Photovoltaic output hybrid probability interval prediction method and system based on parallel deep learning architecture

Abstract

The present invention discloses a photovoltaic output hybrid probability interval prediction method and system based on a parallel deep learning architecture, constructing a photovoltaic output multi-source driving factor set; constructing an original feature matrix based on the photovoltaic output multi-source driving factor set; performing parallel decoupling of the spatiotemporal features of the original feature matrix, inputting the decoupled time features and spatial features into a spatiotemporal feature complementary enhancement module for fusion; inputting the feature-enhanced photovoltaic output spatiotemporal feature matrix into a photovoltaic output benchmark prediction model to obtain a photovoltaic output benchmark prediction result and a corresponding error; inputting the photovoltaic output benchmark prediction result error into a risk-type prediction model respectively, calculating the prediction error risk interval boundary value, and superimposing it on the benchmark prediction result to obtain respective photovoltaic output risk-type prediction results; constructing a photovoltaic output hybrid risk-type prediction framework, inputting the results of two risk-type prediction models for coupling and optimization, and obtaining a photovoltaic output hybrid risk-type prediction result.

Description

Technical Field

[0001] The present invention belongs to the interdisciplinary field of photovoltaic output prediction and artificial intelligence technology, and specifically relates to a photovoltaic output hybrid probability interval prediction method and system based on a parallel deep learning architecture. Background Art

[0002] Current PV output forecasting methods exhibit multi-dimensional, multi-model coupling characteristics, with physical models, statistical methods, machine learning, and hybrid architectures tightly linked through data networks. Traditional benchmark forecasting methods have significant limitations when dealing with complex meteorological conditions. When multi-dimensional spatiotemporal characteristics dynamically interact, single-point forecasting models struggle to quantify potential uncertainties, leading to risky misalignments between forecast results and actual output fluctuations. For example, during extreme weather events, the combination of physical modeling biases in numerical weather forecasts and the static assumptions of statistical models can easily lead to chain propagation of forecast errors, highlighting the urgent need to develop risk-based forecasting to enhance the robustness of grid dispatch.

[0003] There are three major problems with the existing forecasting technology system. First, PV output forecasting is affected by multiple factors, and most existing PV output forecasting studies only consider meteorological and PV output historical sequence factors, failing to fully account for factors influencing PV output. Second, mainstream benchmark forecasting methods lack a parallel decoupling mechanism for spatiotemporal features. While hybrid architectures integrate multi-model features, their stacked series design results in an incomplete decoupling of spatial topological correlations and temporal evolution patterns, hindering improvements in forecast accuracy. Finally, risk-based forecasting modeling has limitations, typically relying on a single model to infer risk probabilities, which can lead to insufficient risk interval coverage in complex scenarios.

[0004] Breaking through these bottlenecks requires two breakthroughs: first, building a parallel decoupling and complementary enhancement mechanism for spatiotemporal features, enabling multidimensional information fusion by deeply decoupling spatial topological correlations from dynamic temporal evolution; and second, developing a hybrid risk-based forecasting framework that integrates the advantages of kernel density estimation for non-distributed assumptions with the temporal dependency capture capabilities of quantile regression. Despite the increasing research on hybrid models, their parallel decoupling architecture design and uncertainty quantification techniques are still immature, making it difficult to meet the power grid system's demand for efficient and accurate simulation. Summary of the Invention

[0005] In order to solve the problems existing in the prior art, the present invention provides a photovoltaic output hybrid probability interval prediction method and system based on a parallel deep learning architecture, which realizes photovoltaic output risk prediction through the integration of regional photovoltaic output multi-source driving factor sets, construction of original feature matrix, parallel decoupling and complementary enhancement module of spatiotemporal features, photovoltaic output benchmark prediction, basic risk prediction, and hybrid risk prediction framework.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A photovoltaic output hybrid probability interval prediction method based on a parallel deep learning architecture, the method comprising:

[0008] S1: Construct a set of multi-source driving factors for photovoltaic output;

[0009] S2: Synchronize the multi-source driving factor set of photovoltaic output under a unified spatiotemporal reference by spatial reference and time dimension, and construct an original feature matrix;

[0010] S3: Construct a feature extraction module based on a parallel deep learning architecture, perform parallel decoupling of the temporal and spatial features of the original feature matrix, and input the decoupled temporal and spatial features into the temporal and spatial feature complementary enhancement module for fusion;

[0011] S4: Construct a photovoltaic output benchmark prediction model based on sparse attention Transformer, input the photovoltaic output spatiotemporal feature matrix after feature enhancement into the photovoltaic output benchmark prediction model, and obtain the photovoltaic output benchmark prediction result and corresponding error;

[0012] S5: Construct a risk-based prediction model based on data-driven and statistical theory. Input the error of the PV output benchmark prediction result into the two risk-based prediction models respectively, calculate the boundary value of the prediction error risk interval, and superimpose it on the benchmark prediction result to obtain the respective PV output risk-based prediction results.

[0013] S6: Construct a hybrid risk-based prediction framework for photovoltaic output, input the results of two risk-based prediction models for coupling, and use a multi-objective optimization algorithm to optimize the coupling parameters to obtain the hybrid risk-based prediction results for photovoltaic output.

