Photovoltaic power prediction method based on DNN-physics combined model

Abstract

A photovoltaic power prediction method based on a DNN-physical combination model includes the following steps: step S1: obtaining historical electricity and meteorological data for the region; step S2: preprocessing the given data set and determining whether the data set is complete; step S3: allocating weight factors for building the DNN model and the physical model. If the data set is complete, the weight factors tend to favor the DNN model. If the data set is incomplete, the weight distribution of the factors allocated to the physical model should tend to favor the physical model; step S4: the DNN model and the physical model adopt a parallel modeling strategy, simultaneously performing model building and verification evaluation; step S5: combining the DNN model and the physical model into a hybrid model for evaluation and optimization; step S6: outputting the photovoltaic power prediction result based on the hybrid model. This invention overcomes the limitations of a single model in data processing and prediction capabilities, realizes adaptive weight allocation under different data set conditions, and thus improves the accuracy and stability of photovoltaic power prediction.

Description

Technical Field

[0001] The present invention belongs to the field of new energy prediction for power systems, and specifically relates to a photovoltaic power prediction method based on a DNN-physics combined model. Background Art

[0002] Solar energy is a clean, renewable, and unlimited energy source. Its power generation process is pollution-free and noise-free, significantly reducing greenhouse gas emissions and environmental pollution, while also improving air quality and conserving water resources. Furthermore, solar power generation can promote economic development, create jobs, and enhance energy security and independence. Photovoltaic power generation, one of the most important forms of solar power generation, has become a research hotspot worldwide. Photovoltaic power generation systems, influenced by multiple natural factors such as solar radiation intensity, temperature, humidity, and cloud cover, exhibit significant randomness and nonlinear characteristics, making the prediction of their output power extremely challenging. Furthermore, this volatility and randomness, exacerbated by constantly changing meteorological and environmental conditions, negatively impact the integration of large-scale photovoltaic power generation into the grid and pose a threat to the stable operation of the power grid.

[0003] The present application differs from the prior art in the following ways:

[0004] Technical comparison with patent CN112507603B "Identification method for extreme scenarios of robust optimization of power systems based on DNN algorithm"

[0005] 1. The input data set of the DNN algorithm constructed by patent CN112507603B mainly consists of box-type uncertainty sets, specifically including m load power and n photovoltaic power generation output parameters. Our DNN model input layer specially designed five neurons to cover key influencing factors such as ambient temperature, system capacity, wind speed, solar irradiance, photovoltaic panel type, and additional ambient humidity. This makes the design close to the actual physical process of photovoltaic power generation and accurately transmits environmental parameters for calculation.

[0006] Second, patent CN112507603B is a method for identifying extreme scenarios for robust optimization of power systems based on a DNN algorithm. This method simplifies the steps involved in solving extreme scenarios and improves their efficiency. Our photovoltaic power prediction method, based on a combined DNN-physics model, combines the physical clarity of the physical model with the high precision of the DNN model, aiming to improve both the accuracy and practicality of photovoltaic power technology predictions.

[0007] Technical comparison with patent CN117708716B "Photovoltaic power prediction method and equipment based on regression and time series fusion"

[0008] 1. Patent CN 117708716B employs a prediction method that combines random forests, histogram gradient boosting regression (HGBR), and deep neural networks (DNNs). The regression results of these models are fused through stacking ensemble technology. We innovatively fuse a physical model with clear physical meaning and a high-precision DNN model. During the prediction process, we dynamically weight the two methods: when data is scarce or DNN performance is limited, we increase the weight of the physical model, using physical principles as a guide to compensate for the limitations of the data-driven model. When data is sufficient, we increase the weight of the DNN model. This strategy ensures efficient and accurate predictions in various scenarios.

[0009] Second, patent CN 117708716B utilizes the Informer model for long-term time series forecasting, leveraging large-scale datasets and long-sequence information to improve forecast accuracy and efficiency. The combined model we designed not only achieves accurate forecasts when data is sufficient, but also demonstrates a powerful forecast enhancement effect when data is scarce, effectively addressing the complexity and uncertainty in photovoltaic power forecasting.

[0010] Technical comparison with patent CN118643328A "Power transfer learning prediction method for new photovoltaic stations based on coupled irradiance and temperature"

[0011] First, patent CN118643328A analyzes how photovoltaic modules absorb solar irradiance, first using physical modeling to extract slope irradiance feature modules. After extracting these feature information through physical modeling, a DNN model is trained to further extract temperature features, which are then combined with other models for prediction. We specifically perform precise physical modeling based on manufacturer-provided data and cleverly integrate this with the DNN model. Using a collaborative strategy, we perform parallel modeling at the same time, then assign weights to combine these models into a hybrid model for power prediction. This approach preserves the accuracy of the physical model while leveraging the DNN's strengths in complex feature extraction. This significantly reduces reliance on field-collected data and improves prediction efficiency and accuracy.

