Photovoltaic output medium and long term prediction method, system and equipment based on fusion of cloud edge architecture

Abstract

This invention discloses a method, system, and device for medium- and long-term photovoltaic (PV) output prediction based on a cloud-edge-device architecture, belonging to the field of new energy power system prediction technology. It includes constructing a three-layer collaborative prediction architecture at the cloud, edge, and terminal levels; collecting historical multi-source data and performing time alignment, missing data repair, and normalization to form a unified feature representation; training multiple prediction models at the cloud and edge levels to obtain multiple basic prediction results; performing fusion processing on the basic prediction results to generate a comprehensive medium- and long-term PV output prediction result; constructing hierarchical trigger criteria based on prediction bias and data distribution changes, and performing edge-side parameter fine-tuning, cloud-edge-device collaborative re-weighting of model fusion weights and adjustment of the prediction model combination structure, as well as full model retraining and synchronous updating. This solution achieves high accuracy, high robustness, and adaptive maintenance in medium- and long-term PV output prediction through cloud-edge-device collaboration and multi-model fusion.

Description

Technical Field

[0001] This invention relates to the field of prediction technology for new energy power systems, and specifically discloses a method, system and equipment for medium- and long-term prediction of photovoltaic output that integrates cloud-edge-device architecture. Background Technology

[0002] As a crucial link in the new power system that directly supplies power to users, the power quality level of the distribution network directly determines the reliability and experience of electricity use for a large number of users. Distributed photovoltaic power has entered a new stage of large-scale integration into the distribution network, and its installed capacity is experiencing explosive growth.

[0003] However, while large-scale distributed photovoltaic (PV) grid integration optimizes the energy structure, its fluctuating and intermittent output, along with its grid connection via power electronic inverters, also presents significant power quality challenges to the distribution network. Particularly on medium- to long-term (weekly, monthly, quarterly, annual, and longer-term) scales, the trend changes and periodic fluctuations in PV output, as well as the uncertainty of extreme weather events, pose greater challenges to grid operation planning, capacity allocation, absorption capacity assessment, and the integration of renewable energy. Traditional dispatching methods relying on short- or ultra-short-term forecasts cannot meet the requirements of medium- to long-term planning and resource allocation for forecast accuracy, stability, and reliability.

[0004] Currently, the mainstream technologies used in the industry for medium- and long-term forecasting mainly include time series models based on historical statistics and regression models based on climate data. However, these methods generally exhibit characteristics of "static modeling" and "posterior fitting," and their predictive capabilities often rely on fixed model structures. They lack sufficient exploration of long-term meteorological trends, cross-seasonal climate models, and the correlations between various forecasting models, leading to a significant decrease in prediction accuracy under seasonal transitions, annual scale changes, or extreme climate disturbances. Specifically, they have the following shortcomings:

[0005] (1) Insufficient model adaptability: A single model is difficult to take into account both short-term fluctuations and medium-to-long-term trends. The model has poor generalization ability and large cross-seasonal prediction errors, which cannot meet the needs of distribution network planning, energy storage configuration and annual consumption assessment.

[0006] (2) Single function and lack of collaborative optimization capability: Traditional methods focus on a single prediction task and cannot simultaneously handle site-level and regional / distribution network-level predictions. They also lack the ability to integrate multi-source information such as meteorological forecasts, historical data, and operational data, and cannot perform cross-model and cross-scale collaborative predictions.

[0007] (3) Lack of uncertainty analysis: Most traditional prediction results only provide point prediction values ​​and cannot provide uncertainty information such as prediction intervals and probability distributions. They cannot support key needs such as power grid planning scenario analysis and reserve capacity assessment, resulting in insufficient safety margin or excessive redundancy in planning strategies.

[0008] In practical applications, various forecasting tools, data sources, and analytical models often operate independently, lacking a unified forecasting framework and cross-model collaboration mechanism. Forecasting results from different time scales and spatial levels are difficult to coordinate and use, leading to fragmentation of the forecasting system and affecting the scientific nature of power grid resource allocation and operation planning.

