Photovoltaic power generation prediction system and method based on smart power grid

Abstract

According to the photovoltaic power generation prediction system and method based on the smart power grid provided by the invention, the efficient fitting advantage of a gradient boosting decision tree model on high-dimensional static characteristics is utilized under a long-period stable weather condition, and the capturing capability of a long-short-term memory network on time sequence nonlinear dynamics is fully exerted in a weather sudden change period; weather categories are judged in real time and weights are dynamically adjusted through an environmental condition classifier, so that the RMSE of a final prediction result in four typical scenes of sunny days, cloudy days, cloudy days and rainy days is obviously reduced, and the RMSE is obviously superior to that of an existing fixed weight combination model; meanwhile, through an automatic data cleaning chain formed by an isolation forest and Lagrange interpolation, the data anomaly rate is controlled to be 0.1% or below, and the dimensional difference is eliminated by adopting minimum-maximum standardization, so that the model training stability is ensured; the parallel hybrid architecture realizes minute-level rolling prediction under the acceleration of the GPU, and the real-time requirement of a provincial power grid dispatching center is met.

Description

Technical Field

[0001] This invention relates to the field of distributed energy management technology for smart grids, and in particular to a photovoltaic power generation prediction system and method based on smart grids. Background Technology

[0002] With the advancement of the "dual carbon" target, the penetration rate of renewable energy, represented by photovoltaics, in the power grid is rapidly increasing. To ensure the safe and economical operation of a high-proportion renewable energy power grid, the dispatch center needs to accurately predict the power generation of each photovoltaic power station for the next few hours to days. Existing technologies typically break down the prediction task into three categories: physical models, statistical models, and machine learning models. Physical models use numerical weather prediction (NWP) as input and combine it with photovoltaic array models for simulation. Their advantage is strong interpretability, but they are insufficient in characterizing micrometeorological processes such as clouds and aerosols, leading to a significant amplification of errors when there are sudden weather changes such as cloudy or rainy weather. Statistical models, such as ARIMA and Kalman filtering, only use historical power sequences for linear extrapolation. They can achieve good results under clear weather conditions, but they are difficult to capture nonlinear changes. Machine learning models, such as BP neural networks and support vector regression, are beginning to introduce multi-dimensional features, but the training sample size is limited, and they mostly use a single model, resulting in poor adaptability to different weather types.

[0003] To compensate for the shortcomings of single models, a "combined forecasting" approach has emerged in recent years. One approach uses simple weighted averaging or fixed-weight voting, which cannot dynamically adjust weights based on weather conditions. Another approach updates weights online using meta-learning or Bayesian methods, improving dynamism but incurring high computational costs, making it difficult to meet the minute-level rolling forecasting requirements of provincial dispatching systems. Furthermore, existing systems generally suffer from data silos: SCADA data from the power grid dispatching center, data from local monitoring systems of photovoltaic power plants, and NWP data released by the meteorological bureau are inconsistent in sampling frequency, time base, and data format, making it difficult to unify model inputs. Simultaneously, handling outliers and missing values ​​still relies on manual experience, lacking automated data cleaning processes. The combined impact of these problems is that during periods of stable weather, the root mean square error (RMSE) of the forecast can be controlled within 5%, while during periods of drastic weather changes, the error rapidly amplifies to 15%-25%, directly leading to dispatching plan deviations, increased reserve capacity, and the risk of power curtailment for renewable energy sources.

[0004] Therefore, there is an urgent need in this field for a photovoltaic power generation prediction method and system that can fully utilize multi-source heterogeneous data from smart grids, adaptively adjust model weights for different weather types, and has high computational efficiency, so as to achieve refined scheduling of high-proportion renewable energy grids. Summary of the Invention

[0005] The purpose of this invention is to provide a photovoltaic power generation prediction system and method based on smart grids to solve the problems existing in the prior art.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] This invention provides a photovoltaic power generation prediction method based on a smart grid, comprising the following steps:

[0008] S1. Data acquisition and spatiotemporal alignment: real-time acquisition of multi-source heterogeneous data from the smart grid dispatch center and the local monitoring system of the photovoltaic power station; timestamp alignment of the acquired multi-source heterogeneous data to ensure that all data have a unified sampling resolution and synchronization benchmark in the time dimension.