[0014] Preferably, the step S1, constructing a set of multi-source driving factors for photovoltaic output, includes:

[0015] S11: Define the research area and systematically collect multi-source heterogeneous data including historical output data of photovoltaic stations, numerical weather forecast data, meteorological data, topographic data, and physical parameters of station equipment;

[0016] S12: constructing a geographic data adapter to obtain the elevation-irradiance correlation matrix of regional photovoltaic stations;

[0017] S13: Build a meteorological data adapter to convert the raw meteorological data into a format that meets the model input and align the meteorological data with the elevation grid of the geographic data;

[0018] S14: Build a device data adapter to convert various device parameters into a form that meets the model input;

[0019] S15: Construct a photovoltaic output simulator based on physical mechanisms, and use the photovoltaic output simulation value driven by the physical mechanism as the input feature of the model.

[0020] Preferably, in S3, a feature extraction module based on a parallel deep learning architecture is constructed to perform parallel decoupling of the spatiotemporal features of the original feature matrix, and the decoupled temporal features and spatial features are input into a spatiotemporal feature complementary enhancement module for fusion, including:

[0021] S31: Construct a parallel decoupling module for spatiotemporal features. The GRU model (TGRU) based on the improved temporal attention mechanism is used to extract the temporal features from the original PV output feature matrix. The spatial features in the original feature matrix are extracted by the convolutional attention module (ICBAM) based on the improved multi-scale recursive attention and dynamic weight coupling mechanism.

[0022] S32: Construct a spatiotemporal feature complementary enhancement module to fuse the temporal and spatial features after decoupling TGRU and ICBAM;

[0023] Among them, TGRU includes:

[0024] Update Gate

[0025] ;

[0026] In the formula, z represents the update gate output, l represents the original feature matrix of the input, σ represents the Sigmoid function, and h prev Represents the hidden state of the previous moment, W z Represents the weight parameter of the update gate;

[0027] Reset Gate

[0028] ;

[0029] Where re represents the reset gate output, W r Represents the weight parameter of the reset gate;

[0030] Temporal Attention

[0031] ;

[0032] ;

[0033] Where, Represents the temporal attention score, U a represents the hidden state projection matrix, V a Represents the attention score matrix, W a Represents the weight parameter of the attention mechanism, l attn Represents the characteristics after the attention mechanism is strengthened;

[0034] Candidate status

[0035] ;

[0036] Where h tilde represents the candidate hidden state, tanh is the hyperbolic tangent function;

[0037] Status Update

[0038] ;

[0039] Where h new Represents the updated hidden state. The hidden state at the final moment is the global representation of the time dimension feature of the photovoltaic output feature set l TGRU ;

[0040] ICBAM includes: multi-scale recursive attention and dynamic weight coupling mechanism;

[0041] Among them, the multi-scale recursive attention includes: multi-scale channel attention module and multi-receptive field spatial attention module;

[0042] The multi-scale channel attention module includes:

[0043] Multi-scale pooling layer:

[0044] Global average pooling: ;

[0045] Global max pooling: ;

[0046] Where td represents the channel index, l avg Represents the global average value of the feature of channel td, Row represents the height of the feature map, Col represents the width of the feature map, row and col are the position index l td,row,col Represents the characteristic value of row row col column in channel td, l max Represents the maximum response value among all eigenvalues of the channel;

[0047] Shared MLP structure: using two layers of full connection, the formula is:

[0048] ;

[0049] ;

[0050] In the formula, MLP(l avg )、MLP(l max ) represent the average and maximum pooling features after nonlinear transformation of MLP structure, are the weight matrices of each fully connected layer, δ is the ReLU activation function, b1 and b2 are the position offset control parameters of each fully connected layer;

[0051] Weight generation:

[0052] ;

[0053] Where M td (l) is the output channel weight matrix, TD is the channel dimension capacity;

[0054] Feature weighting:

[0055] ;

[0056] Where, l channel Represents the output of the channel attention module;

[0057] The multi-receptive field spatial attention module includes:

[0058] Channel Compression: Yes Perform average pooling and maximum pooling in the channel dimension to obtain , represents the average pooled value output by the channel attention module, Represents the maximum pooling value of the channel attention module output;

[0059] Feature stitching:

[0060] ;

[0061] Where, l cat Represents the channel compression feature after feature splicing;

[0062] Convolution kernel design:

[0063] ;

[0064] Where M s (l cat ) represents the concatenated features after convolution operation;

[0065] Spatial weighting:

[0066] ;

[0067] Where, l spatial Features representing the output of the spatial attention module;

[0068] Dynamic weight coupling mechanism:

[0069] A dynamic weight controller is designed to automatically adjust the fusion ratio of channel attention and spatial attention according to input features, improving the adaptability to different scenarios:

[0070] ;

[0071] Where, l ICBAM represents the spatial features extracted by ICBAM, is the coupling parameter of channel attention and spatial attention;

[0072] Construct a spatiotemporal feature complementary enhancement module to fuse the temporal and spatial features after decoupling TGRU and ICBAM, including:

[0073] After aligning the temporal and spatial feature channels, the concatenated features are fed into the 3D convolution kernel to capture the temporal and spatial local correlations:

[0074] ;

[0075] Where, l fused Represents the fusion feature after fusing time and space features;

[0076] The fused feature vector is enhanced using cross-modal residual connections, and the original temporal and spatial features are aligned through one-dimensional convolution and then superimposed on the fused features:

[0077] ;

[0078] Where, Represents the spatiotemporal feature matrix of photovoltaic output after feature enhancement.