[0009] Therefore, there is an urgent need for a medium- to long-term photovoltaic power output prediction method that integrates multiple models, multiple scales, and multiple source data to achieve unified management, dynamic updating, and cross-regional collaborative computing of prediction models. Summary of the Invention

[0010] To address the aforementioned issues, this invention proposes a method, system, and device for medium- to long-term photovoltaic power output prediction that integrates a cloud-edge-device architecture. By introducing a multi-timescale prediction mechanism, the photovoltaic system is transformed from a passive prediction mode that relies on a single model and posterior fitting to an active prediction scheme based on multi-model fusion, cross-scale collaborative computing, interpretability analysis, and uncertainty quantification. This solves the problems of insufficient prediction accuracy, poor model adaptability, inability to simultaneously support site-level and regional-level predictions, and lack of a unified prediction architecture in existing technologies.

[0011] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for medium- and long-term prediction of photovoltaic power output based on a cloud-edge-device architecture, comprising: A three-layer collaborative prediction architecture is constructed, comprising a cloud for performing long-term fusion prediction, an edge for performing regional collaborative prediction, and a terminal for performing raw measurement data preprocessing and feature construction. Collect at least two types of data from historical multi-source data, and perform time alignment, missing data repair, and normalization on the collected multi-source data to form a unified feature representation; Based on the unified feature representation, multiple prediction models are trained on the cloud and the edge respectively, and prediction calculations are performed using each prediction model to obtain corresponding multiple basic prediction results. The multiple basic prediction results are fused to generate a medium- and long-term comprehensive prediction result for photovoltaic power output; Based on the prediction deviation of the medium- and long-term comprehensive forecast results of photovoltaic power output and the changes in the data distribution of the unified feature representation, a hierarchical triggering criterion is constructed. When the first triggering level is determined, the parameters are fine-tuned on the edge side based on the unified feature representation formed by the latest collected multi-source data. When the second triggering level is determined, the cloud and the edge side jointly perform the recalculation of model fusion weights and the adjustment of the prediction model combination structure. When the third triggering level is determined, the deployed multi-type prediction models are retrained, and the updated model parameters and model structure configuration are synchronized to the edge side and the terminal side.

[0012] Optionally, when the cloud is used to perform long-cycle fusion prediction, it employs a long-sequence model based on an attention mechanism to model the long-term time-series dependency of the unified feature representation to generate a first predicted value, and employs a seasonal time series model to model the periodicity of the historical photovoltaic power output data to generate a second predicted value; the first predicted value and the second predicted value are fused to generate the cloud-side basic prediction result.

[0013] Optionally, the attention-based long sequence model includes: Transformer model and Longformer model; The seasonal time series models include: SARIMA model and Prophet model.

[0014] Optionally, when the edge side is used to perform regional collaborative prediction, a regional local prediction model is constructed based on the unified feature representation corresponding to each regional node, and a consistency iteration is performed based on the prediction error feedback between adjacent regional nodes, so that the local prediction results of each regional node are collaboratively corrected during the iteration process to generate the edge side basic prediction results.

[0015] Optionally, the historical multi-source data includes historical photovoltaic power output data, numerical weather forecast data, and climate index data; wherein the numerical weather forecast data includes irradiance, temperature, wind speed, and cloud cover, and the climate index data includes seasonal climate index and long-term sunshine trend index.

[0016] Optionally, the construction of the hierarchical triggering criterion based on the prediction deviation of the medium- and long-term comprehensive prediction results of photovoltaic power output and the data distribution changes of the unified feature representation includes: setting a first error threshold, a second error threshold, a first drift threshold, and a second drift threshold; wherein, the first error threshold is less than the second error threshold, and the first drift threshold is less than the second drift threshold; When the prediction deviation of the medium- and long-term comprehensive forecast of photovoltaic power output is less than the first error threshold, and the change in the data distribution represented by the unified feature is less than the first drift threshold, it is determined to be at the first trigger level; when the prediction deviation of the medium- and long-term comprehensive forecast of photovoltaic power output is between the first error threshold and the second error threshold, or when the change in the data distribution represented by the unified feature is between the first drift threshold and the second drift threshold, it is determined to be at the second trigger level; when the prediction deviation of the medium- and long-term comprehensive forecast of photovoltaic power output is not less than the second error threshold, or when the change in the data distribution represented by the unified feature is not less than the second drift threshold, it is determined to be at the third trigger level.