[0009] S2. Data Preprocessing and Feature Engineering: The aligned data is cleaned, outlier detection and removal are performed using the isolated forest algorithm, and missing data is filled using Lagrange interpolation to generate a standardized dataset. Feature engineering is performed on the standardized dataset to construct a feature matrix for predicting model inputs. The feature matrix contains original data features, as well as time-derived features and physical constraint features generated by transforming the original data. All input features are normalized by using the min-max normalization method to linearly map the feature data of each dimension to the [0,1] interval to eliminate the influence of differences in feature dimensions on the model training process.

[0010] S3. Parallel Hybrid Prediction Model Construction: Based on the feature matrix, two independent sub-models, the first and second, are constructed and trained in parallel. The first sub-model uses a long short-term memory network model, taking historical photovoltaic power data sequences and time-derived features as inputs to capture the deep temporal dependence and nonlinear dynamic change law of photovoltaic power generation, and outputs the first predicted power sequence. The second sub-model uses a gradient boosting decision tree model, taking numerical weather forecast data, smart grid operation status data, and physical constraint features as inputs to mine the complex mapping relationship between high-dimensional static features and photovoltaic output power, and outputs the second predicted power sequence.

[0011] S4. Dynamic classification of environmental conditions: Construct and train an environmental condition classifier. The environmental condition classifier uses the total solar radiation intensity, relative humidity and wind speed in the numerical weather forecast data as input features and classifies them into one of a number of preset typical meteorological categories. The typical meteorological categories include at least sunny, cloudy, overcast and rainy. The environmental condition classifier outputs a corresponding meteorological category label at each forecast time point.

[0012] S5. Adaptive weighted fusion prediction: Based on the meteorological category labels output in step S4, a weight coefficient is dynamically calculated for each prediction time point. The calculation logic of the weight coefficient is as follows: when the meteorological category label is a category with relatively stable radiation changes such as sunny or cloudy, the weight assigned to the gradient boosting decision tree model is increased; when the meteorological category label is a category with drastic power fluctuations such as cloudy or rainy, the weight assigned to the long short-term memory network model is increased. The final photovoltaic power generation prediction value is calculated using a dynamic weighted fusion formula. The calculated final photovoltaic power generation prediction power sequence is output as the prediction result to the smart grid dispatch center.

[0013] Preferably, in step S1, the multi-source heterogeneous data includes at least: historical photovoltaic power data sequences, numerical weather forecast data, and smart grid operation status data.

[0014] Preferably, in step S3, the construction of the long short-term memory network model includes: defining the unit structure of the long short-term memory network, where each unit contains a forget gate, an input gate, and an output gate.

[0015] Preferably, in step S3, the training objective function of the gradient boosting decision tree model consists of a loss function and a regularization term, as shown in the following formula:

[0016] Obj=ΣL(y i ,y hati )+ΣΩ(f k );

[0017] Among them, y i Let y be the true power value of the i-th sample. hati Let f be the model's predicted value for the i-th sample, and L be the loss function used to measure the difference between the predicted and actual values, employing the mean squared error function; k Let be the function representation of the k-th decision tree in the model, and Ω be the regularization term used to control the complexity of the model, defined as Ω(f) = γT + (1 / 2)λ||w|| 2 , where T is the number of leaf nodes in the decision tree, w is the weight vector of the leaf nodes, and γ and λ are regularization coefficients.

[0018] Preferably, in step S4, the environmental condition classifier is a classification model based on support vector machine, which is trained by supervised learning using historically labeled meteorological data. The training process aims to find an optimal hyperplane to maximize the interval between samples of different meteorological categories.

[0019] This invention also provides a photovoltaic power generation prediction system based on a smart grid, including a processor and a memory. The memory stores computer program instructions that are executed by the processor. When the processor is configured to execute the computer program instructions, it performs the functions of the following modules:

[0020] The data acquisition and spatiotemporal alignment module is configured to communicate with the smart grid dispatch center and the local monitoring system of the photovoltaic power station. It is used to acquire historical photovoltaic power data sequences, numerical weather forecast data and smart grid operation status data in real time, and to perform timestamp alignment processing on the acquired multi-source heterogeneous data to ensure that all data have a unified sampling resolution and synchronization benchmark in the time dimension.

