A photovoltaic power prediction method based on an attention dendritic neural network
By using an attention-based dendritic neural network approach, the problems of insufficient robustness and inadequate engineering constraints in photovoltaic power prediction are solved, achieving high-precision and stable photovoltaic power prediction, which is applicable to grid dispatch and photovoltaic power plant management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIANGSU OCEAN UNIV
- Filing Date
- 2026-02-28
- Publication Date
- 2026-06-05
AI Technical Summary
Existing photovoltaic power prediction methods lack robustness in modeling nonlinear, nonstationary, and abrupt power sequences, and do not adequately consider engineering constraints, resulting in prediction results that do not meet actual operational requirements and reducing the engineering applicability of the methods.
A high-precision photovoltaic power prediction method based on attention dendritic neural networks is achieved by constructing time-dependent feature windows, dendritic branch grouping mapping, response intensity-driven attention modulation, and fluctuation-sensing membrane potential modulation, combined with nonnegativity and capacity constraints.
It improves the accuracy and stability of photovoltaic power prediction, accurately depicts the short-term fluctuation characteristics and overall trend of photovoltaic power, and enhances the engineering applicability and generalization ability of the model.
Smart Images
Figure CN122159788A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of photovoltaic power generation technology and artificial intelligence prediction, specifically relating to a photovoltaic power prediction method based on attention dendritic neural network. Background Technology
[0002] As one of the core forms of new energy power generation, photovoltaic power generation is affected by multiple meteorological factors such as solar irradiance, temperature, cloud cover, and wind speed, exhibiting significant nonlinear, non-stationary, and abrupt characteristics. Especially under conditions such as cloud cover and sudden weather changes, photovoltaic power is prone to drastic fluctuations, posing challenges to the safe and stable operation of the power system, grid dispatch optimization, and the absorption of new energy sources.
[0003] Existing photovoltaic power prediction methods mainly include statistical methods, traditional machine learning methods, and deep learning methods. Statistical methods have a simple structure but limited ability to characterize nonlinear features; traditional machine learning methods can capture some nonlinear relationships, but their multi-scale feature extraction capabilities are insufficient. Deep learning methods, such as recurrent neural networks and long short-term memory networks, have been applied in time series prediction. The introduction of attention mechanisms has further improved the ability to capture key features. However, these models use traditional neuron structures and cannot adaptively adjust the model sensitivity according to the intensity of power fluctuations, resulting in poor robustness in power surge scenarios.
[0004] Meanwhile, some existing methods do not fully consider the engineering constraints of photovoltaic power, such as power non-negativity and upper limit constraints on installed capacity, which can easily lead to prediction results that do not meet actual operating requirements, reducing the engineering applicability of the methods. Therefore, there is an urgent need for a prediction method that can accurately characterize the complex characteristics of photovoltaic power, adapt to fluctuating operating conditions, and take into account engineering practicality. (Note: The research content on time series prediction and Transformer-type models is mainly extracted and compiled from publicly published academic literature, including Informer, Autoformer, FEDformer, PatchTST, and TimeXer models. The relevant technical solutions are disclosed in the proceedings of international academic conferences such as NeurIPS, ICML, and ICLR, as well as the arXiv preprint platform. All of these are prior art that can be obtained by those skilled in the art through public channels before the application date.)
[0005]
[11] Zhou H., Zhang S., Peng J., Zhang S., Li J., Xiong H., Zhang W.Informer: Beyond efficient transformer for long sequence time-seriesforecasting[C] / / Proceedings of the AAAI Conference on ArtificialIntelligence, 2021, 35(12): 11106-11115.
[0006]
[12] Wu H., Xu J., Wang J., Long M. Autoformer: Decompositiontransformers with auto-correlation for long-term series forecasting[J].Advances in Neural Information Processing Systems, 2021, 34: 22419-22430.
[0007]
[13] Zhou T., Ma Z., Wen Q., Wang X., Sun L., Jin R. Fedformer:Frequency enhanced decomposed transformer for long-term series forecasting[C] / / Proceedings of the International Conference on Machine Learning (ICML),2022: 27268-27286.
[0008]
[14] Zeng A., Chen M., Zhang L., Xu Q. Are transformers effective fortime series forecasting?[C] / / Proceedings of the AAAI Conference onArtificial Intelligence, 2023, 37(9): 11121-11128.
