A photovoltaic power prediction method and device based on multi-source heterogeneous data

Abstract

The application discloses a photovoltaic power prediction method and device based on multi-source heterogeneous data, and belongs to the technical field of photovoltaic power prediction. The method comprises the following steps: acquiring current meteorological data of a current time step and photovoltaic power sequence data containing historical photovoltaic power and historical meteorological data of historical time steps; inputting the two types of data into a photovoltaic power prediction model, extracting current meteorological features and historical meteorological time sequence features from the current meteorological data and the historical meteorological data by a meteorological feature extraction channel of the model; extracting power time sequence features from the historical photovoltaic power by a power feature extraction channel of the model; and finally predicting target photovoltaic power of a future time step by fusing the three types of features and based on the fused feature sequence. By means of the double independent channels, different characteristics of the meteorological data and the power can be respectively focused on and mined, and adaptive feature extraction is realized, so that the prediction accuracy of the power is improved, and the stable operation of the photovoltaic power station is ensured.

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 based on multi-source heterogeneous data. Background Technology

[0002] When predicting future photovoltaic power, it is possible to capture the changing patterns of photovoltaic power to improve the accuracy of power prediction, thereby providing reliable data support for grid dispatch and photovoltaic power plant operation and maintenance, thus optimizing the efficiency of photovoltaic energy utilization and ensuring the stable operation of photovoltaic power plants.

[0003] However, when traditional technologies predict photovoltaic power, they usually mix meteorological data with historical photovoltaic power data using a single channel, without considering the essential differences between meteorological data and photovoltaic power data. For example, meteorological data is not time-series and usually has a slow-changing characteristic; photovoltaic power data is time-series and fluctuates greatly over time, usually requiring identification and processing of its fluctuation trend over time.

[0004] Therefore, when traditional technologies use a single channel to process the same set of feature extraction logic, they ignore the differentiated characteristics of heterogeneous data sources. This leads to the model being unable to adapt its feature learning and extraction to the different characteristics of meteorological data and photovoltaic power data. Consequently, when making photovoltaic power predictions, the model cannot fully exploit the different characteristics of the two types of heterogeneous data, resulting in low prediction accuracy. This makes it difficult to adapt to the photovoltaic prediction needs under complex meteorological conditions, and it cannot provide reliable prediction basis for grid dispatch and photovoltaic power plant operation and maintenance. As a result, it affects the utilization efficiency of photovoltaic energy and the safe and stable operation of the power grid. Summary of the Invention

[0005] The main purpose of this application is to propose a photovoltaic power prediction method and device based on multi-source heterogeneous data. By designing a dual-channel independent extraction of heterogeneous data sources, it is possible to learn and extract the adaptability features of meteorological data and photovoltaic power data, fully explore the different features of the two types of heterogeneous data, and effectively improve the accuracy of photovoltaic power prediction.

[0006] To achieve the above objectives, one aspect of this application proposes a photovoltaic power prediction method based on multi-source heterogeneous data, the method comprising: Obtain current meteorological data at the current time step and photovoltaic power sequence data to be input; wherein, the photovoltaic power sequence data includes several historical time steps; each historical time step corresponds to a historical photovoltaic power and a historical meteorological data. The current meteorological data and the photovoltaic power sequence data are input into the photovoltaic power prediction model so that the meteorological feature extraction channel in the photovoltaic power prediction model can extract features from the current meteorological data and the historical meteorological data corresponding to each historical time step, and output the current meteorological features and historical meteorological time sequence features. The power feature extraction channel in the photovoltaic power prediction model extracts features from the historical photovoltaic power corresponding to each historical time step, and outputs power time series features to characterize the fluctuation trend of power. The photovoltaic power prediction model fuses the current meteorological features, the historical meteorological time series features, and the power time series features through its feature fusion channel, and outputs a fused feature sequence. The photovoltaic power prediction model outputs the target photovoltaic power corresponding to the future time step based on the fusion feature sequence.

[0007] Furthermore, the training process of the photovoltaic power prediction model includes... A meteorological data sample with the same time step and a photovoltaic power sequence data sample to be trained are used as training samples; wherein, the photovoltaic power sequence data sample includes several historical time steps; each historical time step corresponds to a historical photovoltaic power sample and a historical meteorological data sample. Using each of the training samples as input and the actual photovoltaic power corresponding to each training sample as a supervision label, the photovoltaic power prediction model is iteratively trained until the error between the photovoltaic power prediction result output by the model and the supervision label meets a preset threshold, thereby generating the photovoltaic power prediction model.

[0008] Furthermore, the historical meteorological time series features include: a first convolutional feature and a second convolutional feature; The meteorological feature extraction channel in the photovoltaic power prediction model extracts features from the current meteorological data and the historical meteorological data corresponding to each historical time step, and outputs the current meteorological features and historical meteorological time series features, including: The current meteorological data is convolved through the meteorological feature extraction channel to capture the spatial distribution characteristics of the current meteorological data at the current time step and output the current meteorological features. The historical meteorological data is convolved to capture the spatial distribution characteristics of the historical meteorological data at each historical time step, and the first convolution feature is output; wherein, the first convolution feature is used to characterize the state of historical meteorology at a single time step. The current meteorological features and the first convolutional features are convolved to output short-distance meteorological correlation features corresponding to short time steps and long-distance meteorological correlation features corresponding to long time steps; wherein, the short-distance meteorological correlation features are used to characterize the short-term meteorological change trend between adjacent historical time steps; the long-distance meteorological correlation features are used to characterize the long-term meteorological evolution trend across multiple historical time steps. The short-range meteorological correlation features and the long-range meteorological correlation features are fused to output a second convolutional feature that characterizes the cross-time period correlation pattern between historical meteorological data and current meteorological data.

[0009] Furthermore, the current meteorological data includes: current solar irradiance, current ambient temperature, current ambient humidity, and current cloud cover; the meteorological feature extraction channel includes: causal convolutional layers and dilated convolutional layers; The step of convolving the current meteorological data through the meteorological feature extraction channel to capture the spatial distribution characteristics of the current meteorological data at the current time step and outputting the current meteorological features includes: The causal convolutional layer captures the spatial distribution characteristics of the current solar irradiance, current ambient temperature, current ambient humidity, and current cloud cover data at the current time step, and the dilated convolutional layer performs dilated convolution on each of the spatial distribution characteristics to output the current meteorological features.

[0010] Furthermore, the historical meteorological data includes: historical solar irradiance, historical ambient temperature, historical ambient humidity, and historical cloud cover data; the first convolutional feature includes: historical irradiance features, historical ambient temperature features, historical ambient humidity features, and historical cloud cover features; the dilated convolutional layer corresponds to a dilation factor. The step of convolving the current meteorological features and the first convolutional features to output short-range meteorological correlation features corresponding to short time steps and long-range meteorological correlation features corresponding to long time steps includes: The expansion factor is adjusted exponentially based on each historical time step and a preset multiple to generate the adjusted expansion factor. The receptive field of the current meteorological features, historical irradiance features, and historical cloud cover features is expanded by the expanded convolutional layer and the adjusted expansion factor, respectively, to extract irradiance trend features and cloud cover change trend features. Then, the irradiance trend features and cloud cover change trend features are output as short-range meteorological correlation features. The receptive field of the historical ambient temperature features and the historical ambient humidity features is expanded by the expanded convolutional layer and the adjusted expansion factor, respectively, and the temperature and humidity coupling features used to characterize the cumulative impact of temperature and humidity changes on photovoltaic modules are extracted. Then, the temperature and humidity coupling features are output as the long-distance meteorological correlation features.

[0011] Furthermore, the power timing characteristics include: a first power timing characteristic and a second power timing characteristic; The process of extracting features from the historical photovoltaic power corresponding to each historical time step using the power feature extraction channel in the photovoltaic power prediction model, and outputting power time-series features to characterize the power fluctuation trend, includes: Through the power feature extraction channel, the long-term dependence of each of the historical photovoltaic power is extracted based on the positive time sequence, and the first power time series feature is extracted to characterize the long-term evolution trend of power over time. Through the power feature extraction channel, the short-term dependence of each of the historical photovoltaic power is extracted based on the reverse time order, and a second power time series feature is generated to characterize the short-period fluctuation trend of power over time.

[0012] Furthermore, the feature fusion channel of the photovoltaic power prediction model fuses the current meteorological features, the historical meteorological time-series features, and the power time-series features to output a fused feature sequence, including: The current meteorological features, the first convolutional features, the second convolutional features, the first power time series features, and the second power time series features are fused through the feature fusion channel to generate a first fusion feature for characterizing the trend of photovoltaic power change with irradiance, a second fusion feature for characterizing the trend of photovoltaic power change with cloud cover, a third fusion feature for characterizing the degree of influence of ambient temperature on the extreme value of photovoltaic power, and a fourth fusion feature for characterizing the degree of influence of ambient humidity on the ramp-up of photovoltaic power. Global average pooling is performed on the first fusion feature, the second fusion feature, the third fusion feature, and the fourth fusion feature in the time dimension to generate a first feature distribution that reflects the global information of the meteorological feature extraction channel and a second feature distribution that reflects the global information of the power feature extraction channel. Based on the gating mechanism of the feature fusion channel, the dependency relationship between the first feature distribution and the second feature distribution is captured, and a first weight of the meteorological feature extraction channel and a second weight of the power feature extraction channel are generated; wherein, the first weight is used to measure the degree of influence of meteorological factors on power; and the second weight is used to measure the degree of influence of historical power on power. Based on the first weight and the second weight, the current meteorological feature, the first convolutional feature, the second convolutional feature, the first power time series feature, and the second power time series feature are summed element by element to generate a fused feature sequence.

[0013] Furthermore, the photovoltaic power prediction model includes: a probabilistic modeling network; The step of the photovoltaic power prediction model outputting the target photovoltaic power corresponding to the future time step based on the fused feature sequence includes: The probabilistic modeling network captures long-term temporal dependencies in the fused feature sequence to generate a latent variable sequence; wherein, the latent variable sequence corresponds to a hidden state at several time points, and each hidden state is used to indicate the temporal evolution law of photovoltaic power under the combined influence of meteorological and historical power. The probabilistic modeling network outputs the Gaussian mixture distribution parameters corresponding to the hidden variable sequence based on each hidden state; wherein, the Gaussian mixture distribution parameters include: several Gaussian distributions; each Gaussian distribution includes: Gaussian distribution weights, Gaussian distribution mean, and Gaussian distribution positive variance; Based on the mean of each Gaussian distribution, and the Gaussian distribution weight and positive variance corresponding to each Gaussian distribution mean, a weighted sum of Gaussian distributions is calculated, and the weighted sum of Gaussian distributions is used as the predicted value of the target photovoltaic power corresponding to the future time step.