[0079] Preferably, in S4, the method for constructing a photovoltaic output benchmark prediction model based on sparse attention Transformer includes:

[0080] The standard attention mechanism retained by the decoder:

[0081] ;

[0082] Where, is the query, key, and value matrix, d k is the dimension scaling factor;

[0083] Dynamic Sparse Attention for Encoder:

[0084] ;

[0085] ;

[0086] Where LocalWindow(P) represents the dynamic window centered at position P, GlobalAnchor(P) represents the global anchor point selected according to the importance score, M represents the attention mask matrix, and q represents the candidate key position within the dynamic window or within the range of the anchor point.

[0087] Preferably, S5: constructing a risk-based prediction model based on data-driven and statistical theory, inputting the photovoltaic output benchmark prediction result error into two risk-based prediction models respectively, calculating the prediction error risk interval boundary value, and superimposing it on the benchmark prediction result to obtain respective photovoltaic output risk-based prediction results, including:

[0088] S51: Build a data-driven risk-based prediction model. Based on the quantile regression framework, a multi-quantile loss function is constructed. During the time series modeling process, error quantiles corresponding to different risk levels are simultaneously output and superimposed on the baseline prediction results to obtain a data-driven risk-based prediction result for PV output.

[0089] S52: Construct a risk-based prediction model based on statistical theory. Use adaptive kernel functions and local bandwidth adjustment to improve the traditional kernel density estimation method, resulting in the IKDE method. Superimpose the prediction error risk interval boundaries on the baseline prediction results to obtain a statistically based risk-based prediction result for photovoltaic output.

[0090] Wherein, the S51 includes:

[0091] ;

[0092] Where obj1 represents the IQRLSTM objective function without considering the quantile constraint penalty term, y i Represents the actual observed value of photovoltaic output, Represents the photovoltaic output forecast value of the corresponding quantile, represents the quantile, and J represents the total number of quantiles;

[0093] The improved QRLSTM, i.e. IQRLSTM, takes into account the inherent properties of quantiles and adds a quantile constraint penalty term to the objective function.

[0094] ;

[0095] ;

[0096] ;

[0097] Where dif represents the sum of multi-quantile losses, penalty represents the constraint penalty term, and obj represents the final loss function;

[0098] The S52 includes:

[0099] The photovoltaic output benchmark prediction error is divided according to the timestamp;

[0100] Distance calculation:

[0101]

[0102] In the formula, fac1 and fac2 represent the elements to be calculated;

[0103] Use the binary tree search method to accelerate the nearest neighbor search and convert the photovoltaic output benchmark prediction error sequence Arrange in ascending order, n represents the sample size, the local neighborhood size is win, and for each error point err i , expand forward and backward in the sorted array points, the maximum distance between the error point and the neighborhood boundary point is ;

[0104] Dynamic bandwidth adjustment:

[0105] ;

[0106] ;

[0107] Where, Represents the local density sensitivity factor, which controls the expansion amplitude of the sparse area. represents the global smoothing factor, h global represents the global bandwidth, represents the standard deviation of the data, h i is bandwidth;

[0108] Adaptive kernel function, using improved Gaussian kernel function:

[0109] .

[0110] Preferably, in S6, a photovoltaic output hybrid risk-type prediction framework is constructed, two risk-type prediction model results are input for coupling, and a multi-objective optimization algorithm is used to optimize the coupling parameters to obtain photovoltaic output hybrid risk-type prediction results, including:

[0111] S61: Using the triple mutation strategy of "sine mapping-spiral flight-Lévy flight", an improved single-objective tribal competition and member cooperation optimization algorithm, namely ICTCM, is proposed;

[0112] S62: Use multi-objective adaptation mechanism reconstruction to achieve the transformation from single-objective to multi-objective optimization framework, namely MOICTCM;

[0113] S63: The risk-based prediction results of PV output are obtained by coupling the risk-based prediction results of IKDE and IQRLSTM through MOICTCM;

[0114] Wherein, the S62 includes:

[0115] S621: Multi-objective fitness evaluation and Pareto ranking:

[0116] Target vectorization:

[0117] ;

[0118] Where F represents the target vector, Ang represents the dimension of the target function, and f Ang represents the subgoal with the largest dimension;

[0119] Solution A dominates solution B if and only if:

[0120] ;

[0121] Where Ang represents the objective function dimension, f ang () represents each sub-objective function;

[0122] Define the Pareto dominance relationship: calculate the non-dominated relationship of all solutions, extract the first layer of non-dominated solutions, recursively extract the next layer of non-dominated solutions from the remaining solutions, and divide the population into multiple Pareto frontier layers;

[0123] Within each level of the Pareto front, the crowding of solutions is calculated to maintain diversity:

[0124] ;

[0125] Where CD represents the individual crowding value, O Now represents the current solution in the target space, O Now−1 and O Now+1 Represent the adjacent solutions of the current solution in the target space;

[0126] S622: Multi-objective extension of tribal structure and competition rules:

[0127] Tribe division and role definition: Tribes are divided according to the non-dominated sorting results. Each tribe contains solutions at different Pareto levels. The non-dominated solution with the highest crowding degree in the tribe is the chief.