[0021] The data preprocessing and feature engineering module is configured to clean, impute missing values, and perform feature engineering on the aligned data. Data cleaning uses outlier detection and removal based on the isolated forest algorithm, missing value impute uses the Lagrange interpolation method, and feature engineering is used to construct the feature matrix required for the prediction model.

[0022] The parallelized hybrid prediction model building module is configured to build and train a long short-term memory network model and a gradient boosting decision tree model in parallel based on the feature matrix. The long short-term memory network model is used to capture the time series features of photovoltaic power and output the first predicted power sequence, while the gradient boosting decision tree model is used to mine the mapping relationship between high-dimensional static features and photovoltaic output power and output the second predicted power sequence.

[0023] The dynamic classification module for environmental conditions is configured to build and run an environmental condition classifier, which outputs the corresponding meteorological category label at each forecast time point based on numerical weather forecast data.

[0024] The adaptive weighted fusion prediction module is configured to dynamically calculate a weight coefficient for each prediction time point based on the meteorological category label, and use the dynamic weight coefficient to weight and fuse the first and second prediction power sequences to calculate the final photovoltaic power generation prediction value, and send the final prediction result to the smart grid dispatch center.

[0025] Preferably, the data preprocessing and feature engineering module is further configured to perform min-max normalization on all input features, linearly mapping the feature data of each dimension to a preset numerical range, so as to eliminate the impact of differences in the dimensions of different features on the model performance.

[0026] Preferably, in the parallelized hybrid prediction model building module, the long short-term memory network model contains multiple stacked LSTM layers. The unit structure of each LSTM layer includes a forget gate, an input gate, and an output gate. The three gates work together to control the flow and update of information in the cell state, thereby effectively learning long-term dependencies.

[0027] Preferably, in the parallelized hybrid prediction model construction module, the gradient boosting decision tree model is implemented using a histogram-based optimization algorithm. By discretizing continuous floating-point feature values ​​into k integer buckets to construct a feature histogram, the calculation of the optimal split point is accelerated by the statistical information of the histogram, significantly reducing memory usage and computational complexity.

[0028] Preferably, in the adaptive weighted fusion prediction module, the calculation logic of the weight coefficients is stored as a lookup table or function. This lookup table or function accurately maps each meteorological category label output by the environmental condition dynamic classification module to a preset specific weight value between 0 and 1, thereby achieving a fast and deterministic allocation of fusion weights.

[0029] The present invention achieves the following beneficial technical effects compared to the prior art:

[0030] This invention provides a photovoltaic power generation prediction system and method based on a smart grid. Under long-term stable weather conditions, it leverages the advantages of gradient boosting decision tree models in efficiently fitting high-dimensional static features. During periods of drastic weather changes, it fully utilizes the ability of long short-term memory networks to capture temporal nonlinear dynamics. An environmental condition classifier determines the weather category in real time and dynamically adjusts the weights, resulting in a significant reduction in the RMSE of the final prediction results under four typical scenarios: sunny, cloudy, overcast, and rainy days. This is significantly better than existing fixed-weight combination models. Simultaneously, an automated data cleaning chain consisting of isolated forests and Lagrange interpolation controls the data anomaly rate to below 0.1%, and min-max normalization is used to eliminate dimensional differences, ensuring stable model training. The parallel hybrid architecture, accelerated by GPUs, achieves minute-level rolling predictions, meeting the real-time requirements of provincial power grid dispatch centers. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 The flowchart of the photovoltaic power generation prediction method based on smart grid provided by the present invention is shown. Detailed Implementation

[0033] The serial numbers assigned to components in this document, such as "first" and "second," are used only to distinguish the described objects and have no sequential or technical meaning. The terms "connection" and "linkage" used in this application, unless otherwise specified, include both direct and indirect connections (linkages) and should not be construed as limiting the invention.

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] The purpose of this invention is to provide a photovoltaic power generation prediction system and method based on smart grids to solve the problems existing in the prior art.

[0036] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0037] Example 1:

[0038] This embodiment provides a photovoltaic power generation prediction method based on a smart grid, such as... Figure 1 As shown, it includes:

[0039] Step S1: Data Acquisition and Spatiotemporal Alignment

[0040] The smart grid dispatch center pushes grid frequency, key node voltage amplitude and regional load demand in real time through the IEC-104 protocol, with a sampling period of 1 minute; the local monitoring system of photovoltaic power plants uploads historical photovoltaic power data using the Modbus-TCP protocol, with a sampling period of 5 minutes; the meteorological bureau issues numerical weather forecasts through the FTP interface, with a time resolution of 15 minutes.