[0009]
[15] Wang Y., Wu H., Dong J., Qin G., Zhang H., Liu Y., Long M. Timexer: Empowering transformers for time series forecasting with exogenous variables[J]. Advances in Neural Information Processing Systems, 2024, 37:469-498. Summary of the Invention
[0010] To address the shortcomings of existing photovoltaic power prediction methods in modeling nonlinear, nonstationary, and abrupt power sequences, and their insufficient consideration of engineering constraints, this invention provides a photovoltaic power prediction method based on attention dendritic neural networks. This method achieves high-precision and high-stability prediction of photovoltaic power, and improves the method's engineering adaptability.
[0011] A photovoltaic power prediction method based on attention dendritic neural network includes the following steps:
[0012] (1) Construction of photovoltaic power time series and amplitude constraint processing:
[0013] Collect historical power generation data of photovoltaic power plants on a continuous time scale to construct a photovoltaic power time series:
[0014]
[0015] To eliminate amplitude differences caused by varying installed capacity and operating conditions, amplitude constraints and scaling are applied to the power sequence:
[0016]
[0017] in, To prevent constants with a denominator of zero;
[0018] This step ensures that the input power sequence is within a uniform numerical range, reducing the interference of extreme operating conditions on model training; the historical maximum power value is the maximum measured power generation of the photovoltaic power station within a preset historical time period, which is used as the benchmark value for photovoltaic power normalization processing.
[0019] (2) Construction of time-dependent feature windows:
[0020] Based on the sliding window mechanism, time-dependent input samples are constructed from the normalized power sequence:
[0021]
[0022] in, The time window length is used to characterize the variation characteristics of photovoltaic power within the time range covered by the sliding time window. The time granularity of the sliding time window can be set to minutes or hours according to the data acquisition frequency. Its length corresponds to several continuous time sampling points, which are used to characterize the variation characteristics of photovoltaic power within the time range covered by the sliding time window.
[0023] (3) Grouping mapping modeling of dendritic branches:
[0024] Time window characteristics Divided into Each subinterval corresponds to a dendritic branch:
[0025]
[0026] Each sub-interval represents a segment of power change at different time scales; different dendritic branches correspond to different time sub-intervals within the sliding time window, used to extract the variation characteristics of photovoltaic power at different times within the window.
[0027] The response of the m-th dendritic branch is calculated as follows:
[0028]
[0029] in, For dendritic branch weights, For bias terms, It is a non-linear mapping function.
[0030] This step enables the modeling of photovoltaic power characteristics across different time sub-intervals within a sliding time window;
[0031] (4) Attention modulation mechanism based on response intensity:
[0032] To highlight dendritic branches that contribute significantly to power changes, response intensity-driven attention weights are introduced:
[0033]
[0034] Dendritic branch outputs are then fused after attention weighting:
[0035]
[0036] This attention mechanism does not rely on external query vectors but is directly generated adaptively by the dendritic response amplitude, making it suitable for non-stationary time series modeling. The response intensity of the dendritic branch can be characterized by the numerical amplitude of its output features, which reflects the degree of contribution of the branch to the change in photovoltaic power.
[0037] (5) Wave-sensing membrane potential modulation:
[0038] Considering the strong fluctuation characteristics of photovoltaic power under scenarios such as cloud shading and sudden changes in irradiance, a modulation term based on the time window discreteness is introduced;
[0039]
[0040] in, The mean value is used as the window mean. The dispersion of photovoltaic power within the time window is preferably characterized by the standard deviation, which is used to measure the fluctuation level of the power sequence within the window.
[0041] Incorporating a wave-sensing factor into the membrane potential calculation process:
[0042]
[0043] in, The modulation coefficient, It is a non-linear activation function;
[0044] This step enables the model to adaptively adjust its prediction sensitivity based on the stability of the input sequence.
[0045] (6) Photovoltaic power prediction output:
[0046] Based on membrane potential state Generate the photovoltaic power prediction value for the next time step:
[0047]
[0048] in, Represents the prediction mapping function;
[0049] (7) Constraints and applications of prediction results:
[0050] Apply nonnegativity and capacity constraints to the prediction results:
[0051]
[0052] The final prediction results can be used for grid dispatch, energy storage control, and photovoltaic power plant operation and management.