[0017] Optionally, when the second trigger level is determined, the recalculation of model fusion weights and adjustment of the prediction model combination structure by the cloud and edge sides collaboratively includes: recalculating the fusion weights of each prediction model based on the prediction errors of each prediction model within the most recent preset time period; and performing the enabling, disabling, or replacement operation of the prediction model based on the magnitude and trend of the prediction error of each prediction model; if the prediction error of a certain prediction model is higher than a preset error threshold in multiple consecutive update cycles, the fusion weight of the prediction model is reduced, or the prediction model is temporarily removed from the fusion structure; when the prediction performance of the prediction model recovers to the preset standard, it is added back to the fusion structure.

[0018] Optionally, the fusion processing of multiple basic prediction results includes: using the multiple basic prediction results as input features, training a fusion weight learning model to learn the fusion weights corresponding to each prediction model, and outputting point prediction results; and determining the posterior weights of the models based on the prediction performance of each prediction model on the validation dataset, and weighting each basic prediction result according to the posterior weights of the models to obtain a probability prediction distribution. Obtain a medium- to long-term comprehensive forecast of photovoltaic power output based on the point prediction results and the probability prediction distribution.

[0019] Secondly, the present invention provides a medium- and long-term photovoltaic power output prediction system integrating cloud-edge-device architecture, comprising: An architecture building module is used to construct a three-layer collaborative prediction architecture for cloud, edge, and terminal. The three-layer collaborative prediction architecture includes a cloud for performing long-term fusion prediction, an edge for performing regional collaborative prediction, and a terminal for performing raw measurement data preprocessing and feature construction. The preprocessing module is used to collect at least two types of data from historical multi-source data, and to perform time alignment, missing data repair and normalization on the collected multi-source data to form a unified feature representation. The model prediction module is used to train multiple prediction models on the cloud and the edge side respectively based on the unified feature representation, and to perform prediction calculations using each prediction model to obtain corresponding multiple basic prediction results. The fusion processing module is used to perform fusion processing on the multiple basic prediction results to generate a medium- and long-term comprehensive prediction result for photovoltaic power output. The action execution module is used to construct a hierarchical triggering criterion based on the prediction deviation of the medium- and long-term comprehensive prediction results of photovoltaic power output and the data distribution changes of the unified feature representation. When the first triggering level is determined, the parameters are fine-tuned on the edge side based on the unified feature representation formed by the latest collected multi-source data. When the second triggering level is determined, the cloud and the edge side jointly execute the recalculation of model fusion weights and the adjustment of the prediction model combination structure. When the third triggering level is determined, the deployed multi-type prediction models are retrained, and the updated model parameters and model structure configuration are synchronized to the edge side and the terminal side.

[0020] Thirdly, the present invention provides an electronic device, the electronic device comprising: At least one processor; and a memory communicatively connected to said at least one processor; The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the method described in any one of the first aspects.

[0021] Compared with the closest existing technology, the present invention has the following advantages: This application proposes a method, system, and device for medium- and long-term photovoltaic power output forecasting that integrates cloud-edge-device architecture. The proposed three-layer forecasting architecture based on cloud-edge-device collaboration effectively solves the problems of fragmented forecasting architecture, independent model operation, and inability to accommodate different time scales in existing technologies. By establishing a hierarchical system of long-term multi-model fusion in the cloud, collaborative forecasting in edge regions, and real-time data preprocessing at the terminal, a unified working link is formed for medium- and long-term trend forecasting, short- and medium-term regional forecasting, and local data updates, realizing the systematization and integrated processing of forecasting tasks.

[0022] This invention fundamentally solves the problems of traditional forecasting methods, such as single model, insufficient feature utilization, and lack of cross-model collaboration, by constructing a multi-source data fusion and multi-model collaborative mode. It unifies the processing of multi-source data, including historical output, weather forecasts, and climate indices, and combines multiple forecasting models to construct a comprehensive forecasting system. Through fusion processing, it achieves integrated output of point forecasts and probabilistic forecasts, significantly improving the model's adaptability and generalization ability.

[0023] The hierarchical triggering model update mechanism employed in this invention solves the problems of traditional prediction systems lacking adaptive maintenance capabilities and experiencing a decline in model accuracy over long-term operation. By setting three-level criteria based on prediction bias and data distribution drift, a multi-level update process is achieved, from rapid fine-tuning at the edge, cloud-edge collaborative reweighting, to full retraining in the cloud. This enables the prediction system to continuously respond to data changes, achieving a balance between prediction accuracy, stability, and cost-effectiveness. Attached Figure Description

[0024] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0025] Figure 1 This is a flowchart of a method for medium- and long-term prediction of photovoltaic power output that integrates cloud-edge-device architecture, provided by the present invention. Figure 2 This is a schematic diagram of the structure of the photovoltaic power output medium- and long-term prediction system with integrated cloud-edge-device architecture provided by the present invention; Figure 3 This is a diagram of the internal structure of the electronic device provided by the present invention. Detailed Implementation

[0026] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of the present invention and are therefore merely examples, and should not be construed as limiting the scope of protection of the present invention.