[0012] Second, patent CN118643328A first requires obtaining relevant parameter data through a slope irradiance feature extraction module to predict photovoltaic power. However, in our designed DNN-physics combined model, the physics model does not rely on large amounts of data to operate effectively. It relies solely on readily available manufacturer data sheets, significantly reducing the burden of data collection and improving the model's ease of application and practicality.

[0013] Third, patent CN118643328A constructs a prediction model by integrating multiple feature extraction modules. It first pre-trains the model on data from PV plant A, then adaptively fine-tunes the model parameters based on data from PV plant B, ultimately accurately predicting PV plant B's power generation in future time periods. Our DNN-physics combined model, based on deep learning of historical data using a DNN model and combining it with a precise description of PV panels using a physical model, can more effectively capture the dynamic changes in PV plant output power and effectively address the randomness and intermittent nature of meteorological conditions. This approach reduces reliance on massive amounts of historical data. Summary of the Invention

[0014] In response to the above problems, the present invention proposes a photovoltaic power prediction method based on a DNN-physics combined model. This method combines the physical meaning clarity of the physical model with the high-precision characteristics of the DNN model, aiming to achieve a dual improvement in prediction accuracy and practicality. Specifically, the physical model, as a supplement, can provide guidance based on physical principles when data is scarce or the performance of the DNN model is poor, ensuring the stability and reliability of the prediction results. At the same time, the DNN model uses its powerful data processing and feature extraction capabilities to accurately capture the complex characteristics of the photovoltaic system in actual operation, such as spectral loss, dirt accumulation, and component mismatch, thereby correcting the deviation of the physical model under ideal conditions and improving prediction accuracy.

[0015] To achieve the above object, the technical solution adopted by the present invention is:

[0016] A photovoltaic power prediction method based on a DNN-physics combined model is characterized by comprising the following steps:

[0017] Step S1: Obtain historical power and meteorological data for the area;

[0018] Step S2: preprocess the given data set and determine whether the data set is complete;

[0019] Step S3: Assign weight factors to build the DNN model and the physical model. When the dataset is complete, the weight factors tend to favor the DNN model. If the dataset is incomplete, the weight factors assigned to the physical model tend to favor the physical model.

[0020] Step S4: The DNN model and the physical model adopt a parallel modeling strategy, and the model construction and verification evaluation are carried out simultaneously;

[0021] Step S5: The DNN model and the physical model are combined into a hybrid model for evaluation and optimization;

[0022] Step S6: Outputting photovoltaic power prediction results based on the hybrid model.

[0023] As a further improvement to the present invention, step S3 specifically involves assigning weights to the DNN model and the physical model based on the completeness of the dataset. When the dataset is complete, the weighting factors favor the DNN model more to leverage its strong data-fitting capabilities. When the dataset is incomplete, the weighting factors favor the physical model to leverage its predictive stability and reliability based on physical principles. This dynamic weighting strategy helps improve overall prediction performance.

[0024] As a further improvement of the present invention, the specific step S4 is specifically:

[0025] Using a parallel modeling approach, the physical model and the DNN model are built together. Based on the data provided by the manufacturer, the physical model is built as follows:

[0026]

[0027] As a further improvement of the present invention, the DNN model is used for photovoltaic power prediction. Specifically, the model input layer includes five neurons specifically designed to account for key influencing factors such as ambient temperature, system capacity, wind speed, solar irradiance, photovoltaic panel type, and additional ambient humidity. Input data is divided into numerical and categorical types. The categorical data is encoded using a OneHotEncoder and then input. This data is processed and transformed in the DNN hidden layer. The neuron outputs are normalized using an activation function and passed layer by layer to form a data stream. Finally, the output layer generates a photovoltaic power prediction result to verify the accuracy of the model.

[0028] As a further improvement of the present invention, step S5 is specifically as follows: organically integrating the two, giving full play to the advantages of the DNN model in processing complex nonlinear relationships, and the accuracy of the physical model in explaining physical phenomena, effectively reducing the limitations of each of them, and then strictly verifying the accuracy of the hybrid model.