[0053] The present invention has the following advantages:
[0054] Data Collection and Preprocessing: Historical power generation data from different photovoltaic application scenarios are collected as experimental samples, including publicly available photovoltaic power plant data and power data from photovoltaic power plants operating under actual grid conditions. The data acquisition time resolution can be 15 minutes or 1 hour, and operational information at the corresponding time scale is acquired simultaneously. Missing value imputation, outlier removal, and smoothing are performed on the raw photovoltaic power data to ensure data integrity and consistency. To reduce the impact of different weather conditions, installed capacity, and operating status on model training, this embodiment standardizes the photovoltaic power sequence, and the calculation method is as follows:
[0055]
[0056] in, The sample mean. This represents the sample standard deviation.
[0057] During the prediction phase, an invariant standardization operation is performed on the model output to recover the true power scale:
[0058]
[0059] Model Construction: The prediction model is constructed using the ADNM-PV (Photovoltaic Power Prediction Method Based on Attention Dendritic Neural Network) proposed in this patent. The model takes historical photovoltaic power time series as input and constructs supervised learning samples using a sliding time window approach. Input features are first grouped and mapped through a dendritic branch structure, with different branches corresponding to different time scales or power change segments to characterize the differences in photovoltaic power during short-term fluctuations and medium-term changes. Subsequently, an attention modulation mechanism based on dendritic response intensity is introduced to dynamically weight the outputs of each branch, thereby highlighting the feature branches that contribute more to power changes.
[0060] Feature Fusion and Prediction Output: In the dendritic feature fusion stage, a unified feature representation is formed through weighted aggregation. Furthermore, a modulation factor based on the time window fluctuation is introduced to nonlinearly adjust the fusion result, enhancing the model's response to sudden changes in photovoltaic power and unstable operating conditions. Finally, the model generates the photovoltaic power prediction result for the target time step based on the modulated membrane potential output. The prediction process can be expressed as follows:
[0061]
[0062] in, This represents the nonlinear mapping function constructed using the ADNM-PV method.
[0063] Model optimization and parameter update: During the model training phase, mean squared error is used as the primary optimization objective. Simultaneously, multiple indicators are introduced as joint constraints to improve prediction stability. The loss function is defined as follows:
[0064]
[0065] in, , These are weighting coefficients used to balance the influence of different error indicators.
[0066] Meanwhile, the model parameters are updated iteratively, and the update rule can be expressed as follows:
[0067]
[0068] in, For learning rate, This represents the change in the parameter gradient during the current iteration.
[0069] Results Evaluation and Validation: Multiple evaluation metrics were used to comprehensively evaluate the prediction results, including mean squared error (MSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and correlation coefficient (R). Validation was performed experimentally on the Belgian photovoltaic dataset and the island photovoltaic dataset (see Table 1 and...). Figure 1 In this embodiment, the ADNM-PV method outperforms mainstream prediction models such as Informer, Autoformer, FEDformer, PatchTST, and TimeXer in all evaluation metrics. Experimental results show that this method can accurately characterize the short-term fluctuations and overall trends of photovoltaic power, and maintains high prediction accuracy and stability even under complex meteorological conditions and non-stationary operating environments.
[0070] Implementation Results: As can be seen from the above embodiments, the ADNM-PV photovoltaic power prediction method proposed in this patent can effectively reduce the prediction error of non-stationary photovoltaic power sequences, improve the prediction accuracy and model generalization ability under different application scenarios, and provide reliable power prediction basis for grid dispatch, energy storage system control and new energy power generation management. It has good engineering application value and promotion prospects. Attached Figure Description
[0071] Figure 1 This is a flowchart of the present invention;
[0072] Figure 2 This is a visualization of the 24-hour photovoltaic power generation forecast results for islands, as presented in this invention.
[0073] Figure 3This is a visualization of the 24-hour photovoltaic (PV) forecast results in Belgium, as presented in this invention. Detailed Implementation
[0074] The specific technical solutions of the present invention are further described below to enable those skilled in the art to further understand the present invention, without constituting a limitation on its rights. To verify the accuracy and stability of the photovoltaic power prediction method based on attention dendritic neural networks proposed in this invention in new energy power generation prediction tasks, a large number of numerical experiments and model comparison analyses were conducted.