[0027] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0028] This invention provides a method, system, and device for medium- and long-term forecasting of photovoltaic power output based on a cloud-edge-device architecture. The embodiments of this invention are described below with reference to the accompanying drawings.

[0029] Example 1: As Figure 1 As shown, Embodiment 1 of the present invention provides a method for medium- and long-term prediction of photovoltaic power output based on a cloud-edge-device architecture. This method specifically includes the following steps: S101 constructs a three-layer collaborative prediction architecture for cloud, edge, and terminal. The three-layer collaborative prediction architecture includes a cloud for performing long-term fusion prediction, an edge for performing regional collaborative prediction, and a terminal for performing raw measurement data preprocessing and feature construction. S102 collects at least two types of data from historical multi-source data, performs time alignment, missing data repair, and normalization on the collected multi-source data, and forms a unified feature representation; S103 Based on the unified feature representation, multiple prediction models are trained on the cloud and the edge respectively, and prediction calculations are performed using each prediction model to obtain corresponding multiple basic prediction results; S104 performs a fusion process on the multiple basic prediction results to generate a medium- and long-term comprehensive prediction result for photovoltaic power output; S105 constructs a hierarchical triggering criterion based on the prediction deviation of the medium- and long-term comprehensive forecast results of photovoltaic power output and the changes in the data distribution of the unified feature representation. When the first triggering level is determined, the parameters are fine-tuned on the edge side based on the unified feature representation formed by the latest collected multi-source data. When the second triggering level is determined, the cloud and the edge side jointly perform the recalculation of model fusion weights and the adjustment of the prediction model combination structure. When the third triggering level is determined, the deployed multi-type prediction models are retrained, and the updated model parameters and model structure configuration are synchronized to the edge side and the terminal side.

[0030] In one embodiment, step S101 of constructing a cloud-edge-device three-layer collaborative prediction architecture includes: Step A1: Establish a long-term fusion prediction model in the cloud: A set of prediction models is constructed in the cloud with the objective of minimizing long-term prediction errors. The objective function is defined as follows:

[0031] In the formula, This represents the overall objective function of the optimization problem. For time scale sets, Indicates a specific time scale The set of all corresponding time points; Weights for each time scale, To contribute practically, These are cloud-based predictions.

[0032] Solving the objective function yields cloud prediction parameters and model combination schemes at different time scales.

[0033] The long-term fusion prediction model includes: a long-series model based on an attention mechanism and a seasonal time series model. The long-series model based on an attention mechanism includes: the Transformer model and the Longformer model. The seasonal time series models include: SARIMA model and Prophet model.

[0034] Step A2: Use an attention-based long sequence model to model the long-term time-series dependency of the unified feature representation to generate a first predicted value; and use a seasonal time series model to model the periodicity of the historical photovoltaic power output data to generate a second predicted value; fuse the first predicted value and the second predicted value to generate a cloud-side basic prediction result.

[0035] In a specific embodiment, constructing a long-period fusion prediction model based on attention mechanism (Transformer / Longformer) and SARIMA includes the following steps: For each prediction region, the input feature sequence is used. Construct a Transformer-Longformer model and output long-period predictions. :

[0036] in, Represents the mathematical function mapping relationship of long sequence models (Transformer / Longformer) based on attention mechanism; The above indicates The complete set of trainable parameters in the model; Simultaneously, a seasonal ARIMA (SARIMA) time series model is used to characterize the periodic trend, resulting in... The cloud will weight the two accordingly. Fusion generation of long-term benchmark forecasts:

[0037] in, This refers to the final long-term baseline forecast value generated on the cloud side. These are predictions from attention-based long sequence models (Transformer / Longformer). These are the predicted values ​​from the Seasonal Time Series Model (SARIMA). Step A3: Construct a regional collaborative computation mechanism on the edge side. On the edge side, each regional node constructs a local prediction model based on LSTM, and its output is:

[0038] in, The function mapping representing the Long Short-Term Memory (LSTM) network model. For area code, Let be the predicted photovoltaic output value generated by the i-th regional node on the edge side. Let be the feature vector of the i-th region at the current time t; For the i-th region from t The historical feature sequence from time L to time t, where L is the length of the time window; For the function mapping of the Extreme Gradient Boosting (XGBoost) model; Each edge node achieves collaborative correction of regional predictions through consensus iteration:

[0039] in, Let j be the set of neighborhoods of region i, and let j belong to... The number of any adjacent region in the set; The consistency weight between regions is dynamically adjusted through the prediction error feedback between adjacent regional nodes. This represents the final edge-side prediction value output by the i-th region node after the (k+1)-th consensus iteration collaborative correction; This represents the current predicted value of the i-th region node in the k-th iteration. This represents the current predicted value of neighboring region j in the k-th iteration; When the edge side is used to perform regional collaborative prediction, a regional local prediction model is constructed based on the unified feature representation corresponding to each regional node, and a consistency iteration is performed based on the prediction error feedback between adjacent regional nodes, so that the local prediction results of each regional node are collaboratively corrected during the iteration process to generate the edge side basic prediction results.

[0040] Step A4: Construct a terminal-side data preprocessing and cloud-edge-terminal data interaction mechanism: For the raw measurement data Sliding window feature construction and real-time cleaning are performed to obtain preprocessed results:

[0041] in, For window length, For denoising, missing data filling and outlier correction operators; This represents W consecutive historical raw data points; The feature vectors generated on the terminal side are directly used as inputs to the edge and cloud models. Simultaneously, cloud-edge-terminal data interaction latency constraints are defined.

[0042] in, The total communication latency required for a complete data interaction to be completed in a cloud-edge-device three-layer architecture. This refers to the network transmission latency between the cloud and the edge. This refers to the network transmission latency between the edge side and the terminal side. This is the maximum allowable communication delay threshold for the system.

[0043] Ensure that preprocessed data and prediction results can interact within the allowable latency to achieve stable collaborative operation of the three-tier architecture.

[0044] Step A5: Terminal-side data preprocessing and feature construction On the terminal side, sliding window feature extraction is performed on the raw measurement data, followed by denoising, missing data filling, outlier correction, and normalization to generate a terminal-side feature vector, which is then uploaded to the edge side and the cloud.

[0045] In one embodiment, the construction process of multi-source data fusion and multi-model collaboration for medium- and long-term forecasting is described by detailing steps S102-S104 in Embodiment 1.

[0046] Step B1: Construct a unified feature representation for multi-source data. The historical multi-source data includes historical photovoltaic power output data, numerical weather prediction data, and climate index data; wherein, the numerical weather prediction data includes irradiance, temperature, wind speed, and cloud cover, and the climate index data includes seasonal climate index and long-term sunshine trend index.

[0047] Specifically, collecting historical photovoltaic power output data. Numerical weather forecast Long-term climate index Construct a unified feature vector from multiple data sources:

[0048] in, The feature transformation parameters are normalized and aligned with time to ensure that all data sources participate in modeling on the same time axis and scale, thus forming a unified feature representation. This represents the time from time tL to time t; L is the historical backtracking step size, tL is the window start time, and t is the current time; Step B2: For unified features SARIMA, Prophet, and other models were trained separately to obtain the prediction outputs of each model. :

[0049] in, Let m be the predicted output value of the m-th basic model at the target time t+h. For the function mapping relationship of the m-th basic model, This represents a unified feature representation sequence from time tL to time t; Let m be the set of all trainable parameters of the m-th base model. This represents a statistical linear model based on the differenced autoregressive moving average. XGBoost represents an additive model based on decomposable trends, seasonality, and holiday effects, while XGBoost represents an ensemble learning model based on gradient boosting decision trees.

[0050] Each model estimates its parameters by minimizing its respective loss function:

[0051] in, For the function mapping relationship of the m-th basic model, Let t represent the actual photovoltaic output at the target prediction time t+h, where t represents the time sample.

