodiments described above.

[0075] Figure 4 A schematic diagram of the photovoltaic power prediction device with dynamic adjustment and feedback correction of the fusion model provided in the embodiments of this application is shown. For ease of explanation, only the parts related to the embodiments of this application are shown, and are described in detail below: like Figure 4 As shown, the photovoltaic power prediction device 4, which integrates model dynamic adjustment and feedback correction, includes: The first output module 41 is used to acquire a first time series and input the first time series into a first photovoltaic power prediction model constructed based on a backpropagation neural network, and output a first predicted photovoltaic power series. The first time series includes irradiance and temperature within 24 hours of the day to be predicted. The second output module 42 is used to acquire the second time series and input the second time series into the second photovoltaic power prediction model constructed based on the long short-term memory network, and output the second predicted photovoltaic power series. The second time series includes the historical photovoltaic power in the 24 hours of N days before the date to be predicted. The label determination module 43 is used to determine a first category label for a first predicted photovoltaic power sequence and a second category label for a second predicted photovoltaic power sequence; The power determination module 44 is used to determine the target predicted photovoltaic power sequence for the day to be predicted based on the first category label and the second category label.

[0076] This application provides a photovoltaic power prediction device that integrates dynamic adjustment and feedback correction of the model. It acquires a first time series and inputs it into a first photovoltaic power prediction model constructed based on a backpropagation neural network, outputting a first predicted photovoltaic power sequence. The first time series includes irradiance and temperature within 24 hours of the day to be predicted. It then acquires a second time series and inputs it into a second photovoltaic power prediction model constructed based on a long short-term memory network, outputting a second predicted photovoltaic power sequence. The second time series includes historical photovoltaic power within 24 hours of N days before the day to be predicted. It determines a first category label for the first predicted photovoltaic power sequence and a second category label for the second predicted photovoltaic power sequence. Based on the first and second category labels, it determines the target predicted photovoltaic power sequence for the day to be predicted. This application improves prediction accuracy by integrating the advantages of two different models and comprehensively utilizing various types of data features. Furthermore, by using meteorological and historical power data, it provides rich information for prediction from different perspectives. Compared to prediction methods using only a single data source, it can more comprehensively consider various factors affecting photovoltaic power, thereby improving prediction accuracy. Simultaneously, by determining the first category label of the first predicted photovoltaic power sequence and the second category label of the second predicted photovoltaic power sequence, it provides additional classification information for the prediction results. These category labels can reflect the characteristics and reliability of the prediction results from different models. Based on these category labels, the fusion strategy can be adjusted more flexibly to further optimize the target predicted photovoltaic power sequence, while providing more targeted support for photovoltaic system operation decisions.

[0077] In one possible implementation, prior to the first output module, the device may further include a tag system construction module, which can be used for: Obtain historical photovoltaic power data, which includes historical photovoltaic power curves for multiple days; Historical photovoltaic power data is clustered to obtain a power category labeling system, which includes labels for sunny days, cloudy days, and rainy days.

[0078] In one possible implementation, the label determination module can specifically be used for: The first predicted photovoltaic power sequence is clustered, and the clustering results are mapped to a power category label system to obtain the first category label; The second predicted photovoltaic power sequence is clustered, and the clustering results are mapped to the power category label system to obtain the second category label.

[0079] In one possible implementation, the power determination module can specifically be used for: Determine if the first category label and the second category label are consistent; If the first category label and the second category label are the same, then the first predicted photovoltaic power sequence and the second predicted photovoltaic power sequence are weighted and fused to obtain the target predicted photovoltaic power sequence; If the first category label and the second category label are inconsistent, a bidirectional feedback correction mechanism is triggered to correct the first time series and the second time series. After the bidirectional feedback mechanism reaches the preset convergence condition, a weighted fusion calculation is performed on the first predicted photovoltaic power series corresponding to the corrected first time series and the second predicted photovoltaic power series corresponding to the corrected second time series to obtain the target predicted photovoltaic power series.