[0014] Furthermore, the method also includes: Each time the photovoltaic power prediction model is trained, its parameters are updated according to the following loss function: ; ; in, The value of the loss function. The negative log-likelihood of the probabilistic modeling network is used to measure the likelihood matching between the Gaussian distribution output by the probabilistic modeling network and the actual value of photovoltaic power. Mean squared error loss is used to indicate the error between the photovoltaic power prediction output by the model and the actual photovoltaic power. for The corresponding loss balance weight, for The corresponding loss balance weight.

[0015] To achieve the above objectives, another aspect of this application proposes a photovoltaic power prediction device based on multi-source heterogeneous data, the device comprising: The data acquisition module is used to acquire current meteorological data at the current time step and photovoltaic power sequence data to be input; wherein, the photovoltaic power sequence data includes several historical time steps; each historical time step corresponds to a historical photovoltaic power and a historical meteorological data. The photovoltaic power prediction module is used to input the current meteorological data and the photovoltaic power sequence data into the photovoltaic power prediction model. The meteorological feature extraction channel in the photovoltaic power prediction model extracts features from the current meteorological data and the historical meteorological data corresponding to each historical time step, outputting current meteorological features and historical meteorological time-series features. The power feature extraction channel in the photovoltaic power prediction model extracts features from the historical photovoltaic power corresponding to each historical time step, outputting power time-series features characterizing the power fluctuation trend. The feature fusion channel in the photovoltaic power prediction model fuses the current meteorological features, the historical meteorological time-series features, and the power time-series features, outputting a fused feature sequence. Based on the fused feature sequence, the photovoltaic power prediction model outputs the target photovoltaic power corresponding to the future time step.

[0016] The embodiments of this application include at least the following beneficial effects: This application provides a photovoltaic power prediction method and apparatus based on multi-source heterogeneous data. The method acquires current meteorological data at the current time step and photovoltaic power sequence data containing several historical time steps, specifying that each historical time step corresponds to a historical photovoltaic power and a historical meteorological data point. Simultaneously, the current meteorological data and photovoltaic power sequence data are input into a photovoltaic power prediction model, which processes the heterogeneous data sources separately through its built-in independent feature extraction channels. Specifically, through the meteorological feature extraction channel in the model, features are extracted only from the current meteorological data and the historical meteorological data corresponding to each historical time step, outputting current meteorological features and historical meteorological time-series features. Through the power feature extraction channel in the model, features are extracted from each historical photovoltaic power in the photovoltaic power sequence data, outputting only the power time-series features used to characterize the power fluctuation trend. Finally, through the model's feature fusion channel, the current meteorological features, historical meteorological time-series features, and power time-series features are fused, outputting a fused feature sequence. The photovoltaic power prediction model then outputs the target photovoltaic power corresponding to the future time step based on this fused feature sequence. Unlike traditional technologies, this invention features an independent meteorological feature extraction channel that focuses on mining the gradual changes in meteorological data, unaffected by the instantaneous fluctuations in photovoltaic power data, ensuring the purity and specificity of meteorological features. It also features an independent power feature extraction channel that focuses on capturing the temporal fluctuation trends of photovoltaic power data, unaffected by the gradual changes in meteorological data, thus enabling precise mining of the inherent variation patterns of power. Therefore, this invention, through its dual-channel independent extraction design of heterogeneous data sources, achieves adaptive feature learning and extraction for both meteorological and photovoltaic power data, fully mining the different characteristics of these two types of heterogeneous data, effectively improving the accuracy of photovoltaic power prediction, adapting to the photovoltaic prediction needs under complex meteorological conditions, providing reliable prediction basis for grid dispatch and photovoltaic power plant operation and maintenance, thereby optimizing photovoltaic energy utilization efficiency and ensuring the safe and stable operation of photovoltaic power plants and the power grid. Attached Figure Description

[0017] Figure 1 This is a schematic flowchart of a photovoltaic power prediction method based on multi-source heterogeneous data provided in an embodiment of this application; Figure 2 This is a schematic diagram of the prediction method based on an end-to-end prediction framework provided in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of a photovoltaic power prediction device based on multi-source heterogeneous data provided in an embodiment of this application; Figure 4 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0019] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various concepts, but unless otherwise stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the words “if,” “when,” or “in response to a determination” as used herein may be interpreted as “when…” or “when…” or “in response to a determination.”

[0020] As used in this application, the terms "several", "each", etc., "several" include one, two or more, "each" refers to each of the corresponding plurality, and "any" refers to any one of the plurality.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0022] Traditional photovoltaic (PV) power forecasting technologies often employ a single channel to process meteorological data and historical PV power data together, without considering their fundamental differences. For example, meteorological data lacks strong temporal sequence and exhibits gradual changes, while historical PV power data exhibits strong temporal sequence and fluctuates significantly over time, requiring precise identification of its fluctuation trends. Therefore, because traditional technologies ignore the differentiated characteristics of heterogeneous data sources and use the same feature extraction logic for mixed processing, the models cannot adapt to the different characteristics of meteorological and PV power data in terms of feature learning and extraction. Ultimately, the inability to fully exploit the core features of these two types of heterogeneous data results in low PV power forecasting accuracy, making it difficult to adapt to complex meteorological conditions. This fails to provide reliable data for grid dispatching and PV power plant operation and maintenance, and also affects the efficiency of PV energy utilization and the safe and stable operation of the power grid.

[0023] In view of this, this application provides a photovoltaic power prediction method and device based on multi-source heterogeneous data. This scheme achieves adaptive feature learning and extraction of meteorological data and photovoltaic power data through the design of dual-channel independent extraction of heterogeneous data sources. It can independently and fully explore the core change patterns of the two types of heterogeneous data, effectively improve the accuracy of photovoltaic power prediction, adapt to the prediction needs under complex meteorological conditions, provide a reliable basis for grid dispatch and photovoltaic power plant operation and maintenance, thereby optimizing the utilization efficiency of photovoltaic energy and ensuring the safe and stable operation of photovoltaic power plants and the grid.

[0024] Figure 1 This is an optional flowchart of a photovoltaic power prediction method based on multi-source heterogeneous data provided in an embodiment of this application. Figure 1 The method may include, but is not limited to, steps S1 to S2: Step S1: Obtain the current meteorological data at the current time step and the photovoltaic power sequence data to be input; wherein, the photovoltaic power sequence data includes several historical time steps; each historical time step corresponds to a historical photovoltaic power and a historical meteorological data. Step S2: Input the current meteorological data and the photovoltaic power sequence data into the photovoltaic power prediction model, so that the meteorological feature extraction channel in the photovoltaic power prediction model can extract features from the current meteorological data and the historical meteorological data corresponding to each historical time step, and output the current meteorological features and historical meteorological time sequence features. The power feature extraction channel in the photovoltaic power prediction model extracts features from the historical photovoltaic power corresponding to each historical time step, and outputs power time series features to characterize the fluctuation trend of power. The photovoltaic power prediction model fuses the current meteorological features, the historical meteorological time series features, and the power time series features through its feature fusion channel, and outputs a fused feature sequence. The photovoltaic power prediction model outputs the target photovoltaic power corresponding to the future time step based on the fusion feature sequence.

[0025] Steps S1 to S2 as shown in the embodiments of this application split the two heterogeneous data sources, meteorological data and photovoltaic power data, into two independent feature extraction channels for separate processing. After fusion, the target photovoltaic power for the future time step is predicted. This realizes the differentiation and adaptive feature learning of multi-source heterogeneous data, and solves the feature confusion problem caused by the traditional method of mixing meteorological and power data in the same channel. The present invention can ultimately output a more accurate and stable future photovoltaic power.

[0026] In this embodiment of the invention, meteorological data and historical photovoltaic power data in photovoltaic power sequence data are completely different heterogeneous data. Therefore, when current meteorological data and photovoltaic power sequence data are input into the photovoltaic power prediction model, the meteorological feature extraction channel in the photovoltaic power prediction model focuses only on processing meteorological data and learns the changing patterns of meteorology itself, thereby independently extracting current meteorological features and historical meteorological time series features. The power feature extraction channel in the photovoltaic power prediction model focuses only on processing power sequence data and can learn the fluctuation and trend patterns of power itself, thereby independently extracting power time series features used to characterize the fluctuation trend of power. The above two channels do not interfere with each other and each is adapted to the characteristics of the corresponding data source.

[0027] Therefore, this invention extracts features from multi-source heterogeneous data through dual independent channels of meteorology and power, realizing differentiated learning of time series data with different characteristics. This avoids feature interference and pattern confusion between heterogeneous data sources, improves the purity and accuracy of feature extraction, and thus makes the subsequent output of future photovoltaic power prediction results more accurate. This effectively improves the accuracy of photovoltaic power prediction and can meet the photovoltaic prediction needs under complex meteorological conditions.

[0028] Step S1, as the data preparation stage of the prediction process, mainly involves acquiring complete and valid model input data. In some embodiments, the scope and dimensions of data collection can be clearly defined first. For example, one time step is one moment, and the current meteorological data refers to the real-time meteorological parameters collected at the current moment. Furthermore, the photovoltaic power sequence data to be input is the measured power data of photovoltaic power plants collected in historical moments. The photovoltaic power sequence data to be input is divided into multiple historical time steps, and each historical time step is synchronously bound to the historical photovoltaic power and historical meteorological data at the corresponding moment to ensure the one-to-one correspondence of the time series data.

[0029] Optionally, the time step is the smallest time unit for data collection and prediction in photovoltaic power plants. It can be set by the actual collection frequency of the power plant, such as 1 minute, 15 minutes, 30 minutes, or 1 hour, representing the time interval of data recording and serving as the basic unit for constructing time series. The current time step refers to the reference moment for photovoltaic power prediction. The model must predict the photovoltaic power at future moments after this moment based on this moment and historical data. Example: If the time step is 15 minutes, and this embodiment of the invention aims to predict the photovoltaic power at 12:15 on March 19, 2025, then 12:00 on March 19, 2025 is the current time step. The real-time meteorological data collected at this moment can also indicate weather forecast data, i.e., weather data from 12:00 on March 19, 2025 to 12:15 on March 19, 2025, such as solar irradiance, ambient temperature, ambient humidity, and cloud cover. This is the input for the model to perceive the current meteorological state and can be used to deduce the driving effect of future weather on power.