[0041] Among them, the historical photovoltaic power data sequence is defined as {P hist(t-n) ,...,P hist(t-1)}, where P hist The historical actual output power of the photovoltaic power station is t, where t is the current time and n is the preset length of the historical time series.

[0042] Numerical weather forecast data is defined as {NWP(t+1),...,NWP(t+m)}, where NWP is a multidimensional weather forecast vector containing the total solar radiation intensity, ambient temperature, relative humidity, wind speed and wind direction for the next m forecast time steps, and m is the forecast duration;

[0043] The smart grid operation status data is defined as {Grid(tk),...,Grid(t-1)}, where Grid is the grid state vector, which includes the grid frequency, key node voltage amplitude and regional load demand for k historical time steps.

[0044] The system uses a unified time base (UTC+8) to perform linear interpolation resampling on multi-source data to a 5-minute granularity, thus completing timestamp alignment.

[0045] Step S2: Data Preprocessing and Feature Engineering

[0046] An isolated forest algorithm was used to detect anomalies in the three-dimensional data of power, meteorology, and power grid status, with an anomaly threshold set at 0.05. Detected anomalies were filled using Lagrange interpolation with an interpolation window length of 7. Subsequently, all features were normalized to the minimum-maximum range, mapping to the interval [0,1]. In the feature engineering stage, in addition to the original variables, time-derived and physically constrained features such as "past 1-hour moving average power," "cosine of solar altitude angle," and "wind speed × solar radiation coupling term" were constructed, forming a matrix X of dimension D.

[0047] Step S3: Parallel Hybrid Prediction Model Construction

[0048] In the Long Short-Term Memory (LSTM) sub-model, two stacked LSTM layers are set up, each with 128 hidden units, and the time step n is 12 (i.e., 1h). The input gate, forget gate, output gate weight matrices W and bias vector b are initialized using Xavier.

[0049] In this embodiment, the forget gate (f t The information discarded from the cell state is determined by the following formula: f t =σ(W f ·[h t-1 ,x t ]+b f Input gate (i) t The following formula determines how much new information should be added to the cell state: i t =σ(W i ·[h t-1 ,x t ]+b i And create a new candidate value vector C through the tanh layer. tilda =tanh(W c ·[h t-1 ,x t ]+b c Update cell state: C t =f t *C t-1 +i t *C tilda Output gate (o) t The hidden state of the output is determined by the following formula: t =σ(W o ·[h t-1 ,x t]+b o The final hidden state h t =o t *tanh(C t ); where σ is the Sigmoid activation function, W and b are the weight matrix and bias vector of the corresponding gate, respectively, and h t-1 x is the hidden state from the previous time step. t This represents the input features at the current moment.

[0050] The gradient boosting decision tree sub-model is implemented using LightGBM, and the formula is as follows:

[0051] Obj=ΣL(y i ,y hati )+ΣΩ(f k );

[0052] Among them, y i Let y be the true power value of the i-th sample. hati Let f be the model's predicted value for the i-th sample, and L be the loss function used to measure the difference between the predicted and actual values, employing the mean squared error function; k Let be the function representation of the k-th decision tree in the model, and Ω be the regularization term used to control the complexity of the model, defined as Ω(f) = γT + (1 / 2)λ||w|| 2 Where T is the number of leaf nodes in the decision tree, w is the weight vector of the leaf nodes, and γ and λ are regularization coefficients; the number of decision trees is set to 800, the maximum depth is 8, the learning rate is 0.05, the loss function is mean squared error, and the regularization term Ω has γ = 0.1 and λ = 0.01. The two sub-models are trained in parallel on the GPU, the training set and validation set are divided in an 8:2 ratio, and the number of early stopping rounds is set to 50.

[0053] Step S4: Dynamic Classification of Environmental Conditions

[0054] A support vector machine classifier was trained using three-dimensional features: total solar radiation, relative humidity, and wind speed. The kernel function used was RBF with a penalty coefficient of 10 and a kernel width of 0.5. Training samples consisted of 15-minute meteorological observations and manually labeled weather categories (sunny, cloudy, overcast, rainy) from the past three years. The classifier outputs the weather category label for the current prediction time.