[0128] Dynamic adjustment of member loyalty:

[0129] ;

[0130] In the formula, te represents the current number of iterations, r te represents the loyalty of members during the te-th iteration, Represents the maximum number of iterations;

[0131] Multi-target velocity update formula:

[0132] ;

[0133] Where p best is the best historical position of the individual, g best For the tribal chief position, X random is a randomly selected non-dominated solution from other tribes, represents the speed of the mem-th member of the num-th tribe in the dim dimension in the te+1-th iteration, The inertia coefficient representing the velocity, Represents the position of the member in the teth iteration, c1 and c2 represent the experience factor and compliance factor respectively, 、 Represents the chaos loyalty of each member, c3 represents the escape factor, represents the chaotic random factor;

[0134] Integration of spiral flight strategy: Introducing spiral exploration into position update and searching near the Pareto frontier:

[0135] ;

[0136] Where, represents the logarithmic spiral shape constant, Represents a random number between [-1, 1];

[0137] S623: External Archive Management and Retention:

[0138] Archive update rule: merge the parent and child populations, filter non-dominated solutions and add them to the temporary archive. If the temporary archive exceeds the capacity SI archive , eliminate solutions from low to high according to the congestion degree;

[0139] Levy flight perturbation: Apply Levy perturbation to the archived solution every Ge generation, jump out of the local optimum on a global scale, and discover potential solutions:

[0140] ;

[0141] Where, Represent the archive groups in the te and te+1 iterations respectively, represents the step size parameter, 、 Obeying a random normal distribution, The value range is (0, 2], Represents the gama function;

[0142] Clash of Clans and Escape Mechanics:

[0143] Conflict triggering conditions: Every generation, two tribes are randomly selected. A and Tribe B , and compare their hypervolume contributions HV:

[0144] ;

[0145] Escape strategy: The escape strategy guides the losing tribe to learn from the winning tribe and accelerates convergence. The speed of the losing tribe members is updated as follows:

[0146] ;

[0147] Where, represents the speed of the failed tribe A member during the te-th iteration, represents the update speed of tribe A members, c4 represents the escape factor, Represents the update random factor, X B,chief Represents the position of the chief in the winning tribe B, represents the position of the tribe A member in the teth iteration;

[0148] The S63 includes:

[0149] ;

[0150] ;

[0151] ;

[0152] In the formula, PINAW represents the average bandwidth index of the prediction interval, PICP represents the coverage index of the prediction interval, represents the established risk level, ω represents the risk type prediction result coupling parameter of IKDE and IQRLSTM, Up mix 、Up IKDE 、Up IQRLSTM Respectively represent the upper bound of mixed risk type prediction, the upper bound of IKDE risk type prediction, the upper bound of IQRLSTM risk type prediction, Low mix 、Low KDE 、Low QRLSTM They represent the lower bounds of mixed risk type prediction, IKDE risk type prediction, and IQRLSTM risk type prediction, respectively. 、 Represents the maximum upper bound and minimum lower bound of the prediction interval under a given risk level, Dis represents the maximum range of the prediction interval boundary, 、 Represents the upper and lower bounds of the prediction interval at the i-th sample point under a given risk level. Mo is a Boolean value. When the actual photovoltaic output value is within the risk-type prediction interval, Mo takes the value of 1, otherwise it takes the value of 0. represents the Mo value corresponding to the i-th sample point under a given risk level. The meanings of other variables remain consistent with the previous text.

[0153] The present invention also provides a photovoltaic output hybrid probability interval prediction system based on a parallel deep learning architecture, the system is used to implement the aforementioned method, the system includes: a data set construction module, a matrix construction module, a feature fusion module, a first prediction module, a second prediction module and a third prediction module;

[0154] The data set construction module is used to construct a set of multi-source driving factors for photovoltaic output;

[0155] The matrix construction module is used to construct an original feature matrix by synchronizing the set of photovoltaic output multi-source driving factors under a unified time and space reference through spatial reference and time dimension and feature splicing;

[0156] The feature fusion module is used to construct a feature extraction module based on a parallel deep learning architecture, perform parallel decoupling of the spatiotemporal features of the original feature matrix, and input the decoupled temporal and spatial features into the spatiotemporal feature complementary enhancement module for fusion;

[0157] The first prediction module is used to construct a photovoltaic output benchmark prediction model based on sparse attention Transformer, input the photovoltaic output spatiotemporal feature matrix after feature enhancement into the photovoltaic output benchmark prediction model, and obtain the photovoltaic output benchmark prediction result and corresponding error;

[0158] The second prediction module is used to build a risk-based prediction model based on data-driven and statistical theory, input the photovoltaic output benchmark prediction result error into the two risk-based prediction models respectively, calculate the prediction error risk interval boundary value, and superimpose it on the benchmark prediction result to obtain the respective photovoltaic output risk-based prediction results;

[0159] The third prediction module is used to construct a photovoltaic output hybrid risk-type prediction framework, input the results of two risk-type prediction models for coupling, and use a multi-objective optimization algorithm to optimize the coupling parameters to obtain a photovoltaic output hybrid risk-type prediction result.

[0160] Compared with the prior art, the present invention has the following beneficial effects:

[0161] 1. Construct a set of multi-source driving factors for photovoltaic output:

[0162] By building a multi-source data adapter and a physics-based PV output simulator, we can effectively integrate a multi-dimensional set of influencing factors, including geographic location information, equipment parameters, physically simulated output, meteorological data, and historical PV output. Compared to traditional factor sets that only consider meteorological conditions and historical PV output, this multi-source factor set can more comprehensively reflect the various key factors affecting the power generation performance of PV power plants. Geographic location information determines the conditions and methods of solar radiation received at a specific location, while equipment parameters incorporate the electrical characteristics and mechanical installation conditions of PV modules into the analysis. The physics-based PV output simulator can combine actual meteorological data to provide simulation results with a certain degree of accuracy.