[0030] Historical time steps refer to the continuous historical moments that have been collected up to the current time step. They serve as the data source for the model to learn the historical patterns of weather and power changes. Example: If the current time step is 2025-03-19 09:00, the historical time step can be set to the previous 24 consecutive 15-minute steps, i.e., 2025-03-19 11:45, 11:30, 11:15…08:00. Each time point is an independent historical time step; these historical time steps together constitute a historical time series, allowing the model to learn the long-term / short-term patterns of weather changes and power fluctuations, such as the diurnal cycle of rising and falling irradiance and the response pattern of power to sudden changes in cloud cover.

[0031] Furthermore, each historical time step corresponds to a historical photovoltaic power and a historical meteorological data point. This means that each historical time step is strictly bound to two sets of synchronously collected data, achieving precise temporal alignment between meteorological data and photovoltaic power data. This is a prerequisite for ensuring that the model learns the true physical laws.

[0032] For any given historical time step (e.g., 2025-03-19 11:45), two sets of data will be collected simultaneously: Historical photovoltaic power: The actual power output of the photovoltaic power station at that moment, such as 480kW; Historical meteorological data: Meteorological factor data collected synchronously at this time, such as historical solar irradiance = 720W / m², historical ambient temperature = 25℃, historical ambient humidity = 40%, and historical cloud cover = 20%. In some embodiments of the present invention, a sliding window technique may be employed to construct time-dependent time-series samples. The window length is set to... If the step size is s, then the continuous time series can be divided into several subsequences, which can be represented as follows: ; In the formula, This represents continuous data within a window starting from time t.

[0033] As an illustration, the segmentation of the time window described above not only ensures the continuity of the input samples but also fully captures historical information and time dependencies, providing sufficient data support for subsequent model training.

[0034] In some embodiments, the present invention can also complete missing data in historical photovoltaic power sequences, resulting in: Each of the aforementioned historical photovoltaic power includes: a power value or a preset identifier; wherein the preset identifier is used to indicate that the historical photovoltaic power value is missing; Therefore, inputting the current meteorological data and the photovoltaic power sequence data into the photovoltaic power prediction model includes: When it is determined that the preset identifier does not exist in the photovoltaic power sequence data, the current meteorological data and the photovoltaic power sequence data are input into the photovoltaic power prediction model; When it is determined that the preset identifier exists in the photovoltaic power sequence data, the historical time step corresponding to each preset identifier is used as the historical time step to be filled. For each historical time step to be filled, several adjacent time steps are marked in the photovoltaic power sequence data; wherein, the time interval between the adjacent time step and the historical time step to be filled is less than a preset time threshold, and the historical photovoltaic power of the adjacent time step corresponds to a power value. For each of the historical time steps to be filled, the average power value is calculated based on the power value corresponding to each of the adjacent time steps and the total number of adjacent time steps, and the average power value is replaced with the preset identifier corresponding to the historical time step to be filled. After determining that all the historical time steps to be filled have been replaced, the current meteorological data and the photovoltaic power sequence data that have been replaced are input into the photovoltaic power prediction model.

[0035] In illustratively, in this embodiment of the invention, the missing historical photovoltaic power is preprocessed by filling in the average of adjacent effective time steps. This can repair the power data missing problem caused by sensor failure, communication interruption and other scenarios without destroying the time series data structure or introducing additional errors, thus ensuring the integrity and availability of the model input from the data source.

[0036] Because photovoltaic power has a strong time-series dependence, directly removing missing data would cause historical time series breaks, disrupting the power's own fluctuation trend and evolution. This invention retains all historical time steps through mean filling, maintaining the integrity of the time series structure and providing a reliable data foundation for subsequent power feature extraction channels to accurately capture power fluctuation trends. Specifically, this invention only selects effective adjacent power values ​​with a close interval to the time step to be filled to calculate the mean. It fully utilizes the characteristics of photovoltaic power changing slowly and having local continuity in a short period of time, making the filling result closer to the true power value. This avoids unreasonable data introduced by traditional simple interpolation or random filling methods, ensuring the authenticity of the data to the greatest extent.

[0037] Since each historical time step is bound to the corresponding historical meteorological data and historical photovoltaic power, if the missing power is not filled in, it will cause the meteorological and power data to be misaligned, which will destroy the temporal matching during dual-channel feature extraction. The embodiments of the present invention can ensure that the meteorological and power data always correspond one-to-one by filling in the missing power, which supports the subsequent independent and accurate extraction of features by dual channels.

[0038] As an illustration, in more complex or sparse data structures, this invention can use the K-Nearest Neighbors (KNN) algorithm to impute missing values ​​for missing data. If we find the K nearest neighbors in the sample and fill in the gaps using the average of these K data values, then we have: ; in, The power average can be used as replacement data for the historical time step to be filled, which will yield... Replace with the preset identifier corresponding to the historical time step to be filled. For the sample and missing data Power values ​​corresponding to adjacent time steps, illustratively. K represents the power value corresponding to the historical photovoltaic power at adjacent time steps, and K is the total number of adjacent time steps.

[0039] In some embodiments, before inputting the current meteorological data and the photovoltaic power sequence data into the photovoltaic power prediction model, the present invention may also perform anomaly detection, normalization, missing data imputation and sample construction processing.

[0040] For anomaly detection, embodiments of this invention can use current meteorological data and photovoltaic power sequence data as input data. Statistical methods are employed to clean the input data, and abnormal data that does not conform to the expected range is removed by setting reasonable upper and lower thresholds. To ensure that the data samples possess good statistical characteristics, this invention uses 3... σThe principle for detecting outliers is based on the following criteria: ; In the formula, μ is the data mean, σ is the standard deviation, and x is the data to be input.

[0041] Indicative, 3 σ The principle refers to an outlier detection method based on the statistical properties of normal distribution.

[0042] When the value of the input data x exceeds If the data is abnormal, it will be removed.

[0043] Furthermore, for normalization, to eliminate the influence of different units and numerical ranges on model training, this invention can also use the Z-score normalization method to standardize each data point, the expression of which is: ; It is understandable that the above formula can convert each data value of the input data x into a standard normal distribution value with a mean of zero and a standard deviation of one. As an illustration, Z-score normalization, also known as standardization, is a preprocessing method that linearly transforms raw data into a standard normal distribution.

[0044] Furthermore, for missing parts in the data, the present invention can also use a linear interpolation method to fill in the missing parts. For the missing part in the input data sequence x, the present invention can use a linear interpolation method to fill in the missing parts. Missing data The following formula can be used for estimation: ; In the formula, and For missing data Known data at an adjacent moment or time step.

[0045] This invention enables the standardization and integrity processing of raw data, solving core problems such as missing, abnormal, inconsistent units, or irregular time sequences of raw data. It provides high-quality data support for dual-channel differentiated feature learning, further ensuring the accuracy and stability of subsequent feature extraction, fusion, and photovoltaic power prediction. This makes the model more adaptable to the complex data scenarios in the actual operation of photovoltaic power plants, and gives it stronger practicality and generalization ability.

[0046] For step S2, in some embodiments, the training process of the photovoltaic power prediction model includes... A meteorological data sample with the same time step and a photovoltaic power sequence data sample to be trained are used as training samples; wherein, the photovoltaic power sequence data sample includes several historical time steps; each historical time step corresponds to a historical photovoltaic power sample and a historical meteorological data sample. Using each of the training samples as input and the actual photovoltaic power corresponding to each training sample as a supervision label, the photovoltaic power prediction model is iteratively trained until the error between the photovoltaic power prediction result output by the model and the supervision label meets a preset threshold, thereby generating the photovoltaic power prediction model.

[0047] To illustrate, before conducting actual photovoltaic power prediction based on the photovoltaic power prediction model, the photovoltaic power prediction model needs to be iteratively trained. This allows the model to learn the mapping rules of meteorological feature extraction, power feature extraction, multi-feature fusion, and future power prediction through labeled samples, so that the deviation between the predicted value output by the model and the actual photovoltaic power value reaches the preset requirements.

[0048] First, standardized labeled samples are constructed for model training. Each sample contains two parts: model input and supervision label. The model input consists of meteorological data samples and photovoltaic power sequence data samples consistent with actual predictions. The supervision label is the actual photovoltaic power value at the corresponding future time step. The constructed labeled samples are divided into training set and validation set according to a preset ratio. The training set is used for iterative updates of model parameters, and the validation set is used for performance evaluation after each round of training, ensuring that the datasets do not overlap.

[0049] Next, the parameters of the photovoltaic power prediction model modules are initialized, including the meteorological feature extraction channel, the power feature extraction channel, the feature fusion channel, and the prediction module in the model responsible for outputting the target photovoltaic power based on the fused feature sequence. Initial values ​​are assigned to the learnable parameters of each channel / module, and the network hyperparameters are set to ensure that each module can complete the forward calculation process of feature extraction, fusion, and power prediction according to the preset logic.

[0050] The labeled samples in the training set are grouped according to a preset batch size and input into the photovoltaic power prediction model to be trained batch by batch. The entire forward computation is completed according to the actual prediction process. Specifically, the meteorological feature extraction channel extracts features from the meteorological data samples in the samples and the historical meteorological data corresponding to the historical time step of each sample, and outputs meteorological feature samples and historical meteorological time series feature samples. The power feature extraction channel extracts features from the photovoltaic power sequence data samples in the training samples and outputs power time series feature samples that represent the power fluctuation trend. The feature fusion channel fuses the above meteorological feature samples, historical meteorological time series feature samples, and power time series feature samples and outputs fused feature sequence samples. Based on the fused feature sequence samples, the model outputs the photovoltaic power prediction value for the future time step corresponding to the batch of samples.

[0051] Using the supervised label of the samples, i.e. the actual photovoltaic power value at the future time step, the photovoltaic power prediction value output by the model is compared with the actual value. The loss value of the batch of samples is calculated by using a preset loss function, such as root mean square error or mean absolute error. The magnitude of the loss value is used to quantify the degree of deviation between the model prediction value and the actual value. The smaller the loss value, the smaller the prediction deviation.

[0052] Finally, based on the calculated loss value, the gradient values ​​of the learnable parameters of each module of the model are calculated using the gradient descent algorithm. Gradient backpropagation is performed in the order of prediction module, feature fusion channel, power feature extraction channel, and meteorological feature extraction channel. Based on the gradient values, the parameters of each channel / module are adjusted and updated in a targeted manner, so that the model can output prediction results that are closer to the actual values ​​in the next round of training, thereby gradually reducing the loss value.

[0053] After each full iteration of the training set, the validation set samples are completely input into the current model to be trained. The photovoltaic power prediction values ​​of the validation set are output according to the above forward calculation process, and the loss value and prediction accuracy of the validation set are calculated. If the validation set loss value continues to decrease and the prediction accuracy steadily improves, it indicates that the model has good generalization ability, and the next round of training continues. If the validation set loss value increases and the prediction accuracy decreases, it indicates that the model is overfitting, and the model hyperparameters are adjusted in time to suppress overfitting.