[0075] This patent selects five representative advanced prediction models from the literature as comparison objects, including mainstream time series prediction models such as Informer
[11] , Autoformer
[12] , FEDformer
[13] , PatchTST
[14] and TimeXer
[15] . These models have high representativeness and good prediction performance in the field of photovoltaic power and time series prediction. They can comprehensively compare the prediction effects from different modeling ideas and structural characteristics, so as to objectively evaluate the advantages and applicability of the method of this invention. Since there is currently no unified standard test dataset for photovoltaic power prediction, this patent selects two representative photovoltaic power datasets as experimental samples. On the one hand, publicly released photovoltaic power generation data is used as a verification sample. This data covers photovoltaic output on a continuous time scale and includes complete solar cycle and seasonal variation characteristics. On the other hand, historical power data of island-type photovoltaic power plants from the actual power grid operation environment are selected. This data is significantly affected by weather changes and operating conditions, and has the characteristics of frequent power fluctuations and strong non-stationarity, which can truly reflect the prediction difficulty in engineering application scenarios. In terms of experimental setup, the photovoltaic power data was divided chronologically, with the first 70% of the historical data used as the training set and the remaining 30% as the test set to ensure the temporal consistency of the prediction process and the objectivity of the results. All comparison models adopted the same data preprocessing procedure, input time window length, and prediction step size, and were trained and tested under the same hardware environment and hyperparameter configuration conditions to ensure the fairness and repeatability of the experimental results.
[0076] To comprehensively evaluate the predictive performance of each model,
[0077] This patent selects four key indicators: mean squared error (MSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and correlation coefficient (R). MSE and MAE measure the magnitude of prediction deviation, MAPE reflects the relative proportion of prediction error, and R characterizes the degree of linear correlation between the predicted result and the actual value. By calculating these evaluation indicators on photovoltaic power datasets from different sources, the prediction accuracy, stability, and generalization ability of the method of this invention under various meteorological conditions and complex operating environments can be systematically evaluated.
[0078] Indicator 1: MSE:
[0079] (1)
[0080] Indicator 2: MAE
[0081] (2)
[0082] Indicator 3: MAPE:
[0083] (3)
[0084] Indicator 4: R:
[0085]
[0086] in The total number of samples, This is the actual value. These are predicted values.
[0087] The experimental results are shown in the table below:
[0088]
[0089] As shown in Table 1, in the comparative experiments with different models, the ADNM-PV model proposed in this patent outperforms the comparative models in terms of MSE, MAE, MAPE, and correlation coefficient R, exhibiting lower prediction error and higher correlation. On the Belgian photovoltaic dataset and the island photovoltaic dataset, the ADNM-PV method significantly reduces the overall prediction error and significantly improves the correlation coefficient compared to mainstream models such as TimeXer, indicating that the method has good prediction stability and generalization ability under different data sources and operating conditions. Furthermore, in the 15-minute short-timescale prediction task, the ADNM-PV method maintains a stable advantage on both datasets, effectively characterizing the rapid fluctuations in photovoltaic power, demonstrating that the method of this invention has strong adaptability and engineering application value under non-stationary and abrupt operating conditions.
[0090] Depend onFigure 1 It can be seen that photovoltaic power exhibits obvious periodic fluctuations over time. The ADNM-PV model proposed in this patent can effectively characterize the short-term fluctuations and overall trends of photovoltaic power, and its prediction results maintain a high degree of consistency with the actual observation sequence in both the training and prediction phases. In contrast, other comparative models show varying degrees of prediction bias during power peak changes and sudden fluctuations, and their stability decreases within the prediction interval. The ADNM-PV method, however, maintains relatively stable and accurate prediction performance, demonstrating strong robustness and engineering applicability.
[0091] In summary, the ADNM-PV model proposed in this patent can effectively reduce prediction errors and significantly improve prediction accuracy and stability in complex and non-stationary photovoltaic power time-series data. This method can provide reliable power prediction support for photovoltaic power plant operation and management, grid dispatch, and renewable energy consumption, and has good engineering application value and promotion significance.