[0163] 2. Realize parallel decoupling and complementary enhancement of the spatiotemporal characteristics of photovoltaic output:

[0164] By synergistically integrating the complementary strengths of TGRU and ICBAM, a parallel decoupling architecture for spatiotemporal features is constructed, effectively separating the dynamic time series variations and spatial heterogeneity of PV output forecast inputs. Compared to traditional stacked hybrid architectures, which suffer from a single feature transfer path and difficulty achieving dynamic feedback, this architecture more accurately captures the complex dynamics of PV output influenced by sudden time-frequency fluctuations and spatial heterogeneity.

[0165] 3. Improve the timeliness of photovoltaic output benchmark forecasts:

[0166] The Sparse Attention Transformer uses sparse patterns to quickly lock onto key features, optimizing computational efficiency and focusing on dynamic features. This enhances response to sudden changes in scenarios and significantly improves the timeliness of baseline forecasts. The traditional Transformer's global self-attention algorithm has high computational complexity, making it difficult to handle the high-dimensional, long-term time series data (such as meteorological data and equipment status time series) required for photovoltaic output forecasts. Sparse Attention significantly reduces computational complexity by using local window constraints.

[0167] 4. Enhance the reliability of PV output risk prediction:

[0168] IKDE fits the probability density function of the baseline PV output forecast errors at the same historical timestamp and inversely solves the cumulative distribution function to obtain the upper and lower bounds of the error value. IQRLSTM constructs a multi-quantile loss function, synchronously outputting error quantiles corresponding to different risk levels during the time series modeling process. The PV output forecast error risk interval boundaries obtained by the two risk-based forecasting methods are superimposed on the baseline forecast results to generate their respective PV output risk-based forecast ranges. The MOICTCM method is used to optimize the coupling parameters of the two basic risk-based forecast results, balancing the dual requirements of risk interval coverage and risk interval width. BRIEF DESCRIPTION OF THE DRAWINGS

[0169] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0170] Figure 1 This is a flow chart of a method for predicting photovoltaic output hybrid probability intervals based on a parallel deep learning architecture according to an embodiment of the present invention;

[0171] Figure 2 A schematic diagram of a process for constructing a multi-source driver set for photovoltaic output according to an embodiment of the present invention;

[0172] Figure 3 This is a schematic diagram of the process of parallel decoupling and complementary enhancement of spatiotemporal features according to an embodiment of the present invention;

[0173] Figure 4 The present invention provides a flowchart for constructing a risk prediction model library based on data-driven and statistical theory. DETAILED DESCRIPTION

[0174] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0175] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0176] Example 1

[0177] As can be seen from the background technology,

[0178] Insufficient consideration of factors influencing photovoltaic output: While existing methods base their modeling on meteorological parameters and historical output data, they lack the systematic integration of multidimensional correlated factors. This is manifested in three missing mechanisms. First, geographic parameters (latitude, longitude, and altitude) directly influence the spatiotemporal distribution of solar radiation flux by altering the atmospheric mass coefficient and surface reflectivity. Second, equipment technical parameters (module photovoltaic conversion efficiency, array inclination / azimuth) determine the system's effective capture rate of incident radiation energy, but a dynamic coupling relationship is not established in existing models. Furthermore, while physics-based irradiance distribution simulations (including shadowing and diffuse reflection components) can provide spatially resolved input source terms, this high-precision data has not yet been effectively incorporated into the prediction system. This fragmentation of multidimensional influencing mechanisms makes it difficult for the model to accurately characterize the output response characteristics of photovoltaic power plants, ultimately significantly weakening the prediction accuracy and reliability under complex operating conditions.

[0179] Disharmony in decoupling of spatiotemporal features: Due to the single feature transfer path, the traditional stacked hybrid architecture makes it difficult to achieve dynamic mutual feedback between the spatial heterogeneity characteristics of the meteorological field and the time-frequency mutation characteristics of the output sequence, resulting in the cumulative amplification of prediction errors under complex meteorological conditions. Specifically, the meteorological field has significant spatial heterogeneity, that is, the meteorological conditions in different regions may vary greatly. This spatial heterogeneity has an important impact on output forecasting. However, the traditional stacked hybrid architecture often only considers a single feature transfer path in the feature transfer process, ignoring the dynamic mutual feedback relationship between the spatial heterogeneity characteristics of the meteorological field and the time-frequency mutation characteristics of the output sequence. This makes it difficult for the prediction model to accurately capture the changes in the spatial heterogeneity characteristics of the meteorological field under complex meteorological conditions, which in turn causes the cumulative amplification of prediction errors.

[0180] Risk-Based Forecasting Methods Are Fragmented: Current research on PV output risk interval forecasting relies on the separate application of probability distribution modeling and time-series quantile forecasting, hindering the effective coupling of statistical inference with dynamic evolutionary characteristics. Probability distribution modeling captures the stochastic nature of output fluctuations through nonparametric density estimation, but its static modeling paradigm cannot account for the time-series lag effects of meteorological factors. While time-series quantile forecasting can directly construct dynamic risk intervals based on historical data, it lacks explicit constraints on the output probability distribution morphology (such as multimodality and heteroscedasticity). The two methods are inherently interrelated in constructing risk intervals. Probability distribution constraints enhance the statistical rigor of interval confidence, while time-series dynamic modeling ensures the interval's ability to track the evolution of weather processes. However, the isolated application frameworks of existing methods lead to a decoupling of the two information flows. This can lead to either over-reliance on historical distributions, leading to conservative decisions in power plant regulation, or one-sided tracking of time-series fluctuations, amplifying the risk of failures caused by extreme weather events. This lack of system coupling creates a dual dilemma for interval forecasting in complex climate scenarios: coverage deviation and reliability degradation.