[0054] Repeat the above batch iteration process of forward computation, loss calculation, parameter update, and validation set evaluation. When any of the following conditions are met, the model training is considered to have converged and the iteration is stopped: the loss values ​​of both the training set and the validation set tend to stabilize for several consecutive rounds, and the prediction accuracy reaches the preset index; the model completes the preset number of training rounds, and the prediction accuracy of the validation set meets the actual application requirements.

[0055] The parameters of the model after training convergence are saved. All optimized parameters of the meteorological feature extraction channel, power feature extraction channel, feature fusion channel and prediction module are saved in a unified manner to form a photovoltaic power prediction model that can be directly deployed for predicting future photovoltaic power in real-world scenarios.

[0056] In some embodiments, the trained photovoltaic power prediction model undergoes the following refined extraction process when actually outputting current meteorological characteristics and historical meteorological time series characteristics.

[0057] The historical meteorological time series features actually include: a first convolutional feature and a second convolutional feature; Therefore, the meteorological feature extraction channel in the photovoltaic power prediction model extracts features from the current meteorological data and the historical meteorological data corresponding to each historical time step, and outputs the current meteorological features and historical meteorological time series features. The specific process includes: The current meteorological data is convolved through the meteorological feature extraction channel to capture the spatial distribution characteristics of the current meteorological data at the current time step and output the current meteorological features. The historical meteorological data is convolved to capture the spatial distribution characteristics of the historical meteorological data at each historical time step, and the first convolution feature is output; wherein, the first convolution feature is used to characterize the state of historical meteorology at a single time step. The current meteorological features and the first convolutional features are convolved to output short-distance meteorological correlation features corresponding to short time steps and long-distance meteorological correlation features corresponding to long time steps; wherein, the short-distance meteorological correlation features are used to characterize the short-term meteorological change trend between adjacent historical time steps; the long-distance meteorological correlation features are used to characterize the long-term meteorological evolution trend across multiple historical time steps. The short-range meteorological correlation features and the long-range meteorological correlation features are fused to output a second convolutional feature that characterizes the cross-time period correlation pattern between historical meteorological data and current meteorological data.

[0058] It is understood that, in this embodiment of the invention, the convolutional layer of the meteorological feature extraction channel performs spatial dimension convolution calculation on the current meteorological data, captures the spatial distribution characteristics of each meteorological factor in the current meteorological data at the current time step and the correlation characteristics between factors, and outputs the current meteorological features with regular dimensions and strong feature representation. The convolutional layer of the meteorological feature extraction channel performs convolutional processing on the historical meteorological data corresponding to each historical time step step, respectively capturing the spatial distribution characteristics of the meteorological data at each historical time step step, and integrating and outputting the first convolutional feature; wherein, the first convolutional feature is used to accurately characterize the instantaneous state of historical meteorology at a single time step, and completely preserve the original feature information of meteorology at each historical moment. The current meteorological features are fused with the first convolutional features through convolution calculation. By adapting the convolution logic to different time scales, short-term and long-term meteorological time series correlation patterns are mined, and then the short-distance meteorological correlation features corresponding to the short time step and the long-distance meteorological correlation features corresponding to the long time step are output. Schematic, the short-distance meteorological correlation features are used to characterize the short-term meteorological change trend between adjacent historical time steps, which can focus on the short-term fluctuation characteristics of meteorological factors; while the long-distance meteorological correlation features are used to characterize the long-term meteorological evolution trend over multiple historical time steps, which can focus on the long-term accumulation and gradual change characteristics of meteorological factors. By dimensional alignment and feature fusion of the short-distance and long-distance meteorological correlation features, and by integrating the short-term fluctuations and long-term evolution patterns of meteorology through element-by-element fusion, a second convolutional feature is output to characterize the cross-time-period correlation pattern between historical meteorological data and current meteorological data. This second convolutional feature can then be used to completely reconstruct the temporal evolution link of meteorology from history to the present.

[0059] It is understandable that meteorological data is an external driving factor for photovoltaic power output. It includes instantaneous states at a single time step, such as irradiance and cloud cover at a certain moment, as well as short-term fluctuations between adjacent moments, such as a sudden increase in cloud cover or a rapid decrease in irradiance. It also possesses long-term evolution over multiple moments, such as a continuous rise in ambient temperature and gradual changes in humidity. Simultaneously, changes in historical meteorology gradually transmit to the current meteorology, forming cross-period evolutionary relationships. This invention extracts current meteorological features, first convolutional features, and second convolutional features in a hierarchical manner, corresponding to the instantaneous state of meteorology, a single-step historical state, and the cross-period evolutionary patterns between history and the present, respectively. This closely matches the actual physical change characteristics of meteorological data, achieving a comprehensive and thorough characterization of meteorological information. It avoids the problem that traditional single-feature extraction cannot fully preserve meteorological patterns, allowing the model to learn more realistic meteorological driving information.

[0060] Furthermore, the impact of different meteorological factors on photovoltaic power manifests at different time scales. Short-term abrupt changes in factors such as solar irradiance and cloud cover directly lead to sudden increases and decreases in photovoltaic power, which are the core causes of short-term power fluctuations. Long-term cumulative changes in factors such as ambient temperature and humidity gradually affect the power generation efficiency of photovoltaic modules, which are important reasons for slow power drift and ramp-up. This invention captures short-term fluctuations in irradiance and cloud cover by extracting short-distance meteorological correlation features, and captures long-term cumulative temperature and humidity by extracting long-distance meteorological correlation features. This allows meteorological features to accurately match the response patterns of photovoltaic power to different meteorological factors, enabling subsequent models to predict short-term fluctuations and long-term changes in power based on corresponding meteorological features, significantly improving the accuracy of power change prediction under complex meteorological conditions.

[0061] In a preferred embodiment, the meteorological feature extraction channel of the present invention uses a Temporal Convolutional Network (TCN) as the backbone network, wherein the temporal convolutional network includes causal convolutional layers and dilated convolutional layers. Therefore, when outputting the current meteorological feature, the first convolutional feature, and the second convolutional feature, processing is actually performed simultaneously through the causal convolutional layers and the dilated convolutional layers.

[0062] Optionally, the current meteorological data specifically includes: current solar irradiance, current ambient temperature, current ambient humidity, and current cloud cover data. These four parameters serve as core meteorological factors affecting photovoltaic power output. Historical meteorological data may specifically include: historical solar irradiance, historical ambient temperature, historical ambient humidity, and historical cloud cover data. The first convolutional feature includes: historical irradiance features, historical ambient temperature features, historical ambient humidity features, and historical cloud cover features. Furthermore, the dilated convolutional layer of this invention corresponds to a dilation factor.

[0063] Therefore, the step of convolving the current meteorological data through the meteorological feature extraction channel to capture the spatial distribution characteristics of the current meteorological data at the current time step and outputting the current meteorological features includes: The causal convolutional layer captures the spatial distribution characteristics of the current solar irradiance, current ambient temperature, current ambient humidity, and current cloud cover data at the current time step. The dilated convolutional layer then performs dilation convolution on each of these spatial distribution characteristics to output the current meteorological features. Optionally, when dilating the spatial distribution characteristics using the dilation convolutional layer, multiple time nodes can be added to learn and extract the causal relationships of temporal changes in meteorology, thereby deriving the final current meteorological features.

[0064] Furthermore, the step of convolving the current meteorological features and the first convolutional features to output short-range meteorological correlation features corresponding to short time steps and long-range meteorological correlation features corresponding to long time steps includes: The expansion factor is adjusted exponentially based on each historical time step and a preset multiple to generate the adjusted expansion factor. The receptive field of the current meteorological features, historical irradiance features, and historical cloud cover features is expanded by the expanded convolutional layer and the adjusted expansion factor, respectively, to extract irradiance trend features and cloud cover change trend features. Then, the irradiance trend features and cloud cover change trend features are output as short-range meteorological correlation features. The receptive field of the historical ambient temperature features and the historical ambient humidity features is expanded by the expanded convolutional layer and the adjusted expansion factor, respectively, and the temperature and humidity coupling features used to characterize the cumulative impact of temperature and humidity changes on photovoltaic modules are extracted. Then, the temperature and humidity coupling features are output as the long-distance meteorological correlation features.

[0065] In illustrative terms, photovoltaic power prediction is a time-series extrapolation task. The power at future moments can be derived from the meteorological and power data at current and historical moments. This invention, through the characteristics of causal convolutional layers, can provide the output at the current moment that depends only on the input at the current and historical moments. Through this layer, all meteorological data are processed synchronously to ensure that all extracted meteorological features conform to the true time-series causal laws. This allows the meteorological and power correlation law learned by the model to fit the actual physical mechanism, laying the foundation for accurate prediction of photovoltaic power in the future.

[0066] Furthermore, the temporal correlations of meteorological data exhibit short-term and long-term scale differences. It is necessary to capture both short-term fluctuations between adjacent time steps and to mine long-term evolution across multiple time steps. Traditional convolutional methods require increasing the number of network layers to expand the receptive field, which can easily lead to a surge in parameters and model overfitting. The dilated convolutional layer of this invention allows for exponential adjustment of the dilation factor by a preset factor. This achieves an exponential expansion of the receptive field without increasing the number of convolutional kernel parameters or deepening the network layers. It can capture short-term correlations of irradiance and cloud cover with a small dilation factor, and also capture long-term accumulations of temperature and humidity with a large dilation factor, thus adapting to the extraction needs of multi-scale meteorological features.

[0067] Therefore, through the simultaneous processing of the aforementioned causal convolutional layers and dilated convolutional layers, the integrated extraction of spatial distribution characteristics and temporal correlation characteristics can be achieved. The impact of meteorological data on photovoltaic power is reflected not only in the spatial distribution and magnitude of various meteorological factors at a single time step (e.g., high irradiance and low cloud cover at a certain moment correspond to high power), but also in the temporal trend of meteorological factors (e.g., a continuous increase in irradiance corresponds to a power ramp-up). The simultaneous processing of causal convolutional layers and dilated convolutional layers can simultaneously extract single-step spatial distribution characteristics and multi-step temporal correlation characteristics in a single convolution calculation, eliminating the need for separate step-by-step processing of the two types of characteristics. This simplifies the feature extraction process, reduces computational overhead, and allows for deep integration of spatial and temporal features, avoiding feature fragmentation caused by independent extraction. This ensures that the extracted meteorological features more comprehensively reflect the combined driving effect of meteorological factors on photovoltaic power.