[0029] As a further improvement to the present invention, step S6 specifically involves comprehensively analyzing the performance of the hybrid model using evaluation metrics and optimizing the model accordingly. Ultimately, the model successfully outputs accurate photovoltaic power prediction results, demonstrating its superior performance in the field of photovoltaic power prediction.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] This method combines the powerful data processing capabilities of DNNs with the precision of physical models to significantly improve the accuracy and stability of photovoltaic power forecasts. By leveraging DNN's deep learning of historical data with the precise description of photovoltaic panels by physical models, this combined model can better capture the dynamic characteristics of photovoltaic power plant output power and effectively address the random and intermittent effects of meteorological conditions such as solar radiation and temperature. Furthermore, this method reduces reliance on large amounts of historical data and is suitable for power forecasting of both new and existing photovoltaic plants, providing strong support for the safe and stable operation of power systems and energy management. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a flow chart of the method of the present invention;

[0033] Figure 2 This is a DNN model framework diagram for photovoltaic prediction in one embodiment of the present invention;

[0034] Figure 3 is the corresponding photovoltaic output curve in one embodiment of the present invention. DETAILED DESCRIPTION

[0035] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0036] Flowchart as Figure 1 As shown, a photovoltaic power prediction method based on a DNN-physics combined model is characterized by comprising the following steps:

[0037] Step S1: Obtain historical power and meteorological data for the area;

[0038] Step S2: preprocess the given data set and determine whether the data set is complete;

[0039] Step S3: Assign weight factors to build the DNN model and the physical model. When the dataset is complete, the weight factors tend to favor the DNN model. If the dataset is incomplete, the weight factors assigned to the physical model tend to favor the physical model.

[0040] Step S3 specifically involves assigning weights to the DNN model and the physical model based on the completeness of the dataset. When the dataset is complete, the weighting factor favors the DNN model to leverage its strong data-fitting capabilities. When the dataset is incomplete, the weighting factor favors the physical model to leverage its predictive stability and reliability based on physics principles. This dynamic weighting strategy helps improve overall prediction performance.

[0041] Step S4: The DNN model and the physical model adopt a parallel modeling strategy, and the model construction and verification evaluation are carried out simultaneously;

[0042] The specific step S4 is to build the physical model and the DNN model together using a parallel modeling method. According to the data provided by the manufacturer, the physical model is built as follows:

[0043]

[0044] The DNN model is used for photovoltaic power prediction, and its DNN model framework is shown in the figure below. Figure 2 As shown in the figure, the model input layer has five specially designed neurons to account for key influencing factors such as ambient temperature, system capacity, wind speed, solar irradiance, PV panel type, and additional ambient humidity. Input data is divided into numerical and categorical types, with categorical data encoded using a OneHotEncoder before input. The data is processed and transformed in the DNN hidden layer, and the neuron outputs are normalized using activation functions. This data is then passed layer by layer to form a data flow. Finally, the output layer generates PV power prediction results to verify the model's accuracy.

[0045] Step S5: The DNN model and the physical model are combined into a hybrid model for evaluation and optimization;

[0046] Step S5 specifically involves organically integrating the two. By leveraging the DNN model's strengths in handling complex nonlinear relationships and the physical model's accuracy in explaining physical phenomena, the limitations of each are effectively mitigated. Subsequently, the hybrid model undergoes rigorous accuracy verification.

[0047] Step S6: Outputting photovoltaic power prediction results based on the hybrid model.

[0048] Step S6 specifically involves comprehensively analyzing the performance of the hybrid model using evaluation indicators and optimizing the model accordingly. Ultimately, the model successfully outputs accurate photovoltaic power prediction results, demonstrating its superior performance in the field of photovoltaic power prediction.

[0049] As a specific embodiment of the present invention, the present invention provides a photovoltaic power prediction method based on a DNN-physics combined model, comprising the following steps:

[0050] Step S1: Obtain historical power and weather data for the area:

[0051] In this example, we address the photovoltaic power forecasting problem for a specific region. To predict local photovoltaic power levels, we first collect comprehensive historical regional electricity and meteorological data, including temperature, humidity, cloud cover, and solar irradiance. We also record parameters such as photovoltaic panel type and capacity, and wind speed at the installation site. Data collection is performed with a 20-minute resolution to capture subtle changes and provide a rich and accurate foundation for subsequent model training.

[0052] Step S2: Preprocess the given data set and determine whether the data set is complete:

[0053] In this example, after obtaining the raw dataset, a series of preprocessing steps, such as data cleaning and removing missing values, are required to improve data quality and model training effectiveness. Furthermore, the dataset's completeness must be assessed to verify that it contains all necessary parameters and information, as well as the continuity of the time series.

[0054] Step S3: Assign weight factors to build the DNN model and the physical model. When the dataset is complete, the weight factors tend to favor the DNN model. If the dataset is incomplete, the weight factors tend to favor the physical model.