[0055] Step S5: Adaptive weighted fusion prediction

[0056] According to the preset lookup table, the weight coefficients for sunny, cloudy, overcast, and rainy days are set to 0.8, 0.7, 0.3, and 0.2, respectively. Based on the meteorological category label C(t) output in step S4, a weight coefficient α(t) is dynamically calculated for each prediction time point t. The calculation logic of the weight coefficient α(t) is as follows: when C(t) is a category with relatively stable radiation changes, such as sunny or cloudy, the weight assigned to the gradient boosting decision tree model is increased; when C(t) is a category with drastic power fluctuations, such as overcast or rainy days, the weight assigned to the long short-term memory network model is increased.

[0057] The final predicted photovoltaic power generation value P is calculated using the following dynamic weighted fusion formula. final(t) :P final(t) =α(t)*P LGBM(t) +(1-α(t))*P LSTM(t) Among them, P LGBM(t) To gradient boost the output value of the decision tree model at time point t, P LSTM(t) The output value of the Long Short-Term Memory network model at time point t; the calculated final photovoltaic power generation prediction sequence {P} final(t+1) ,...,P final(t+m) The prediction results are output to the smart grid dispatch center.

[0058] Example 2:

[0059] This embodiment provides a photovoltaic power generation prediction system based on a smart grid. The system hardware consists of two x86 servers, each configured with 2×NVIDIA A100 GPUs and 512GB RAM. On the software side, the data acquisition and spatiotemporal alignment module uses a Kafka message queue to access heterogeneous data streams; the data preprocessing and feature engineering module uses Python 3.9 + scikit-learn 1.1; the parallel hybrid prediction model construction module uses TensorFlow 2.9 to implement LSTM and LightGBM 3.3 to implement GBDT; the environmental condition dynamic classification module calls LIBSVM 3.25; and the adaptive weighted fusion prediction module is implemented in C++ with a latency of <100ms.

[0060] Specifically, including:

[0061] The data acquisition and spatiotemporal alignment module is configured to communicate with the smart grid dispatch center and the local monitoring system of the photovoltaic power station. It is used to acquire historical photovoltaic power data sequences, numerical weather forecast data and smart grid operation status data in real time, and to perform timestamp alignment processing on the acquired multi-source heterogeneous data to ensure that all data have a unified sampling resolution and synchronization benchmark in the time dimension.

[0062] The data preprocessing and feature engineering module is configured to clean, impute missing values, and perform feature engineering on the aligned data. Data cleaning uses outlier detection and removal based on the isolated forest algorithm, missing value impute uses Lagrange interpolation, and feature engineering is used to construct the feature matrix required for the prediction model. It is also configured to perform min-max normalization on all input features, linearly mapping the feature data of each dimension to a preset numerical range to eliminate the impact of differences in feature units on model performance.

[0063] The parallelized hybrid prediction model building module is configured to construct and train a Long Short-Term Memory (LSTM) network model and a Gradient Boosting Decision Tree (GBR) model in parallel based on the feature matrix. The LSTM network model captures the time-series features of photovoltaic power and outputs a first predicted power sequence, while the GBR model mines the mapping relationship between high-dimensional static features and photovoltaic output power, outputting a second predicted power sequence. The LSTM network model contains multiple stacked LSTM layers, each with a forget gate, input gate, and output gate. These three gates work together to control the flow and updating of information in the cell state, effectively learning long-term dependencies. The GBR model employs a histogram-based optimization algorithm, discretizing continuous floating-point feature values ​​into k integer buckets to construct a feature histogram. When calculating the optimal split point, the statistical information of the histogram is used to accelerate the process, significantly reducing memory usage and computational complexity.

[0064] The dynamic classification module for environmental conditions is configured to build and run an environmental condition classifier, which outputs the corresponding meteorological category label at each forecast time point based on numerical weather forecast data.

[0065] The adaptive weighted fusion prediction module is configured to dynamically calculate weight coefficients for each prediction time point based on meteorological category labels, and use these dynamic weight coefficients to weight and fuse the first and second predicted power sequences to calculate the final photovoltaic power generation prediction value. The final prediction result is then sent to the smart grid dispatch center. The calculation logic of the weight coefficients is stored as a lookup table or function, which precisely maps each meteorological category label output by the dynamic classification module of environmental conditions to a preset specific weight value between 0 and 1, thereby achieving a fast and deterministic allocation of fusion weights.