[0080] In one possible implementation, the power determination module can also be used for: The second time series is corrected using the first predicted photovoltaic power series. The first time series is corrected using the second predicted photovoltaic power series. Using the corrected first time series and the second time series, the first photovoltaic power prediction model and the second photovoltaic power prediction model are iteratively solved until the first predicted photovoltaic power sequence corresponding to the corrected first time series and the second predicted photovoltaic power sequence corresponding to the corrected second time series converge or the preset maximum number of iterations is reached.

[0081] In one possible implementation, the power determination module can also be used for: Using the power curve of the first predicted photovoltaic power sequence as the first matching template, select the first historical day corresponding to the power curve that matches the first matching template, and extract the historical actual power sequence within 24 hours of N days before the first historical day. The corrected second time series is determined by using the historical actual power at each time point and the historical photovoltaic power in the second time series.

[0082] In one possible implementation, the power determination module can also be used for: For each time point N days prior to the first historical date, the historical actual power at that time point is averaged with the historical photovoltaic power at that time point in the second time series to obtain the corrected historical photovoltaic power at that time point in the second time series.

[0083] In one possible implementation, the power determination module can also be used for: Using the power curve of the second predicted photovoltaic power sequence as the second matching template, the second historical day corresponding to the power curve that matches the second matching template is selected, and the historical actual irradiance and historical actual temperature within 24 hours of the second historical day are extracted. The second historical day is the day before the N consecutive days that a power curve that matches the second matching template exists. By using the historical actual irradiance and historical actual temperature at each time point and the irradiance and temperature in the first time series, the corrected first time series is determined.

[0084] In one possible implementation, the power determination module can also be used for: For each time point of the second historical day, the average value of the historical actual irradiance at that time point and the irradiance at that time point in the first time series is calculated to obtain the average irradiance at that time point. Similarly, the average value of the historical actual temperature at that time point and the temperature at that time point in the first time series is calculated to obtain the average temperature at that time point. The average irradiance and average temperature at that time point are used as the corrected irradiance and temperature at that time point in the first time series.

[0085] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0086] Those skilled in the art will recognize that the templates, units, and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0087] If the module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of the above embodiments of the photovoltaic power prediction method with dynamic adjustment and feedback correction of various fusion models. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory, a random access memory, an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0088] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A photovoltaic power prediction method integrating dynamic model adjustment and feedback correction, characterized in that, include: A first time series is obtained and input into a first photovoltaic power prediction model constructed based on a backpropagation neural network, and a first predicted photovoltaic power series is output. The first time series includes irradiance and temperature within 24 hours of the day to be predicted. A second time series is obtained and input into a second photovoltaic power prediction model constructed based on a long short-term memory network. The second predicted photovoltaic power series is output, and the second time series includes the historical photovoltaic power within 24 hours of N days before the date to be predicted. Determine a first category label for the first predicted photovoltaic power sequence, and determine a second category label for the second predicted photovoltaic power sequence; Based on the first category label and the second category label, the target predicted photovoltaic power sequence for the day to be predicted is determined.

2. The photovoltaic power prediction method based on dynamic adjustment and feedback correction of the fusion model according to claim 1, characterized in that, Prior to obtaining the first time series, the method further includes: Acquire historical photovoltaic power data, which includes historical photovoltaic power curves for multiple days; The historical photovoltaic power data is clustered to obtain a power category labeling system, which includes sunny day labels, cloudy day labels, and rainy day labels.

3. The photovoltaic power prediction method based on dynamic adjustment and feedback correction of the fusion model according to claim 2, characterized in that, Determining the first category label of the first predicted photovoltaic power sequence and determining the second category label of the second predicted photovoltaic power sequence includes: The first predicted photovoltaic power sequence is clustered, and the clustering result is mapped to the power category label system to obtain the first category label; The second predicted photovoltaic power sequence is clustered, and the clustering result is mapped to the power category label system to obtain the second category label.