[0055] In this example, when the dataset is relatively complete, the DNN model, due to its strong data-fitting capabilities, is given a higher weight, leveraging the patterns it has learned from large-scale data. For example, when the dataset is complete, the weight factors for the DNN model and the physical model are 0.9352 and 0.0648, respectively. Conversely, when the dataset is incomplete or contains many missing values, the physical model receives more weight due to its predictive stability and reliability based on physical principles. This dynamic weight allocation strategy aims to balance the strengths of the two models to adapt to prediction needs under different dataset conditions. For example, when the dataset is incomplete, the weight factors for the DNN model and the physical model are 0.8087 and 0.1913, respectively.

[0056] Step S4: The DNN model and the physical model adopt a parallel modeling strategy, and the model construction and verification evaluation are carried out simultaneously:

[0057] In this example, after determining the weighting factors for the DNN model and the physical model, a parallel modeling strategy was used to build the model. The two models are trained independently during the training process, but will be combined and optimized later to achieve better prediction results.

[0058] Step S5: The DNN model and the physical model are combined into a hybrid model for evaluation and optimization:

[0059] In this example, the hybrid model's predictive performance is verified by comparing it with actual observed data to assess its accuracy and stability. If the prediction results are biased or insufficient, the hybrid model's parameters need to be adjusted and optimized to improve its predictive ability.

[0060] Step S6: Based on the hybrid model, output the photovoltaic power prediction result:

[0061] Figure 3The following are the prediction curves (the red line represents the fit) and model evaluation metrics for an incomplete dataset. After training and optimization in the above steps, the hybrid model based on the DNN-physics combined model is now capable of accurately predicting PV power. At this point, new meteorological data and PV system parameters can be input to perform real-time predictions using the hybrid model. The prediction results will include PV output power values ​​for a future period (such as hourly or daily), as well as the corresponding prediction error range.

[0062] The above description is merely a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent variation based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A photovoltaic power prediction method based on a DNN-physics combined model, characterized by: The steps include: Step S1: Obtain historical power and meteorological data for the region; Step S2: preprocess the given data set and determine whether the data set is complete; Step S3: Allocate weight factors for building the DNN model and the physical model. When the dataset is complete, the weight factors tend to favor the DNN model. If the dataset is incomplete, the weight factors tend to favor the physical model. Specifically, step S3 includes allocating weight factors of the DNN model and the physical model based on the completeness of the data set. When the data set is complete, the weight factor is more inclined towards the DNN model to take advantage of its strong data fitting ability. When the data set is incomplete, the weight factor of the physical model is increased to take advantage of its prediction stability and reliability based on physical principles. Step S4: The DNN model and the physical model adopt a parallel modeling strategy, and the model construction and verification evaluation are carried out simultaneously; The specific step S4 is specifically as follows: Using a parallel modeling approach, the physical model and the DNN model are built together. Based on the data provided by the manufacturer, the physical model is built as follows: ； Step S5: The DNN model and the physical model are combined into a hybrid model for evaluation and optimization; Step S6: Outputting photovoltaic power prediction results based on the hybrid model.

2. The photovoltaic power prediction method based on the DNN-physical combined model according to claim 1, characterized in that: The DNN model is used for photovoltaic power prediction. Specifically, the model input layer is designed to contain 5 neurons, corresponding to the five key factors of air temperature, capacity, wind speed, solar irradiance and PV panel type, as well as the additional key influencing factor of ambient humidity. The input data is divided into numerical and categorical types. The categorical data is encoded by OneHotEncoder and input. The data is processed and transformed in the DNN hidden layer, and the neuron output is standardized by the activation function and passed layer by layer to form a data stream. Finally, the photovoltaic power prediction result is generated in the output layer to verify the accuracy of the model.

3. The photovoltaic power prediction method based on the DNN-physical combined model according to claim 1, characterized in that: The step S5 specifically includes: organically integrating the two, giving full play to the advantages of the DNN model in processing complex nonlinear relationships, and the accuracy of the physical model in explaining physical phenomena, effectively alleviating the limitations of each of them, and then strictly verifying the accuracy of the hybrid model.

4. The photovoltaic power prediction method based on the DNN-physical combined model according to claim 1, characterized in that: The step S6 specifically includes: comprehensively analyzing the performance of the hybrid model through evaluation indicators, and optimizing and adjusting the hybrid model accordingly. Ultimately, the hybrid model successfully outputs accurate photovoltaic power prediction results, demonstrating its superior performance in the field of photovoltaic power prediction.

Images