[0066] The aforementioned system operated for 30 days in a provincial dispatch center, averaging 288 rolling forecasts per day with an average single forecast time of 42 seconds. Compared to existing fixed-weight combination models, it reduced the error by 41.7% on days with sudden changes in weather (rainy weather), reduced reserve capacity by 9.2%, and prevented any power curtailment events caused by forecasting errors.

[0067] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0068] It should be noted that the components mentioned in the above embodiments are all general standard parts or components known to those skilled in the art. Their structures and principles can be learned by those skilled in the art through technical manuals or conventional experimental methods.

[0069] This invention has illustrated its principles and implementation methods using specific examples. The descriptions of these embodiments are merely illustrative of the method and its core ideas; furthermore, those skilled in the art will recognize that modifications may be made to the specific implementation methods and application scope based on the principles of this invention. Therefore, the content of this specification should not be construed as limiting the invention.

Claims

1. A photovoltaic power generation prediction method based on smart grids, characterized in that, Includes the following steps: S1. Data acquisition and spatiotemporal alignment: real-time acquisition of multi-source heterogeneous data from the smart grid dispatch center and the local monitoring system of the photovoltaic power station; timestamp alignment of the acquired multi-source heterogeneous data to ensure that all data have a unified sampling resolution and synchronization benchmark in the time dimension. S2. Data Preprocessing and Feature Engineering: The aligned data is cleaned, outlier detection and removal are performed using the isolated forest algorithm, and missing data is filled using Lagrange interpolation to generate a standardized dataset. Feature engineering is performed on the standardized dataset to construct a feature matrix for predicting model inputs. The feature matrix contains original data features, as well as time-derived features and physical constraint features generated by transforming the original data. All input features are normalized by using the min-max normalization method to linearly map the feature data of each dimension to the [0,1] interval to eliminate the influence of differences in feature dimensions on the model training process. S3. Parallel Hybrid Prediction Model Construction: Based on the feature matrix, two independent sub-models, the first and second, are constructed and trained in parallel. The first sub-model uses a long short-term memory network model, taking historical photovoltaic power data sequences and time-derived features as inputs to capture the deep temporal dependence and nonlinear dynamic change law of photovoltaic power generation, and outputs the first predicted power sequence. The second sub-model uses a gradient boosting decision tree model, taking numerical weather forecast data, smart grid operation status data, and physical constraint features as inputs to mine the complex mapping relationship between high-dimensional static features and photovoltaic output power, and outputs the second predicted power sequence. S4. Dynamic classification of environmental conditions: Construct and train an environmental condition classifier. The environmental condition classifier uses the total solar radiation intensity, relative humidity and wind speed in the numerical weather forecast data as input features and classifies them into one of a number of preset typical meteorological categories. The typical meteorological categories include at least sunny, cloudy, overcast and rainy. The environmental condition classifier outputs a corresponding meteorological category label at each forecast time point. S5. Adaptive weighted fusion prediction: Based on the meteorological category label output in step S4, a weight coefficient is dynamically calculated for each prediction time point; The calculation logic for the weight coefficients is as follows: when the meteorological category label is sunny or cloudy with relatively stable radiation changes, the weights assigned to the gradient boosting decision tree model are increased; when the meteorological category label is cloudy or rainy with drastic power fluctuations, the weights assigned to the long short-term memory network model are increased; the final photovoltaic power generation prediction value is calculated using a dynamic weighted fusion formula; and the calculated final photovoltaic power generation prediction power sequence is output as the prediction result to the smart grid dispatch center.

2. The photovoltaic power generation prediction method based on smart grid according to claim 1, characterized in that, In step S1, the multi-source heterogeneous data includes at least: historical photovoltaic power data series, numerical weather forecast data, and smart grid operation status data.

3. The photovoltaic power generation prediction method based on smart grid according to claim 1, characterized in that, In step S3, the construction of the long short-term memory network model includes: defining the unit structure of the long short-term memory network, where each unit contains a forget gate, an input gate, and an output gate.