[0181] The present invention provides a photovoltaic output hybrid probability interval prediction method based on a parallel deep learning architecture, such as Figure 1 As shown, the following steps are included:

[0182] S1. Constructing a set of multi-source driving factors for photovoltaic output: First, systematically collect the research data required in the photovoltaic station area, pre-process the multi-source heterogeneous data, and construct a set of factors affecting photovoltaic output prediction. Using a photovoltaic output simulator based on physical mechanisms, a numerical solution for photovoltaic output driven by numerical weather forecast data is obtained. The calculation results are incorporated into the photovoltaic output prediction feature set to construct the original feature matrix containing spatial topological correlation information and temporal dynamic evolution characteristics. The specific process is as follows: Figure 2 shown.

[0183] S11. Define the study area and collect heterogeneous data from multiple sources, including historical output data of photovoltaic stations in the area, numerical weather forecast data, meteorological data, topographic data, and physical parameters of station equipment.

[0184] S12. Geographic Data Adapter: To quantify the coupling mechanism between the geographic attributes of PV sites and the solar radiation energy field, this method constructs a spatial feature fusion architecture based on the digital elevation model (DEM) and the irradiance distribution matrix. First, the elevation raster data and the irradiance sampling matrix are aligned through a spatial interpolation algorithm to establish a grid system with a unified geographic coordinate reference. Matrix operations are performed on the elevation matrix AL (representing the topographic relief at the site) and the irradiance matrix RA (reflecting the radiation flux at the site) to obtain the elevation-irradiance correlation matrix, in which each element represents the elevation-irradiance correlation at the PV site. The elevation-irradiance correlation matrix introduces spatial topological constraints, and each element contains information about the surrounding grid points, which enhances the spatial continuity and correlation of the data. The construction process of the elevation-irradiance correlation matrix is as follows:

[0185] 1. Normalize the elevation matrix and irradiance matrix to the same scale

[0186] ( )

[0187] ( )

[0188] Where AL scaled , RA scaled Represent the normalized elevation and irradiance matrices respectively.

[0189] 2. Weighted average processing

[0190] A weight is assigned to each element based on the Kendall correlation between elevation, irradiance and photovoltaic output, and then a weighted average is performed.

[0191] ( )

[0192] Where RAL represents the matrix of preliminary fusion of elevation and irradiance information, w AL 、w RA They represent the elevation and irradiance matrix weight parameters based on correlation allocation.

[0193] 3. Introducing spatial topological constraints

[0194] In the process of constructing the elevation-irradiance correlation matrix, spatial topological constraints are introduced so that each element YS ma It contains not only the elevation and irradiance information of the current location, but also the information of the surrounding grid points.

[0195] ( )

[0196] Among them YS ma represents the elevation-irradiance correlation matrix element corresponding to the target photovoltaic station at position (m, a), Qu ma represents the neighborhood of the target PV station, Indicates the location of other photovoltaic stations in the neighborhood of the target photovoltaic station, RAL ma 、 Represents the elements corresponding to the target photovoltaic station and its neighboring stations in the matrix of the preliminary fusion of elevation and irradiance information, is the spatial weight coefficient of each grid point and photovoltaic station in the neighborhood.

[0197] S13. Meteorological Data Adapter: This module converts raw meteorological data (such as irradiance and temperature) into a format suitable for the PV output forecast model and ensures that the data aligns with the elevation grid of the geographic data. Input data includes the raw meteorological data, target grid parameters, and geographic data containing elevation information. The raw meteorological data may be in a latitude-longitude grid or in some projected coordinate system. First, the raw data coordinate system is determined and the necessary projection transformations are performed to integrate the data into a unified coordinate system. Subsequently, bilinear interpolation is used to grid the meteorological data to the target resolution, ensuring that the meteorological data at each grid point aligns with the elevation grid of the geographic data. Bilinear interpolation estimates the value of an unknown point using the values of the four nearest known data points. For each meteorological variable, the four nearest data points surrounding the PV site are found in the raw data. The values of these four points are then averaged to obtain the meteorological data value for the PV site.

[0198] S14. Device Data Adapter: This adapter uses a dynamic sliding window to segment PV station equipment parameters into time series. Time-varying data retains the original time series records (e.g., PV module temperature, panel received irradiance), while time-invariant data is broadcasted in an aligned time window (e.g., module tilt, inverter model). This adapter explicitly characterizes the coupled state of the device's dynamic operation and static properties, providing input features that integrate physical constraints for the prediction model.

[0199] S15. Physics-based PV Output Simulator: Using PV output simulation values driven by physical mechanisms as model input features essentially builds a dual-channel collaborative modeling paradigm of physical constraints and data-driven models. The PV output benchmark simulation values output by the physical simulator contain strong prior knowledge of the solar radiation transmission equation and the thermoelectric conversion mechanism. Under extreme weather conditions, the physical constraint deduction mechanism constructed by the numerical weather forecast (NWP) based on the atmospheric motion equations can effectively compensate for the extrapolation limitations of purely data-driven models in oversampled spatial scenarios. The calculation process is as follows:

[0200] ( )

[0201] ( )

[0202] ( )

[0203] ( )

[0204] Where, SP g,t represents the physical simulation photovoltaic output of the g-th photovoltaic power station at time t, α is the solar radiation intensity rate, is the solar radiation intensity of the numerical weather forecast at time t, G stc is the solar radiation intensity under standard test conditions, 1000W / m 2 , is the solar panel temperature at time t, is the temperature of the numerical weather forecast at time t, C noc The normal operating temperature of the solar panel is usually , C stc The temperature under standard test conditions is 25℃, P stc is the photovoltaic output under standard test conditions, β is the heat loss efficiency of the photovoltaic cell, is the area of photovoltaic panels in the g-th photovoltaic power station, SP t The sum of the outputs of all photovoltaic power stations, NM PV is the number of photovoltaic power stations.