[0052] Step B3: Use LSTM and Transformer to model long sequence dependencies and generate predicted values:

[0053] The weighted loss function to be minimized during training is determined by the following formula:

[0054] in, , The loss weights are used for deep models. For the LSTM (Long Short-Term Memory) network model at the target time The predicted output value for t+h This represents the predicted output force of the Transformer model at the target time t+h; , These represent the computational logic of the LSTM and Transformer model architectures, respectively. , These represent the complete set of trainable parameters for the LSTM model and the Transformer model, respectively. Step B4: Achieve multi-model collaborative fusion based on fusion processing The outputs of each base model are used to form the input vector. Point predictions are obtained by learning the fusion weights through a second-level learner (such as linear regression / XGBoost): ; in, This represents the input feature vector composed of the predicted values ​​of all M base models; Let represent the independent prediction value of the m-th base model at the target time t+h, and M represent the total number of base models participating in the fusion. This represents the dot product of the weight vector and the feature vector. Let w represent the transpose of the fusion weight vector w, and b be the bias term; This represents the final predicted value of photovoltaic power output after stacking, generalization, and fusion. Furthermore, the model's posterior weights are used to weight each basic prediction result to obtain the probability prediction distribution:

[0055] in, Let represent the posterior probability density function of photovoltaic output at the target time t+h, given a historical dataset D. These serve as posterior weights for the model, enabling multi-model collaboration and uncertainty characterization. Let f(m) represent the predicted probability density function of the m-th base model's own output.

[0056] The point prediction results and the probability prediction distribution together constitute the medium- and long-term comprehensive prediction results of photovoltaic power output.

[0057] In one embodiment, a detailed description is provided based on step S105 in Embodiment 1 above, describing a hierarchical triggering model update mechanism based on prediction bias and changes in data distribution.

[0058] Step D1: Construct hierarchical triggering criteria Define the prediction bias of the medium- and long-term comprehensive forecast results of photovoltaic power output. Changes in data distribution with unified feature representation Construct hierarchical functions :

[0059] In the formula, The first error threshold, For the second error threshold, The first drift threshold, This is the second drift threshold.

[0060] Wherein, the first error threshold is less than the second error threshold, and the first drift threshold is less than the second drift threshold; When the prediction deviation of the medium- and long-term comprehensive forecast of photovoltaic power output is less than the first error threshold, and the change in the data distribution represented by the unified feature is less than the first drift threshold, it is determined to be at the first trigger level; when the prediction deviation of the medium- and long-term comprehensive forecast of photovoltaic power output is between the first error threshold and the second error threshold, or when the change in the data distribution represented by the unified feature is between the first drift threshold and the second drift threshold, it is determined to be at the second trigger level; when the prediction deviation of the medium- and long-term comprehensive forecast of photovoltaic power output is not less than the second error threshold, or when the change in the data distribution represented by the unified feature is not less than the second drift threshold, it is determined to be at the third trigger level.

[0061] D2, Level 1 Trigger: Performs lightweight parameter fine-tuning on the edge side. When trigger level = 1, the cloud model structure is preserved, and only on the edge side is the model fine-tuned based on the incremental dataset. Update local parameters using gradients:

[0062] Enables rapid tracking of slight distribution changes.

[0063] In the formula, This is the new set of parameters for the edge-side model after fine-tuning and updating, where η is the learning rate; This is the error function, used to measure the model's prediction bias on new datasets; The loss function L represents the loss function with respect to the model parameters. gradient, This represents the old set of parameters of the edge-side model before the current fine-tuning operation. This is a unified feature representation formed from the latest collected multi-source data. This indicates that the model parameters currently used to calculate the loss are the current parameter values ​​on the edge side.

[0064] D3, Level 2 Trigger: Perform cloud-edge collaborative model reweighting and structural adjustment. When trigger level = 2, utilize the errors of each model within the most recent time window. Recalculate the fusion weights:

[0065] It enables or disables some models based on the error mode (such as turning off models with severe performance degradation) to achieve adaptive adjustment of the fusion structure.

[0066] In the formula, The new fusion weights are the recalculated weights for the m-th base model. This represents the basic models involved in the integration. It is the reciprocal of the error of the j-th model.

[0067] D4, Level 3 Trigger: Execute full data retraining and policy update in the cloud. When trigger level = 3, it triggers full historical data retraining in the cloud. Retrain the deep model and fusion layer parameters:

[0068] It is also updated synchronously to edge and terminal nodes to form a new round of baseline models, thereby maintaining the accuracy and stability of the photovoltaic medium- and long-term prediction system during long-term operation.

[0069] In the formula, This represents the value of the independent variable that minimizes the objective function. This means that the object of optimization is all the trainable parameters of the model.