[0068] Furthermore, in this embodiment of the invention, irradiance trend features and cloud cover change trend features are specifically extracted as short-range meteorological correlation features, and temperature and humidity coupling features are extracted as long-range meteorological correlation features, which are highly matched with the response characteristics of photovoltaic power. Specifically, from the perspective of the power generation physical mechanism of photovoltaic modules, solar irradiance and cloud cover are instantaneous and short-term influencing factors: a short-term sudden increase / decrease in irradiance will directly lead to a rapid change in photovoltaic power, and a short-term sudden change in cloud cover, such as a rapid passage of cloud clusters, will instantly block sunlight, causing a sudden drop in power; while ambient temperature and ambient humidity are cumulative and long-term influencing factors. Temperature does not directly determine the power, but sustained high temperatures will reduce the photoelectric conversion efficiency of photovoltaic modules, and humidity will work synergistically with temperature to exacerbate the heat loss of the modules. The effects of both need to be accumulated over multiple time steps to be manifested.

[0069] Therefore, this invention categorizes meteorological factors into short-range and long-range categories based on their impact scale and extracts corresponding features. This approach aligns with the physical response patterns of photovoltaic power to different meteorological factors, allowing the extracted meteorological features to accurately reflect the actual impact patterns of each factor on power. This provides the most physically relevant feature support for subsequent power prediction.

[0070] The short / long-distance meteorological correlation features extracted in this embodiment of the invention correspond one-to-one with the short-term fluctuation trend and long-term evolution trend of photovoltaic power, providing clear feature dimensions for the adaptive weighted fusion of subsequent feature fusion channels. Illustratively, during periods of clear skies and stability, photovoltaic power exhibits strong autoregressive properties, and the model can automatically reduce the weight of meteorological features, relying primarily on power features. Under conditions of cloud cover and sudden changes in irradiance, the model can automatically increase the weight of short-term irradiance and cloud cover features to accurately capture power fluctuations. Furthermore, under conditions of continuous accumulation of temperature and humidity, the model can automatically increase the weight of temperature and humidity coupled features to adapt to the slow drift of power, significantly improving the model's adaptability to different operating conditions.

[0071] In summary, this invention uses causal convolutional layers and dilated convolutional layers as the backbone network for meteorological feature extraction channels and processes them synchronously. At the same time, based on the differentiated influence mechanism of meteorological factors on photovoltaic power, it extracts short / long-distance meteorological correlation features at different scales. This ensures both the temporal rationality and multi-scale nature of meteorological feature extraction, and also ensures a high degree of matching between the extracted meteorological features and the physical response law of photovoltaic power. This provides meteorological-driven feature support for subsequent heterogeneous feature fusion and high-precision prediction of photovoltaic power.

[0072] In a preferred embodiment, the temporal convolutional network (TCN) can be calculated according to the following formula at the 1st... Output features of the layer at time step t : ; In the formula, For the first The k-th weight of the convolutional kernel, The kernel size is the convolution kernel size. For the first The expansion factor of the layer It is a two-dimensional matrix representing the meteorological time-series feature data fed into the TCN backbone network of the temporal convolutional network, and is the source input of the entire meteorological forecasting branch.

[0073] As the number of layers increases, the expansion factor grows exponentially, and the receptive field expands rapidly. Final output characteristics of the meteorological channel. for: ; In the formula, These are the characteristic dimensions of the meteorological channel.

[0074] Optionally, the backbone network of the meteorological feature extraction channel of this invention can also extract strong driving instantaneous fluctuation features to capture the direct mapping and instantaneous impact of irradiance and cloud cover on photovoltaic power. Utilizing the causal convolutional properties of TCN, power surges caused by cloud cover can be accurately identified.

[0075] The backbone network of the meteorological feature extraction channel in this invention can also extract the cumulative effect features of weakly driven environments to capture the indirect impact of temperature and humidity on the photoelectric conversion efficiency and performance loss of photovoltaic modules. The expanded receptive field is enlarged using the extended convolution of the TCN to extract the cumulative evolution trend of temperature and humidity over long periods.

[0076] The backbone network of the meteorological feature extraction channel of the present invention can also extract time-domain heterogeneous meteorological correlation features to extract the nonlinear coupling relationship of multidimensional meteorological factors on the same time section, and characterize the comprehensive evolution state of external driving factors under complex meteorological environment.

[0077] In some embodiments, the power timing characteristics include: a first power timing characteristic and a second power timing characteristic; The process of extracting features from the historical photovoltaic power corresponding to each historical time step using the power feature extraction channel in the photovoltaic power prediction model, and outputting power time-series features to characterize the power fluctuation trend, includes: Through the power feature extraction channel, the long-term dependence of each of the historical photovoltaic power is extracted based on the positive time sequence, and the first power time series feature is extracted to characterize the long-term evolution trend of power over time. Through the power feature extraction channel, the short-term dependence of each of the historical photovoltaic power is extracted based on the reverse time order, and a second power time series feature is generated to characterize the short-period fluctuation trend of power over time.

[0078] In illustrative terms, this embodiment of the invention employs a two-way time series feature mining strategy that extracts long-term dependencies in forward time sequence and extracts short-term dependencies in reverse time sequence. This strategy can capture long-term and short-term dependencies in strong time series data and, in accordance with the principle that photovoltaic power is affected by different meteorological factors, identify the long-term evolution trend of power over time and the short-term fluctuation trend of power over time.

[0079] Specifically, this invention, through bidirectional extraction in both forward and reverse order, can fully explore historical photovoltaic power sequences from two perspectives. The forward order perspective captures the long-term evolution from the past to the present, while the reverse order perspective captures the local correlation patterns from the present to the past. Compared with traditional unidirectional time series feature extraction, this invention can uncover more comprehensive and detailed power time series patterns, avoid the omission of some patterns by unidirectional extraction, and significantly improve the overall characterization ability of power features.

[0080] Meanwhile, the bidirectional feature extraction forms a precise heterogeneous complement to the short / long-distance meteorological correlation features of the previous meteorological feature extraction channel; the meteorological channel extracts short-term meteorological features of irradiance and cloud cover and long-term coupling features of temperature and humidity, while the power channel extracts long-term power trend features and short-term local power features. The two channels explore the patterns of different factors and different scales from both meteorological and power dimensions, so that the feature sequence after fusion can simultaneously possess the causal laws driven by meteorology and the temporal laws of power itself.

[0081] Optionally, the first power time series characteristic is a macroscopic time series inertia characteristic, due to the strong physical inertia and diurnal cycle regularity of photovoltaic power output (such as smooth ramps and drops). The forward extraction simulates the passage of natural time, aiming to capture the long-term evolution trend of power over time.

[0082] The second power time series characteristic is the micro-fluctuation backtracking characteristic, because photovoltaic power is often affected by instantaneous disturbances (such as the aftereffects of cloud cover). By extracting back the history from the current moment, it is possible to more accurately capture the short-term fluctuation residuals that are closest to the present and have the greatest impact on the future.

[0083] In a preferred embodiment, the present invention preferably uses a Bidirectional Long Short-Term Memory (BiLSTM) network as the backbone network of the power channel. BiLSTM consists of two independent hidden layers: a forward LSTM and a backward LSTM. These layers process the sequence from both directions, and the hidden states from both directions are concatenated as the final output. This ensures that the output at each time step simultaneously contains both historical and future information, thereby significantly enhancing the model's sensitivity to transient changes in the power sequence, such as local extrema and the start of a ramp-up.

[0084] The update process of a forward LSTM is as follows: ; In the formula, , , These are the input gate parameters, forget gate parameters, and output gate parameters, respectively. For memory units, It is in a positive hidden state; , , , The weights for the input gate, forget gate, output gate, and memory unit are parameters that are automatically learned during model training and are determined by the hidden state dimension of the previous time step and the input feature dimension of the current time step.

[0085] The inverse LSTM processes the sequence in reverse order to obtain the inverse hidden state. The final output is a concatenated vector of bidirectional hidden states. : ; Therefore, the final output characteristic map of the power channel is denoted as: Then we have: ; In the formula, For the characteristic dimension of the power channel, This refers to the power time series characteristic data fed into the BiLSTM network.

[0086] In this invention, the meteorological channel is based on causal convolution and dilated convolution, which are adapted to the temporal causal laws and multi-scale characteristics of meteorological data. It extracts differentiated correlation features of irradiance / cloud cover and temperature and humidity in short / long time intervals, which conforms to the physical response law of photovoltaics to different meteorological factors. The power channel is based on the bidirectional long short-term memory network BiLSTM, which captures the long and short-term dependencies of power sequences through bidirectional extraction, thus avoiding the feature mixing problem of traditional isomorphic processing.

[0087] In some embodiments, the feature fusion channel of the present invention achieves feature fusion through three steps: global pooling, gated weight allocation, and element-wise weighting. Then: The feature fusion channel of the photovoltaic power prediction model fuses the current meteorological features, the historical meteorological time-series features, and the power time-series features to output a fused feature sequence, including: The current meteorological features, the first convolutional features, the second convolutional features, the first power time series features, and the second power time series features are fused through the feature fusion channel to generate a first fusion feature for characterizing the trend of photovoltaic power change with irradiance, a second fusion feature for characterizing the trend of photovoltaic power change with cloud cover, a third fusion feature for characterizing the degree of influence of ambient temperature on the extreme value of photovoltaic power, and a fourth fusion feature for characterizing the degree of influence of ambient humidity on the ramp-up of photovoltaic power. Global average pooling is performed on the first fusion feature, the second fusion feature, the third fusion feature, and the fourth fusion feature in the time dimension to generate a first feature distribution that reflects the global information of the meteorological feature extraction channel and a second feature distribution that reflects the global information of the power feature extraction channel. Based on the gating mechanism of the feature fusion channel, the dependency relationship between the first feature distribution and the second feature distribution is captured, and a first weight of the meteorological feature extraction channel and a second weight of the power feature extraction channel are generated; wherein, the first weight is used to measure the degree of influence of meteorological factors on power; and the second weight is used to measure the degree of influence of historical power on power. Based on the first weight and the second weight, the current meteorological feature, the first convolutional feature, the second convolutional feature, the first power time series feature, and the second power time series feature are summed element by element to generate a fused feature sequence.

[0088] In a schematic embodiment of the present invention, the current meteorological feature, the first convolutional feature, and the second convolutional feature of the meteorological feature class are first fused with the first power time-series feature and the second power time-series feature of the power feature class to obtain fused features with multiple coupling relationships. Then, redundant information is compressed and the overall feature distribution is captured through global pooling. Finally, weight allocation and fine fusion are performed, which reduces computational overhead and ensures that the adaptive generation of weights is supported by global data, avoiding imbalance in the fusion ratio caused by local feature deviations.