Claims
1. A photovoltaic power prediction method based on attention dendritic neural network, characterized in that: Includes the following steps: (1) Construction of photovoltaic power time series and amplitude constraint processing: Collect historical power generation data of photovoltaic power plants on a continuous time scale to construct a photovoltaic power time series: To eliminate amplitude differences caused by varying installed capacity and operating conditions, amplitude constraints and scaling are applied to the power sequence: in, To prevent constants with a denominator of zero; This step ensures that the input power sequence is within a uniform numerical range, reducing the interference of extreme conditions on model training; (2) Construction of time-dependent feature windows: Based on the sliding window mechanism, time-dependent input samples are constructed from the normalized power sequence: in, Indicates the length of the time window, used to depict the trajectory of photovoltaic power changes over a recent time range: (3) Grouping mapping modeling of dendritic branches: Time window characteristics Divided into Each subinterval corresponds to a dendritic branch: Each sub-interval represents a segment of power change at different time scales; The response of the m-th dendritic branch is calculated as follows: in, For dendritic branch weights, For bias terms, It is a nonlinear mapping function; This step enables parallel feature extraction of photovoltaic power at different time scales; (4) Attention modulation mechanism based on response intensity: To highlight dendritic branches that contribute significantly to power changes, response intensity-driven attention weights are introduced: Dendritic branch outputs are then fused after attention weighting: This attention mechanism does not depend on external query vectors, but is directly generated adaptively by the dendritic response amplitude, making it suitable for non-stationary time series modeling; (5) Wave-sensing membrane potential modulation: Considering the strong fluctuation characteristics of photovoltaic power under scenarios such as cloud shading and sudden changes in irradiance, a modulation term based on the time window discreteness is introduced; in, The mean of the window; Incorporating a wave-sensing factor into the membrane potential calculation process: in, The modulation coefficient, It is a non-linear activation function; This step enables the model to adaptively adjust its prediction sensitivity based on the stability of the input sequence. (6) Photovoltaic power prediction output: Based on membrane potential state Generate the photovoltaic power prediction value for the next time step: in, Represents the prediction mapping function; (7) Constraints and applications of prediction results: Apply nonnegativity and capacity constraints to the prediction results: The final prediction results can be used for grid dispatch, energy storage control, and photovoltaic power plant operation and management.
2. The photovoltaic power prediction method based on attention dendritic neural network according to claim 1, characterized in that: The amplitude constraint and normalization process in step (2) includes using the maximum measured power generation of the photovoltaic power station within a preset historical time period as the normalization benchmark value to perform scale mapping on the photovoltaic power data in order to eliminate amplitude differences caused by different installed capacities and operating conditions.
3. The photovoltaic power prediction method based on attention dendritic neural network according to claim 1, characterized in that: The time-dependent input samples in step (2) are constructed by a sliding time window of fixed length. The length of the sliding time window is several consecutive time sampling points, corresponding to minute-level or hour-level time granularity, which is used to characterize the variation characteristics of photovoltaic power within the time range covered by the sliding time window.
4. The photovoltaic power prediction method based on attention dendritic neural network according to claim 1, characterized in that: The dendritic branches mentioned in step (3) are multiple parallel feature processing units. Each dendritic branch corresponds to a different time sub-interval within the sliding time window, which is used to extract the variation features of photovoltaic power at different time positions.
5. The photovoltaic power prediction method based on attention dendritic neural network according to claim 1, characterized in that: The response intensity mentioned in step (4) is the numerical amplitude of the dendritic branch output feature. The attention modulation mechanism generates corresponding weight coefficients based on the numerical value of the dendritic branch output feature to highlight the dendritic branches that contribute significantly to the change in photovoltaic power.
6. The photovoltaic power prediction method based on attention dendritic neural network according to claim 1, characterized in that: The modulation factor based on the degree of fluctuation of the time window in step (5) is calculated from the standard deviation of the photovoltaic power data within the input sliding time window, and is used to enhance the model's ability to respond to power abrupt changes and non-stationary changes.
7. The photovoltaic power prediction method based on attention dendritic neural network according to claim 1, characterized in that: The nonnegativity and capacity constraints mentioned in step (7) are used to ensure that the predicted photovoltaic power result is not less than zero and does not exceed the rated installed capacity of the photovoltaic power station.
8. The photovoltaic power prediction method based on attention dendritic neural network according to any one of claims 1 to 7, characterized in that: The method is applicable to scenarios such as power grid dispatching, energy storage system control, new energy consumption, or photovoltaic power plant operation and management.