[0070] Example 2: Based on the same technical concept, Example 2 of this invention also provides a medium- and long-term photovoltaic power output prediction system integrating cloud-edge-device architecture, such as... Figure 2 As shown, it includes: architecture building module 210, preprocessing module 220, model prediction module 230, fusion processing module 240, and action execution module 250, wherein: Architecture building module 210 is used to build a cloud-edge-device three-layer collaborative prediction architecture, which includes a cloud for performing long-term fusion prediction, an edge for performing regional collaborative prediction, and a device for performing raw measurement data preprocessing and feature construction. Preprocessing module 220 is used to collect at least two types of data from historical multi-source data, and to perform time alignment, missing data repair and normalization on the collected multi-source data to form a unified feature representation; Model prediction module 230 is used to train multiple prediction models on the cloud and the edge side respectively based on the unified feature representation, and use each prediction model to perform prediction calculations to obtain corresponding multiple basic prediction results; The fusion processing module 240 is used to perform fusion processing on the multiple basic prediction results to generate a medium- and long-term comprehensive prediction result for photovoltaic power output. The action execution module 250 is used to construct a hierarchical triggering criterion based on the prediction deviation of the medium- and long-term comprehensive prediction results of photovoltaic power output and the data distribution changes of the unified feature representation. When the first triggering level is determined, the parameters are fine-tuned on the edge side based on the unified feature representation formed by the latest collected multi-source data. When the second triggering level is determined, the cloud and the edge side jointly execute the recalculation of model fusion weights and the adjustment of the prediction model combination structure. When the third triggering level is determined, the deployed multi-type prediction models are retrained, and the updated model parameters and model structure configuration are synchronized to the edge side and the terminal side.

[0071] Example 3: In one embodiment, Example 3 of the present invention also provides an electronic device; the electronic device may be a terminal, and its internal structure diagram may be as follows. Figure 3As shown. The electronic device includes a processor, memory, communication interface, display screen, and input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements the long-term photovoltaic power output prediction method based on the cloud-edge-device architecture described in any one of steps S101 to S105. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the device's casing, or an external keyboard, touchpad, or mouse.

[0072] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0073] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0074] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0075] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0076] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0077] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of the claims of the present invention pending approval.

Claims

1. A method for medium- and long-term prediction of photovoltaic power output integrating cloud-edge-device architecture, characterized in that, include: A three-layer collaborative prediction architecture is constructed, comprising a cloud for performing long-term fusion prediction, an edge for performing regional collaborative prediction, and a terminal for performing raw measurement data preprocessing and feature construction. Collect at least two types of data from historical multi-source data, and perform time alignment, missing data repair, and normalization on the collected multi-source data to form a unified feature representation; Based on the unified feature representation, multiple prediction models are trained on the cloud and the edge respectively, and prediction calculations are performed using each prediction model to obtain corresponding multiple basic prediction results. The multiple basic prediction results are fused to generate a medium- and long-term comprehensive prediction result for photovoltaic power output; Based on the prediction deviation of the medium- and long-term comprehensive forecast results of photovoltaic power output, and the changes in the data distribution of the unified feature representation, a hierarchical triggering criterion is constructed; when the first triggering level is determined, the parameters are fine-tuned on the edge side based on the unified feature representation formed by the latest multi-source data. When the trigger level is determined to be the second level, the weights of the collaborative execution model are recalculated and the combination structure of the prediction model is adjusted. When the trigger level is determined to be the third level, the deployed multi-class prediction models are retrained, and the updated model parameters and model structure configurations are synchronized to the edge side and the terminal side.

2. The method according to claim 1, characterized in that, When the cloud is used to perform long-cycle fusion prediction, it adopts a long sequence model based on an attention mechanism to model the long-term time-series dependency of the unified feature representation to generate a first predicted value, and adopts a seasonal time series model to model the periodicity of the historical photovoltaic power output data to generate a second predicted value; the first predicted value and the second predicted value are fused to generate the cloud-side basic prediction result.

3. The method according to claim 2, characterized in that, The attention-based long sequence models include: Transformer model and Longformer model; The seasonal time series models include: SARIMA model and Prophet model.

4. The method according to claim 1, characterized in that, When the edge side is used to perform regional collaborative prediction, it constructs a regional local prediction model based on the unified feature representation corresponding to each regional node, and performs consistency iteration based on the prediction error feedback between adjacent regional nodes, so that the local prediction results of each regional node are collaboratively corrected during the iteration process to generate the edge side basic prediction results.