[0089] This invention can also generate adaptive gating weights based on the gating mechanism of the feature fusion channel, enabling the model to be dynamically generated based on the global feature distribution of meteorological and power channels, and directly associated with the refined sub-features extracted in the preceding sequence. Compared with traditional fixed weights / manual weighting, its adaptability and rationality are greatly improved. It can be understood that the weights are determined by the global distribution of the currently input meteorological and power features, and can automatically adjust the influence ratio of meteorology / power according to actual operating conditions. For example, when clouds pass or irradiance changes abruptly, the first weight (meteorology) is automatically increased, allowing the model to focus on capturing short-term power fluctuations driven by meteorology; during periods of clear skies and stability, the second weight (power) is automatically increased, allowing the model to make stable predictions based on the autoregressive patterns of historical power; and when temperature and humidity accumulate, the weight ratio of long-term coupled features in meteorology can be specifically increased, thereby achieving on-demand allocation.

[0090] Optionally, when a meteorological sensor malfunctions or meteorological data is abnormal, the abnormal meteorological characteristics will cause its global distribution to deviate from the normal pattern. The gating mechanism of this invention will automatically reduce the first weight, allowing the model to rely on historical power characteristics for prediction, without the need for an additional fault handling module, thus improving the fault tolerance of the system.

[0091] Furthermore, in this embodiment of the invention, the specific meaning of element-wise weighted summation is that for all sub-features of the meteorological channel (such as the current meteorological feature, the first convolutional feature, and the second convolutional feature) and all sub-features of the power channel (such as the first power time-series feature and the second power time-series feature), at the same time step and the same feature dimension, they are multiplied by the adaptive weights of the corresponding channels, such as the current meteorological sub-feature × the first weight or the first time-series power sub-feature × the second weight. Then, the calculation results at each position are summed to finally generate a fused feature sequence with unified dimensions and continuous time sequence. It can be understood that the value of each time step and each feature dimension in the fused feature sequence is a weighted sum of the meteorological sub-feature value and the power sub-feature value at the corresponding position, ensuring the refinement and targeting of the fusion. Element-wise operation allows each time step and each feature dimension to achieve collaborative fusion of meteorological and power features. For example, the short-term feature dimension of irradiance at time t is only weighted and fused with the power irradiance associated feature dimension at time t, accurately preserving the time scale and physical meaning of each sub-feature without severing the refined feature information extracted in the preceding sequence.

[0092] In a preferred embodiment, the present invention dynamically generates first and second weights dimension-wise based on the current input; and targets the output feature dimensions of TCN and BiLSTM. and Inconsistency, meteorological characteristics are incorporated through a fully connected layer. and power characteristics Mapped to the same dimension Then we have: ; in, The weight matrix of the fully connected layer for meteorological feature mapping. The weight matrix of the fully connected layer for power eigenmaps. This represents the raw output characteristics of the Weather Channel (TCN). This represents the raw output characteristics of the power channel (BiLSTM). These are the features obtained by mapping meteorological characteristics through a fully connected layer. The power features are the features mapped by the fully connected layer. This is the bias vector of the fully connected layer for meteorological feature mapping. is the bias vector of the fully connected layer for power feature mapping, and T is the number of time steps in the input sequence (i.e., the sequence length).

[0093] Global average pooling is performed on the T×C dimensional time-series features (T is the number of time steps, C is the feature dimension) output from the meteorological and power channels respectively: the mean is calculated for each feature dimension along the time dimension (T), compressing the feature dimension from T×C to a 1×C dimensional feature vector, which preserves the distribution information of the feature dimension and realizes global aggregation of the time dimension. ; in, The feature vector is a C-dimensional feature vector obtained by global average pooling of meteorological features. The power features are represented by a C-dimensional feature vector obtained after global average pooling. Let be the feature vector of the mapped meteorological features at time step t. Let be the feature vector of the power feature at time step t after mapping.

[0094] Furthermore, and After concatenation, the data is sent to the entry control unit. This unit consists of two fully connected layers: the first layer uses ReLU as the activation function for feature dimensionality reduction; the second layer uses Sigmoid as the activation function, with an output dimension of... weight vector If each element takes values ​​in the range [0,1], then: ; Weight vector Each element of the weight vector represents the contribution ratio of meteorological channel information in the corresponding feature dimension; Each element represents the contribution ratio of probability channel information in the corresponding feature dimension. This is the weight matrix of the first fully connected layer of the gated unit, used for feature dimensionality reduction; The weight matrix of the second fully connected layer of the gated unit maps the dimensionality-reduced features to a C-dimensional vector; This is the bias vector of the first fully connected layer of the gated unit; Let α be the bias vector of the second fully connected layer of the gated unit, and α be the first weight, i.e., the meteorological channel weight vector, with dimension 1. , indicating that each element represents the contribution ratio of meteorological information on the corresponding feature dimension; The second weight, namely the power channel weight vector, is where each element represents the contribution ratio of power information on the corresponding feature dimension.

[0095] Indicatively, the aforementioned weights are entirely determined by the real-time feature distribution of the current input sample, achieving dynamic adaptation based on sample dependence.

[0096] For each time step, perform element-wise weighted summation to obtain the fused feature sequence. Its expression is: ; Schematic, characteristic sequence It retains the respective time-series patterns of weather and power, and also achieves adaptive adjustment of information emphasis based on current input characteristics, specifically as follows: When the photovoltaic power station is in a clear and stable period, the autoregressive property of the historical power series is strong, and the gating network automatically reduces... The model primarily relies on the power channel; when clouds pass overhead or irradiance fluctuates drastically, the gating network automatically adjusts its settings. The contribution weight of the meteorological channel increases. Furthermore, when a meteorological sensor experiences a momentary malfunction or communication interruption, abnormal inputs cause a shift in the distribution of meteorological characteristics, which the gating network can automatically adjust. The decrease approaches 0, causing the model to smoothly degenerate into a pure time-series prediction mode, thus giving the system soft sensor redundancy capability.

[0097] In some embodiments, the photovoltaic power prediction model includes: a probabilistic modeling network; The step of the photovoltaic power prediction model outputting the target photovoltaic power corresponding to the future time step based on the fused feature sequence includes: The probabilistic modeling network captures long-term temporal dependencies in the fused feature sequence to generate a latent variable sequence; wherein, the latent variable sequence corresponds to a hidden state at several time points, and each hidden state is used to indicate the temporal evolution law of photovoltaic power under the combined influence of meteorological and historical power. The probabilistic modeling network outputs the Gaussian mixture distribution parameters corresponding to the hidden variable sequence based on each hidden state; wherein, the Gaussian mixture distribution parameters include: several Gaussian distributions; each Gaussian distribution includes: Gaussian distribution weights, Gaussian distribution mean, and Gaussian distribution positive variance; Based on the mean of each Gaussian distribution, and the Gaussian distribution weight and positive variance corresponding to each Gaussian distribution mean, a weighted sum of Gaussian distributions is calculated, and the weighted sum of Gaussian distributions is used as the predicted value of the target photovoltaic power corresponding to the future time step.

[0098] Indicatively, the fusion feature sequence integrates short / long-term correlation features of meteorological channels, such as short-term abrupt changes in irradiance and long-term accumulation of temperature and humidity, as well as four types of time-series features of the power channel, such as the long-term trend of power with irradiance and the influence of temperature on power extremes. However, these features still need to be further explored to uncover the co-evolutionary laws of meteorological driving and power autoregression, such as the chain reaction of power inertia increase and efficiency decay caused by temperature and humidity accumulation when irradiance continues to climb. In this embodiment of the invention, a probabilistic modeling network is constructed to capture the long-term time-series dependence of the fusion feature sequence, generating a hidden variable sequence containing several hidden states. Each hidden state encodes the photovoltaic power evolution law under the combined influence of meteorological factors and historical power. Then, through time-series modeling, the scattered multi-dimensional features can be linked into a continuous evolutionary link, allowing the model to understand the causal transmission logic between current features and future power, making the prediction results more consistent with the actual change law of photovoltaic power.

[0099] It is understandable that photovoltaic power is highly random and intermittent due to meteorological factors. Traditional point prediction only outputs a single predicted value, which cannot accurately predict photovoltaic power based on the reliability of the predicted value. However, the probabilistic modeling network of this invention can output mixture Gaussian distribution parameters, such as weights, mean, and positive variance, thereby quantifying the uncertainty of photovoltaic power. The multi-Gaussian weighted characteristics of the mixture Gaussian distribution can accurately characterize the multimodal distribution of photovoltaic power, such as the distribution characteristics of power concentrated in the high value range during clear and stable periods, while power is dispersed in the medium and low value range when clouds pass by. This distribution characteristic is closer to the actual distribution law of power than a single Gaussian distribution.

[0100] Optionally, the probabilistic modeling network of this invention employs a combination of hybrid density networks (MDN) and recurrent neural networks (RNN) to perform temporal modeling and multimodal prediction of the fused feature sequence. To ensure that the model can capture long-term dependencies and complex temporal evolution features, this invention first uses a recurrent neural network (RNN) to dynamically model the latent variable sequence.

[0101] The hidden state at time t is obtained by updating the current state using the hidden state from the previous time step and the input of the current hidden variable. Its expression is: ; In the formula, The fusion feature module output by AGFU; and is the corresponding weight matrix; b is the bias term; f(.) is the nonlinear activation function.

[0102] After obtaining the hidden state, in order to describe the uncertainty of the predicted target, this invention introduces a hybrid density network (MDN) into the output layer of the recurrent neural network (RNN). The hybrid density network can normalize the original output value of each distribution component using a soft-max function to obtain the hybrid weights (i.e., Gaussian distribution weights), the expression of which is: ; In the formula, For the first The weights of the Gaussian distributions are given by G, where G is the total number of Gaussian distributions. This represents an exponential function that maps any real number to a positive number, ensuring that the variance / standard deviation is positive. This indicates that the output layer of the model corresponds to the first... The original weight values ​​(unnormalized) of the Gaussian distribution are derived from... Mapped to obtain; Indicates the first The unnormalized original weights of the Gaussian distributed components The sum of the exponents of all the original weights of the Gaussian distribution is used as the normalized denominator; Simultaneously, the model directly outputs the mean of each distribution component, and ensures that the variance is positive through exponential transformation, i.e.: ; in, For the first The variance of each Gaussian distribution component is guaranteed to be positive through exponential transformation; This indicates that the output layer of the model corresponds to the first... The original variance output values ​​of a Gaussian distribution are obtained from... The mapping is obtained.