4. The photovoltaic power generation prediction method based on smart grid according to claim 1, characterized in that, In step S3, the training objective function of the gradient boosting decision tree model consists of a loss function and a regularization term, as shown in the formula: Obj=ΣL(y i ,and hati )+ΣΩ(f k ); Among them, y i Let y be the true power value of the i-th sample. hati Let f be the model's predicted value for the i-th sample, and L be the loss function used to measure the difference between the predicted and actual values, employing the mean squared error function; k Let be the function representation of the k-th decision tree in the model, and Ω be the regularization term used to control the complexity of the model, defined as Ω(f) = γT + (1 / 2)λ||w|| 2 , where T is the number of leaf nodes in the decision tree, w is the weight vector of the leaf nodes, and γ and λ are regularization coefficients.

5. The photovoltaic power generation prediction method based on smart grid according to claim 1, characterized in that, In step S4, the environmental condition classifier is a classification model based on support vector machines. It is trained through supervised learning using historically labeled meteorological data. The training process aims to find an optimal hyperplane to maximize the margin between samples of different meteorological categories.

6. A photovoltaic power generation prediction system based on a smart grid, comprising a processor and a memory, wherein the memory stores computer program instructions to be executed by the processor, characterized in that, When a processor is configured to execute computer program instructions, it performs the functions of the following modules: The data acquisition and spatiotemporal alignment module is configured to communicate with the smart grid dispatch center and the local monitoring system of the photovoltaic power station. It is used to acquire historical photovoltaic power data sequences, numerical weather forecast data and smart grid operation status data in real time, and to perform timestamp alignment processing on the acquired multi-source heterogeneous data to ensure that all data have a unified sampling resolution and synchronization benchmark in the time dimension. The data preprocessing and feature engineering module is configured to clean, impute missing values, and perform feature engineering on the aligned data. Data cleaning uses outlier detection and removal based on the isolated forest algorithm, missing value impute uses the Lagrange interpolation method, and feature engineering is used to construct the feature matrix required for the prediction model. The parallelized hybrid prediction model building module is configured to build and train a long short-term memory network model and a gradient boosting decision tree model in parallel based on the feature matrix. The long short-term memory network model is used to capture the time series features of photovoltaic power and output the first predicted power sequence, while the gradient boosting decision tree model is used to mine the mapping relationship between high-dimensional static features and photovoltaic output power and output the second predicted power sequence. The dynamic classification module for environmental conditions is configured to build and run an environmental condition classifier, which outputs the corresponding meteorological category label at each forecast time point based on numerical weather forecast data. The adaptive weighted fusion prediction module is configured to dynamically calculate a weight coefficient for each prediction time point based on the meteorological category label, and use the dynamic weight coefficient to weight and fuse the first and second prediction power sequences to calculate the final photovoltaic power generation prediction value, and send the final prediction result to the smart grid dispatch center.

7. The photovoltaic power generation prediction system based on a smart grid according to claim 6, characterized in that, The data preprocessing and feature engineering module is also configured to perform min-max normalization on all input features, linearly mapping the feature data of each dimension to a preset numerical range to eliminate the impact of differences in feature dimensions on model performance.

8. The photovoltaic power generation prediction system based on a smart grid according to claim 6, characterized in that, In the parallelized hybrid prediction model building module, the long short-term memory network model contains multiple stacked LSTM layers. The unit structure of each LSTM layer includes a forget gate, an input gate, and an output gate. The three gates work together to control the flow and update of information in the cell state, thereby effectively learning long-term dependencies.

9. The photovoltaic power generation prediction system based on a smart grid according to claim 6, characterized in that, In the parallelized hybrid prediction model building module, the gradient boosting decision tree model is implemented using a histogram-based optimization algorithm. By discretizing continuous floating-point feature values ​​into k integer buckets to construct a feature histogram, the calculation of the optimal split point is accelerated by the statistical information of the histogram, which significantly reduces memory usage and computational complexity.

10. The photovoltaic power generation prediction system based on a smart grid according to claim 6, characterized in that, In the adaptive weighted fusion prediction module, the calculation logic of the weight coefficients is stored as a lookup table or function. This lookup table or function accurately maps each meteorological category label output by the environmental condition dynamic classification module to a preset specific weight value between 0 and 1, thereby achieving a fast and deterministic allocation of fusion weights.

Images