[0205] S2 Original Feature Matrix Construction: Under a unified spatiotemporal reference, the original feature matrix is constructed by synchronizing the spatial reference and time dimensions and performing feature splicing operations to link heterogeneous data from multiple sources, including geographic, meteorological, physical simulation output, and historical output of photovoltaic stations.

[0206] S21 Spatial Alignment: Map geographic data, meteorological data, physical simulation output, and historical output to the same grid, ensuring consistent resolution and range across all feature grids. First, identify the original coordinate system of each data source, which may include geographic coordinate systems or various projected coordinate systems. Use a coordinate transformation algorithm to convert all data to a coordinate system that matches the geographic data. Then, through gridding, bilinear interpolation is used to obtain the values of each feature at each PV site.

[0207] S22 Time Alignment: When geographic data, meteorological data, physical simulation output, and historical output have different time resolutions, all time series data must be aligned at the minimum time granularity. For data with higher time resolution, linear interpolation is used for downsampling. For data with lower time resolution, linear interpolation is also used to fill in missing time points, ensuring that all time series data are aligned in the time dimension.

[0208] S23 Feature Concatenation: Through spatial and temporal alignment, dynamic features such as irradiance, temperature, humidity, and historical output are combined into time series data with a uniform grid and resolution. Static feature parameters (such as module tilt and inverter model) remain constant over time and are broadcast to all time steps, ensuring that static and dynamic features correspond to each other at each time step. Finally, using tensor manipulation techniques, the static and dynamic feature tensors are concatenated along the feature dimension to form the final raw feature matrix. This tensor not only captures the dynamic changes in the time series but also encompasses the spatial distribution characteristics, providing comprehensive data input for the PV output prediction model.

[0209] S3. Parallel decoupling and complementary enhancement of spatiotemporal features: Based on the original photovoltaic output feature matrix, the GRU model (TGRU) with a temporal attention mechanism and the improved convolutional attention module (ICBAM) are used to construct a feature extraction module based on a parallel deep learning architecture to perform parallel decoupling of spatiotemporal features. The decoupled temporal and spatial features are input into the spatiotemporal feature complementary enhancement module for fusion to enhance the feature expression capability. Figure 3 shown.

[0210] S31. Parallel Decoupling of Spatiotemporal Features: We introduce a temporal attention mechanism into the GRU to construct a TGRU model. This extracts temporal features from the original feature matrix and enhances the weights of key time steps. Based on multi-scale recursive attention and dynamic weight coupling, we improve the convolutional attention module to create the ICBAM module, enhancing spatial feature extraction capabilities.

[0211] S311. Temporal Feature Decoupling: The TGRU model is used to separate the temporal dimension from the original feature matrix to better capture the changing patterns of time series. TGRU integrates temporal dependencies with the current state. By encoding temporal patterns in historical states and locating key time points, it can capture current features such as irradiance and temperature, and extract temporal patterns such as "cloud movement is often accompanied by a sharp drop in power." The key steps of TGRU are as follows:

[0212] 1. Update Gate

[0213] ( )

[0214] In the formula, z represents the update gate output, l represents the original feature matrix of the input, σ represents the Sigmoid function, and h prev Represents the hidden state of the previous moment, W z Represents the weight parameter of the update gate.

[0215] 2. Reset the door

[0216] ( )

[0217] Where re represents the reset gate output, W r Represents the weight parameter of the reset gate.

[0218] 3. Temporal Attention

[0219] ( )

[0220] ( )

[0221] Where, Represents the temporal attention score, U a represents the hidden state projection matrix, V a Represents the attention score matrix, W a Represents the weight parameter of the attention mechanism, l attn Represents the features after the attention mechanism is enhanced.

[0222] 4. Candidate status

[0223] ( )

[0224] Where h tilde represents the candidate hidden state, W tilde Represents the candidate hidden state calculation weight, and tanh is the hyperbolic tangent function.

[0225] 5. Status Update

[0226] ( )

[0227] Where h new Represents the updated hidden state. The hidden state at the final moment is the global representation of the time dimension feature of the photovoltaic output feature set l TGRU .

[0228] S312. Spatial Feature Decoupling: We improve the convolutional attention module based on multi-scale recursive attention and dynamic weight coupling mechanisms to construct the ICBAM module, separating the spatial dimension features from the original feature matrix. The single-scale attention in traditional CBAM struggles to capture details and global correlations in complex scenes. We introduce pooling operations into channel attention to generate multi-scale channel weights. Spatial attention constructs a receptive field through dilated convolutions, enhancing cross-scale feature interaction. The ICBAM module is suitable for tasks with drastic lighting changes and variable object scales. The details are as follows:

[0229] 1. Multi-scale Recursive Attention:

[0230] (1) Multi-scale channel attention module

[0231] ①Multi-scale pooling layer

[0232] Global average pooling: ( )

[0233] Global max pooling: ( )

[0234] Where td represents the channel index, l avg Represents the global average value of the feature of channel td, Row represents the height of the feature map, Col represents the width of the feature map, row and col are the position indexes respectively, l td,row,col Represents the characteristic value of row row col column in channel td, l max Represents the maximum response value among all eigenvalues of this channel.