[0103] Ultimately, the predicted target at each time t is the photovoltaic power output value. The conditional probability distribution is represented by a weighted sum of multiple Gaussian distributions, that is, the formula for calculating the weighted sum of Gaussian distributions is: ; in, Indicates the target predicted at time t. The conditional probability density function, i.e. The probability distribution of photovoltaic power; This represents a Gaussian distribution (normal distribution). Indicates that it is in a hidden state. Under the conditions, the first A Gaussian distribution for photovoltaic power The probability density distribution, i.e. Obey For the mean, Let be a normal distribution of variance; where, Indicates the first The mean of each Gaussian distribution component is given by It is obtained by direct mapping and characterizes the power center trend; Indicates the first The variance of each Gaussian distribution component characterizes the degree of power dispersion.

[0104] In illustrative terms, actual photovoltaic power plant operation often involves mixed conditions with multiple overlapping meteorological factors, such as a slow increase in irradiance, accumulation of temperature and humidity, and short-term cloud cover. A single Gaussian distribution cannot characterize the power distribution in such complex scenarios. This invention adapts multiple Gaussian distributions to different sub-scenarios within the mixed conditions. For example, rising irradiance corresponds to a high-power distribution, cloud cover corresponds to a medium-power distribution, and accumulation of temperature and humidity corresponds to a power attenuation distribution. By weighting and integrating the effects of various sub-scenarios, the predicted value can simultaneously reflect the synergistic effect of multiple sub-scenarios, outputting accurate predicted values ​​that reflect both irradiance-driven ramp-up and the effects of cloud cover and temperature / humidity attenuation.

[0105] In some preferred embodiments, the present invention further includes: Each time the photovoltaic power prediction model is trained, its parameters are updated according to the following loss function: ; ; in, The value of the loss function. The negative log-likelihood of the probabilistic modeling network is used to measure the likelihood matching between the Gaussian distribution output by the probabilistic modeling network and the actual value of photovoltaic power. Mean squared error loss is used to indicate the error between the photovoltaic power prediction output by the model and the actual photovoltaic power. for The corresponding loss balance weight, for The corresponding loss balance weight.

[0106] Schematic, the present invention constructs a negative log-likelihood. and mean square error loss The hybrid loss function can achieve simultaneous optimization of probability prediction accuracy and feature extraction effectiveness through the synergy of dual loss terms and the adjustment of balanced weights. This ensures that the Gaussian distribution output by the probabilistic modeling network closely matches the actual distribution of photovoltaic power, improving the accuracy of point prediction and the credibility of uncertainty quantification. It can constrain the features extracted by the model to retain the core information of the original data, avoid distortion and mixing of heterogeneous features, and prevent the error between the output photovoltaic power prediction result and the actual photovoltaic power from being too large. The weighted fusion of the two allows different weights to adapt to the needs of different photovoltaic actual operation scenarios, while supporting the coordinated convergence of the parameters of each module in end-to-end training, and ultimately maximizing the overall prediction performance and engineering practicality of the model.

[0107] Optionally, in order to ensure that the parameters of the meteorological feature extraction channel, power feature extraction channel, feature fusion channel and probabilistic modeling network are updated collaboratively in the same training process to achieve the overall optimal performance, this embodiment of the invention adopts an end-to-end joint training method to globally optimize the parameters of the entire model.

[0108] Optionally, embodiments of the present invention may employ a gradient descent method to adjust the total loss. Backpropagation is performed, and the parameters of the meteorological feature extraction channel, power feature extraction channel, feature fusion channel, and probabilistic modeling network are updated synchronously. This allows the three stages of feature extraction, dynamic fusion, and probabilistic prediction to promote each other and optimize synergistically. No staged pre-training is required, which simplifies the engineering deployment process and significantly improves the overall performance of the model.

[0109] Please see Figure 2 , Figure 2 The entire execution process of this invention, from data input to probabilistic power prediction, is presented. The execution process of each step is as follows: First, input the meteorological time series and the historical power time series; specifically, you can input the current and historical meteorological time series data and the historical photovoltaic power time series data. Each historical time step is synchronously bound to the historical meteorological data and historical photovoltaic power at the corresponding moment, ensuring that the time series of the two types of heterogeneous data correspond one-to-one, providing the original data foundation for subsequent feature extraction and prediction.

[0110] Data preprocessing and sliding window construction are performed. Specifically, the input data can be standardized and time-series samples can be constructed. Preprocessing steps include: detecting and removing outliers, filling in missing photovoltaic power data through mean imputation / linear interpolation, and using Z-score normalization to eliminate the dimensional differences between meteorological and power data, ensuring data integrity and standardization. Sliding window construction involves splitting long time-series data into standardized time-series subsequences, fully preserving the time dependence between meteorological and power data, constructing training / prediction samples that the model can directly process, and providing regular time-series input for subsequent dual-channel feature extraction.

[0111] Next, heterogeneous encoding and feature processing are performed based on asymmetric dual channels. Specifically, meteorological and power channels are used to extract differentiated features from the preprocessed data, adapting to the essential characteristics of the two types of heterogeneous data. The meteorological channel extracts current meteorological features, first convolutional features, and second convolutional features to accurately capture the patterns of gradual changes and multi-scale temporal correlations in meteorological data. The power channel (BiLSTM backbone network) extracts long-term and short-term power trend features along the forward time sequence, such as the first and second power time-series features. The extracted features are subjected to adaptive gating fusion. Specifically, the multi-dimensional meteorological features output from the meteorological channel and the two power time-series features output from the power channel are subjected to adaptive weighted fusion under different operating conditions. The global distribution information of meteorological and power features is generated by global average pooling for each fused feature. Then, the meteorological weights and power weights are dynamically generated by gating mechanism. Finally, the fused feature sequence is obtained by weighted summation of each element. This achieves the synergistic integration of meteorological driving law and power autoregressive law, while also having fault tolerance capability when meteorological data is abnormal.

[0112] Finally, probabilistic modeling is performed based on hybrid density networks (MDN) and recurrent neural networks (RNN) to output the predicted target photovoltaic power. Specifically, probabilistic photovoltaic power prediction is carried out based on the fused feature sequence: first, the long-term time-series dependence of the fused features is captured by RNN to generate a sequence of latent variables; then, the mixture Gaussian distribution parameters are output by MDN to characterize the multimodal distribution of photovoltaic power and the prediction uncertainty; finally, the weighted sum of the Gaussian distribution is calculated to obtain the predicted value of the target photovoltaic power at the future time step, while providing probability distribution information (such as the power corresponding to the 95% confidence interval), thereby providing accurate prediction results and risk decision-making basis for grid dispatch and photovoltaic power plant operation and maintenance.

[0113] In summary, this invention constructs an asymmetric dual-channel heterogeneous feature extraction architecture. When dealing with the different physical properties of meteorological and power sequences, a temporal convolutional network (TCN) based on dilated causal convolution is constructed for the meteorological channel to capture long-range periodic dependencies, and a bidirectional long short-term memory network (BiLSTM) is designed for the power channel to enhance the perception of local abrupt events, thereby realizing the extraction of structurally heterogeneous features and significantly improving the interpretability of the model.

[0114] This invention proposes an adaptive gating fusion channel, which dynamically generates continuous fusion weights dimension by dimension based on input data and input features. This fundamentally transforms the fusion mechanism from static preset to sample-driven, enabling the model to autonomously weigh the contribution ratio of meteorological and power features according to real-time operating conditions. Furthermore, it can automatically degenerate to single-channel mode when the sensor fails, possessing soft sensor redundancy capabilities and can be generalized to various multi-source time-series fusion tasks. This invention also proposes an end-to-end joint optimization framework, which enables the negative log-likelihood gradient of the probabilistic modeling network to propagate back to the meteorological feature extraction channel, power feature extraction channel, and feature fusion channel, driving the coordinated evolution of the three major links of feature extraction, dynamic fusion, and probabilistic modeling. While improving the accuracy of point prediction, it outputs a well-calibrated multimodal probability density distribution, comprehensively enhancing the model's generalization ability and uncertainty quantification level under complex meteorological conditions.

[0115] Please see Figure 3 This application also provides a photovoltaic power prediction device based on multi-source heterogeneous data, which can realize the above-mentioned photovoltaic power prediction method based on multi-source heterogeneous data. The device includes: The data acquisition module is used to acquire current meteorological data at the current time step and photovoltaic power sequence data to be input; wherein, the photovoltaic power sequence data includes several historical time steps; each historical time step corresponds to a historical photovoltaic power and a historical meteorological data. The photovoltaic power prediction module is used to input the current meteorological data and the photovoltaic power sequence data into the photovoltaic power prediction model. The meteorological feature extraction channel in the photovoltaic power prediction model extracts features from the current meteorological data and the historical meteorological data corresponding to each historical time step, outputting current meteorological features and historical meteorological time-series features. The power feature extraction channel in the photovoltaic power prediction model extracts features from the historical photovoltaic power corresponding to each historical time step, outputting power time-series features characterizing the power fluctuation trend. The feature fusion channel in the photovoltaic power prediction model fuses the current meteorological features, the historical meteorological time-series features, and the power time-series features, outputting a fused feature sequence. Based on the fused feature sequence, the photovoltaic power prediction model outputs the target photovoltaic power corresponding to the future time step.

[0116] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0117] It should be noted that the device embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0118] Those skilled in the art will clearly understand that, for convenience and simplicity, the specific working process of the device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0119] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the aforementioned photovoltaic power prediction method based on multi-source heterogeneous data. This electronic device can include any smart terminal such as a tablet computer or in-vehicle computer.

[0120] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0121] Please see Figure 4 , Figure 4 This illustrates the hardware structure of an electronic device according to another embodiment, the electronic device comprising: The processor can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to achieve the technical solutions provided in the embodiments of this application. The memory can be implemented in the form of read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory can store the operating system and other applications. When the technical solutions provided in the embodiments of this application are implemented through software or firmware, the relevant program code is stored in the memory and called and executed by the processor. Input / output interfaces are used to implement information input and output; The communication interface is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, Wi-Fi, Bluetooth, etc.). A bus is used to transfer information between various components of a device, such as processors, memory, input / output interfaces, and communication interfaces. The processor, memory, input / output interface, and communication interface are interconnected within the device via a bus.

[0122] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0123] The memory can be used to store the computer program. The processor implements various functions of the terminal device by running or executing the computer program stored in the memory and calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function, etc.; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device or other volatile solid-state storage device.

[0124] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described photovoltaic power prediction method based on multi-source heterogeneous data.