4. The photovoltaic power prediction method based on dynamic adjustment and feedback correction of the fusion model according to claim 1, characterized in that, The step of determining the target predicted photovoltaic power sequence for the date to be predicted based on the first category label and the second category label includes: Determine whether the first category label and the second category label are consistent; If the first category label and the second category label are the same, then the first predicted photovoltaic power sequence and the second predicted photovoltaic power sequence are weighted and fused to obtain the target predicted photovoltaic power sequence; If the first category label and the second category label are inconsistent, a bidirectional feedback correction mechanism is triggered to correct the first time series and the second time series. After the bidirectional feedback mechanism reaches the preset convergence condition, a weighted fusion calculation is performed on the first predicted photovoltaic power series corresponding to the corrected first time series and the second predicted photovoltaic power series corresponding to the corrected second time series to obtain the target predicted photovoltaic power series.

5. The photovoltaic power prediction method based on dynamic adjustment and feedback correction of the fusion model according to claim 4, characterized in that, The triggering bidirectional feedback correction mechanism includes: The second time series is corrected using the first predicted photovoltaic power series. The first time series is corrected using the second predicted photovoltaic power series. Using the corrected first time series and the second time series, the first photovoltaic power prediction model and the second photovoltaic power prediction model are iteratively solved until the first predicted photovoltaic power sequence corresponding to the corrected first time series and the second predicted photovoltaic power sequence corresponding to the corrected second time series converge or the preset maximum number of iterations is reached.

6. The photovoltaic power prediction method based on dynamic adjustment and feedback correction of the fusion model according to claim 5, characterized in that, The step of correcting the second time series using the first predicted photovoltaic power series includes: Using the power curve of the first predicted photovoltaic power sequence as the first matching template, select the first historical day corresponding to the power curve that matches the first matching template, and extract the historical actual power sequence within 24 hours of N days before the first historical day. The corrected second time series is determined by using the historical actual power at each time point and the historical photovoltaic power in the second time series.

7. The photovoltaic power prediction method based on dynamic adjustment and feedback correction of the fusion model according to claim 6, characterized in that, The step of determining the corrected second time series by using the historical actual power at each time point and the historical photovoltaic power in the second time series includes: For each time point N days prior to the first historical date, the historical actual power at that time point is averaged with the historical photovoltaic power at that time point in the second time series to obtain the corrected historical photovoltaic power at that time point in the second time series.

8. The photovoltaic power prediction method based on dynamic adjustment and feedback correction of the fusion model according to claim 5, characterized in that, The step of correcting the first time series using the second predicted photovoltaic power series includes: Using the power curve of the second predicted photovoltaic power sequence as the second matching template, select the second historical day corresponding to the power curve that matches the second matching template, and extract the historical actual irradiance and historical actual temperature within 24 hours of the second historical day. The second historical day is the day before the N consecutive days that a power curve that matches the second matching template exists. The corrected first time series is determined by using the historical actual irradiance and historical actual temperature at each time point and the irradiance and temperature in the first time series.

9. The photovoltaic power prediction method based on dynamic adjustment and feedback correction of the fusion model according to claim 8, characterized in that, The step of determining the corrected first time series by comparing the historical actual irradiance and historical actual temperature at each time point with the irradiance and temperature in the first time series includes: For each time point of the second historical day, the average value of the historical actual irradiance at that time point and the irradiance at that time point in the first time series is calculated to obtain the average irradiance at that time point. The average value of the historical actual temperature at that time point and the temperature at that time point in the first time series is calculated to obtain the average temperature at that time point. The average irradiance and the average temperature at that time point are used as the irradiance and temperature corresponding to that time point in the corrected first time series.

10. A photovoltaic power prediction device integrating model dynamic adjustment and feedback correction, characterized in that, include: The first output module is used to acquire a first time series and input the first time series into a first photovoltaic power prediction model constructed based on a backpropagation neural network, and output a first predicted photovoltaic power series. The first time series includes irradiance and temperature within 24 hours of the day to be predicted. The second output module is used to acquire the second time series and input the second time series into the second photovoltaic power prediction model constructed based on the long short-term memory network, and output the second predicted photovoltaic power series. The second time series includes the historical photovoltaic power in the 24 hours of N days before the date to be predicted. A label determination module is used to determine a first category label for the first predicted photovoltaic power sequence and a second category label for the second predicted photovoltaic power sequence; A power determination module is used to determine the target predicted photovoltaic power sequence for the day to be predicted based on the first category label and the second category label.

Images