5. The method according to claim 1, characterized in that, The historical multi-source data includes historical photovoltaic power output data, numerical weather forecast data, and climate index data; wherein, the numerical weather forecast data includes irradiance, temperature, wind speed, and cloud cover, and the climate index data includes seasonal climate index and long-term sunshine trend index.

6. The method according to claim 1, characterized in that, The prediction deviation based on the medium- and long-term comprehensive prediction results of photovoltaic power output, and the data distribution changes represented by the unified feature, are used to construct a hierarchical triggering criterion, which includes setting a first error threshold, a second error threshold, a first drift threshold, and a second drift threshold; wherein, the first error threshold is less than the second error threshold, and the first drift threshold is less than the second drift threshold; When the prediction deviation of the medium- and long-term comprehensive forecast of photovoltaic power output is less than the first error threshold, and the change in the data distribution represented by the unified feature is less than the first drift threshold, it is determined to be at the first trigger level; when the prediction deviation of the medium- and long-term comprehensive forecast of photovoltaic power output is between the first error threshold and the second error threshold, or when the change in the data distribution represented by the unified feature is between the first drift threshold and the second drift threshold, it is determined to be at the second trigger level; when the prediction deviation of the medium- and long-term comprehensive forecast of photovoltaic power output is not less than the second error threshold, or when the change in the data distribution represented by the unified feature is not less than the second drift threshold, it is determined to be at the third trigger level.

7. The method according to claim 1, characterized in that, When the second trigger level is determined, the recalculation of model fusion weights and adjustment of the prediction model combination structure jointly performed by the cloud and edge sides include: recalculating the fusion weights of each prediction model based on the prediction errors of each prediction model within the most recent preset time period; and performing the enabling, disabling, or replacement operations of prediction models based on the magnitude and trend of prediction errors of each prediction model; if the prediction error of a certain prediction model is higher than the preset error threshold in multiple consecutive update cycles, the fusion weight of the prediction model is reduced, or the prediction model is temporarily removed from the fusion structure; when the prediction performance of the prediction model recovers to the preset standard, it is added back to the fusion structure.

8. The method according to claim 1, characterized in that, The fusion processing of multiple basic prediction results includes: using the multiple basic prediction results as input features, training a fusion weight learning model to learn the fusion weights corresponding to each prediction model, and outputting point prediction results; and determining the posterior weights of the models based on the prediction performance of each prediction model on the validation dataset, and weighting each basic prediction result according to the posterior weights of the models to obtain a probability prediction distribution. Obtain a medium- to long-term comprehensive forecast of photovoltaic power output based on the point prediction results and the probability prediction distribution.

9. A medium- to long-term photovoltaic power output forecasting system integrating cloud-edge-device architecture, characterized in that, include: An architecture building module is used to construct a three-layer collaborative prediction architecture for cloud, edge, and terminal. The three-layer collaborative prediction architecture includes a cloud for performing long-term fusion prediction, an edge for performing regional collaborative prediction, and a terminal for performing raw measurement data preprocessing and feature construction. The preprocessing module is used to collect at least two types of data from historical multi-source data, and to perform time alignment, missing data repair and normalization on the collected multi-source data to form a unified feature representation. The model prediction module is used to train multiple prediction models on the cloud and the edge side respectively based on the unified feature representation, and to perform prediction calculations using each prediction model to obtain corresponding multiple basic prediction results. The fusion processing module is used to perform fusion processing on the multiple basic prediction results to generate a medium- and long-term comprehensive prediction result for photovoltaic power output. The action execution module is used to construct a graded triggering criterion based on the prediction deviation of the medium- and long-term comprehensive prediction results of photovoltaic power output and the data distribution changes of the unified feature representation; when the first triggering level is determined, the parameters are fine-tuned on the edge side based on the unified feature representation formed by the latest collected multi-source data. When the trigger level is determined to be the second level, the cloud and edge sides will work together to recalculate the model fusion weights and adjust the prediction model combination structure. When the trigger level is determined to be the third level, the deployed multi-class prediction models are retrained, and the updated model parameters and model structure configurations are synchronized to the edge side and the terminal side.

10. An electronic device, characterized in that, The electronic device includes: At least one processor; and a memory communicatively connected to said at least one processor; The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1-8.

Images