[0125] It is understood that the content of the above method embodiments is applicable to the present computer storage medium embodiments. The specific functions implemented by the present computer storage medium embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0126] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described photovoltaic power prediction method based on multi-source heterogeneous data.

[0127] It is understood that the content of the above method embodiments is applicable to the embodiments of this computer program product. The specific functions implemented by the embodiments of this computer program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0128] Those skilled in the art will understand that all or some of the steps, apparatuses, or functional modules / units in the methods disclosed above can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0129] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A photovoltaic power prediction method based on multi-source heterogeneous data, characterized in that, The method includes: Obtain current meteorological data at the current time step and photovoltaic power sequence data to be input; wherein, the photovoltaic power sequence data includes several historical time steps; each historical time step corresponds to a historical photovoltaic power and a historical meteorological data. The current meteorological data and the photovoltaic power sequence data are input into the photovoltaic power prediction model so that the meteorological feature extraction channel in the photovoltaic power prediction model can extract features from the current meteorological data and the historical meteorological data corresponding to each historical time step, and output the current meteorological features and historical meteorological time sequence features. The power feature extraction channel in the photovoltaic power prediction model extracts features from the historical photovoltaic power corresponding to each historical time step, and outputs power time series features to characterize the fluctuation trend of power. The photovoltaic power prediction model fuses the current meteorological features, the historical meteorological time series features, and the power time series features through its feature fusion channel, and outputs a fused feature sequence. The photovoltaic power prediction model outputs the target photovoltaic power corresponding to the future time step based on the fusion feature sequence.

2. The photovoltaic power prediction method based on multi-source heterogeneous data according to claim 1, characterized in that, The training process of the photovoltaic power prediction model includes... A meteorological data sample with the same time step and a photovoltaic power sequence data sample to be trained are used as training samples; wherein, the photovoltaic power sequence data sample includes several historical time steps; each historical time step corresponds to a historical photovoltaic power sample and a historical meteorological data sample. Using each of the training samples as input and the actual photovoltaic power corresponding to each training sample as a supervision label, the photovoltaic power prediction model is iteratively trained until the error between the photovoltaic power prediction result output by the model and the supervision label meets a preset threshold, thereby generating the photovoltaic power prediction model.

3. The photovoltaic power prediction method based on multi-source heterogeneous data according to claim 1, characterized in that, The historical meteorological time series features include: a first convolutional feature and a second convolutional feature; The meteorological feature extraction channel in the photovoltaic power prediction model extracts features from the current meteorological data and the historical meteorological data corresponding to each historical time step, and outputs the current meteorological features and historical meteorological time series features, including: The current meteorological data is convolved through the meteorological feature extraction channel to capture the spatial distribution characteristics of the current meteorological data at the current time step and output the current meteorological features. The historical meteorological data is convolved to capture the spatial distribution characteristics of the historical meteorological data at each historical time step, and the first convolution feature is output; wherein, the first convolution feature is used to characterize the state of historical meteorology at a single time step. The current meteorological features and the first convolutional features are convolved to output short-distance meteorological correlation features corresponding to short time steps and long-distance meteorological correlation features corresponding to long time steps; wherein, the short-distance meteorological correlation features are used to characterize the short-term meteorological change trend between adjacent historical time steps; the long-distance meteorological correlation features are used to characterize the long-term meteorological evolution trend across multiple historical time steps. The short-range meteorological correlation features and the long-range meteorological correlation features are fused to output a second convolutional feature that characterizes the cross-time period correlation pattern between historical meteorological data and current meteorological data.

4. The photovoltaic power prediction method based on multi-source heterogeneous data according to claim 3, characterized in that, The current meteorological data includes: current solar irradiance, current ambient temperature, current ambient humidity, and current cloud cover; the meteorological feature extraction channel includes: causal convolutional layers and dilated convolutional layers; The step of convolving the current meteorological data through the meteorological feature extraction channel to capture the spatial distribution characteristics of the current meteorological data at the current time step and outputting the current meteorological features includes: The causal convolutional layer captures the spatial distribution characteristics of the current solar irradiance, current ambient temperature, current ambient humidity, and current cloud cover data at the current time step, and the dilated convolutional layer performs dilated convolution on each of the spatial distribution characteristics to output the current meteorological features.

5. The photovoltaic power prediction method based on multi-source heterogeneous data according to claim 4, characterized in that, The historical meteorological data includes: historical solar irradiance, historical ambient temperature, historical ambient humidity, and historical cloud cover data; the first convolutional feature includes: historical irradiance feature, historical ambient temperature feature, historical ambient humidity feature, and historical cloud cover feature; the dilated convolutional layer corresponds to a dilation factor. The step of convolving the current meteorological features and the first convolutional features to output short-range meteorological correlation features corresponding to short time steps and long-range meteorological correlation features corresponding to long time steps includes: The expansion factor is adjusted exponentially based on each historical time step and a preset multiple to generate the adjusted expansion factor. The receptive field of the current meteorological features, historical irradiance features, and historical cloud cover features is expanded by the expanded convolutional layer and the adjusted expansion factor, respectively, to extract irradiance trend features and cloud cover change trend features. Then, the irradiance trend features and cloud cover change trend features are output as short-range meteorological correlation features. The receptive field of the historical ambient temperature features and the historical ambient humidity features is expanded by the expanded convolutional layer and the adjusted expansion factor, respectively, and the temperature and humidity coupling features used to characterize the cumulative impact of temperature and humidity changes on photovoltaic modules are extracted. Then, the temperature and humidity coupling features are output as the long-distance meteorological correlation features.

6. The photovoltaic power prediction method based on multi-source heterogeneous data according to claim 5, characterized in that, The power timing characteristics include: a first power timing characteristic and a second power timing characteristic; The process of extracting features from the historical photovoltaic power corresponding to each historical time step using the power feature extraction channel in the photovoltaic power prediction model, and outputting power time-series features to characterize the power fluctuation trend, includes: Through the power feature extraction channel, the long-term dependence of each of the historical photovoltaic power is extracted based on the positive time sequence, and the first power time series feature is extracted to characterize the long-term evolution trend of power over time. Through the power feature extraction channel, the short-term dependence of each of the historical photovoltaic power is extracted based on the reverse time order, and a second power time series feature is generated to characterize the short-period fluctuation trend of power over time.

7. The photovoltaic power prediction method based on multi-source heterogeneous data according to claim 6, characterized in that, The feature fusion channel of the photovoltaic power prediction model fuses the current meteorological features, the historical meteorological time-series features, and the power time-series features to output a fused feature sequence, including: The current meteorological features, the first convolutional features, the second convolutional features, the first power time series features, and the second power time series features are fused through the feature fusion channel to generate a first fusion feature for characterizing the trend of photovoltaic power change with irradiance, a second fusion feature for characterizing the trend of photovoltaic power change with cloud cover, a third fusion feature for characterizing the degree of influence of ambient temperature on the extreme value of photovoltaic power, and a fourth fusion feature for characterizing the degree of influence of ambient humidity on the ramp-up of photovoltaic power. Global average pooling is performed on the first fusion feature, the second fusion feature, the third fusion feature, and the fourth fusion feature in the time dimension to generate a first feature distribution that reflects the global information of the meteorological feature extraction channel and a second feature distribution that reflects the global information of the power feature extraction channel. Based on the gating mechanism of the feature fusion channel, the dependency relationship between the first feature distribution and the second feature distribution is captured, and a first weight of the meteorological feature extraction channel and a second weight of the power feature extraction channel are generated; wherein, the first weight is used to measure the degree of influence of meteorological factors on power; and the second weight is used to measure the degree of influence of historical power on power. Based on the first weight and the second weight, the current meteorological feature, the first convolutional feature, the second convolutional feature, the first power time series feature, and the second power time series feature are summed element by element to generate a fused feature sequence.

8. The photovoltaic power prediction method based on multi-source heterogeneous data according to claim 7, characterized in that, The photovoltaic power prediction model includes: a probabilistic modeling network; The step of the photovoltaic power prediction model outputting the target photovoltaic power corresponding to the future time step based on the fused feature sequence includes: The probabilistic modeling network captures long-term temporal dependencies in the fused feature sequence to generate a latent variable sequence; wherein, the latent variable sequence corresponds to a hidden state at several time points, and each hidden state is used to indicate the temporal evolution law of photovoltaic power under the combined influence of meteorological and historical power. The probabilistic modeling network outputs the Gaussian mixture distribution parameters corresponding to the hidden variable sequence based on each hidden state; wherein, the Gaussian mixture distribution parameters include: several Gaussian distributions; each Gaussian distribution includes: Gaussian distribution weights, Gaussian distribution mean, and Gaussian distribution positive variance; Based on the mean of each Gaussian distribution, and the Gaussian distribution weight and positive variance corresponding to each Gaussian distribution mean, a weighted sum of Gaussian distributions is calculated, and the weighted sum of Gaussian distributions is used as the predicted value of the target photovoltaic power corresponding to the future time step.

9. The photovoltaic power prediction method based on multi-source heterogeneous data according to claim 8, characterized in that, The method further includes: Each time the photovoltaic power prediction model is trained, its parameters are updated according to the following loss function: ； ； in, The value of the loss function. The negative log-likelihood of the probabilistic modeling network is used to measure the likelihood matching between the Gaussian distribution output by the probabilistic modeling network and the actual value of photovoltaic power. Mean squared error loss is used to indicate the error between the photovoltaic power prediction output by the model and the actual photovoltaic power. for The corresponding loss balance weight, for The corresponding loss balance weight.

10. A photovoltaic power prediction device based on multi-source heterogeneous data, characterized in that, The device includes: The data acquisition module is used to acquire current meteorological data at the current time step and photovoltaic power sequence data to be input; wherein, the photovoltaic power sequence data includes several historical time steps; each historical time step corresponds to a historical photovoltaic power and a historical meteorological data. The photovoltaic power prediction module is used to input the current meteorological data and the photovoltaic power sequence data into the photovoltaic power prediction model. The meteorological feature extraction channel in the photovoltaic power prediction model extracts features from the current meteorological data and the historical meteorological data corresponding to each historical time step, outputting current meteorological features and historical meteorological time-series features. The power feature extraction channel in the photovoltaic power prediction model extracts features from the historical photovoltaic power corresponding to each historical time step, outputting power time-series features characterizing the power fluctuation trend. The feature fusion channel in the photovoltaic power prediction model fuses the current meteorological features, the historical meteorological time-series features, and the power time-series features, outputting a fused feature sequence. Based on the fused feature sequence, the photovoltaic power prediction model outputs the target photovoltaic power corresponding to the future time step.

Images