[0235] ② Shared MLP structure: using two layers of full connection, the formula is:

[0236] ( )

[0237] ( )

[0238] Where, MLP(l avg )、MLP(l max ) represent the average and maximum pooling features after nonlinear transformation of MLP structure, are the weight matrices of each fully connected layer, δ is the ReLU activation function, b1 and b2 are the position offset control parameters of each fully connected layer.

[0239] ③Weight generation:

[0240] ( )

[0241] Where M td (l) is the output channel weight matrix, and TD is the channel dimension capacity.

[0242] ④ Feature weighting:

[0243] ( )

[0244] Where, l channel Represents the output of the channel attention module.

[0245] (2) Multi-receptive field spatial attention module

[0246] ① Channel compression: Perform average pooling and maximum pooling in the channel dimension, , represents the average pooled value output by the channel attention module, Represents the maximum pooling value of the channel attention module output.

[0247] ② Feature splicing:

[0248] ( )

[0249] Where, l cat Represents the channel compression feature after feature concatenation.

[0250] ③Convolution kernel design:

[0251] ( )

[0252] Where M s (lcat ) represents the concatenated features after the convolution operation.

[0253] ④ Spatial weighting:

[0254] ( )

[0255] Where, l spatial Features representing the output of the spatial attention module.

[0256] 2. Dynamic Weight Coupling Mechanism

[0257] In traditional CBAM, channel attention and spatial attention are connected in a fixed order, which limits the flexibility of feature interaction. A dynamic weight controller is designed to automatically adjust the fusion ratio of channel attention and spatial attention according to input features, improving the adaptability to different scenarios:

[0258] ( )

[0259] Where, l ICBAM represents the spatial features extracted by ICBAM, is the coupling parameter of channel attention and spatial attention.

[0260] S32. Temporal and spatial feature complementary enhancement module: This module fuses the decoupled temporal and spatial features of TGRU and ICBAM to enhance the expressiveness of these features and improve the accuracy of PV output forecasting.

[0261] S321. Feature Fusion: After aligning temporal and spatial feature channels, the concatenated tensor features are fed into a 3D convolution kernel to capture local spatial and temporal correlations.

[0262] ( )

[0263] Where, l fused Represents the fusion feature after fusing temporal and spatial features.

[0264] S322. Feature Enhancement: Use cross-modal residual connections to enhance the fused feature vector. Align the original temporal and spatial features through one-dimensional convolution and then add them to the fused features to avoid information loss.

[0265] ( )

[0266] Where, Represents the spatiotemporal feature matrix of photovoltaic output after feature enhancement.

[0267] S4. Photovoltaic Output Benchmark Forecast Model: Because photovoltaic output is affected by both local meteorological fluctuations (such as cloud cover) and long-term seasonal trends, a fixed window size is insufficient to adapt to these changes. Therefore, a dynamic sparse window is introduced to construct a photovoltaic output benchmark forecast model based on a sparse attention Transformer. This sparse attention mechanism retains highly relevant positions in the computation. Sparse attention is only used in the encoder to process long input sequences, while the decoder retains standard attention. This reduces computational cost and improves long sequence modeling capabilities. The key steps are as follows:

[0268] 1. Standard attention mechanism retained by the decoder:

[0269] ( )

[0270] Where, is the query, key, and value matrix, d k is the dimension scaling factor.

[0271] 2. Dynamic Sparse Attention of Encoder

[0272] ( )

[0273] ( )

[0274] Where LocalWindow(P) represents the dynamic window centered at position P, GlobalAnchor(P) represents the global anchor point selected according to the importance score, M represents the attention mask matrix, and q represents the candidate key position within the dynamic window or within the range of the anchor point.

[0275] S5. Risk-based prediction model library based on data-driven and statistical theory: The improved QRLSTM (IQRLSTM) relies on the quantile regression framework to directly output the photovoltaic output error quantiles corresponding to different risk levels during the time series modeling process. Finally, the error quantile prediction results output by it are superimposed with the benchmark prediction value to obtain the risk-based prediction result. The traditional kernel density estimation method is improved using adaptive kernel functions and local bandwidth adjustment to establish the IKDE method. IKDE derives the cumulative probability distribution curve by fitting the probability density function of the photovoltaic output benchmark prediction error data of the same timestamp, thereby extracting the error threshold based on the specified risk level. The upper and lower boundaries of the error value are inversely solved through the cumulative distribution function, and superimposed on the benchmark prediction result to generate the photovoltaic output risk-based prediction range. IQRLSTM learns the dynamic evolution law of the photovoltaic output error distribution through a deep learning model, while IKDE focuses on the statistical inference of the posterior probability distribution of photovoltaic output, and relies on the independent and identically distributed assumption of the benchmark prediction error. The contents of each part are as follows Figure 4 shown.

[0276] S51 Data-Driven Risk-Based Forecasting Model: IQRLSTM leverages the quantile regression framework to construct a multi-quantile loss function, synchronously outputting error quantiles corresponding to different risk levels during the time series modeling process. This model dynamically captures the temporal dependencies of PV output forecast errors by optimizing the quantile loss function, enabling end-to-end prediction of upper and lower error quantiles. Finally, the quantile error prediction output from IQRLSTM is superimposed with the baseline prediction to obtain a risk-based PV output forecast. This model can be transformed into an optimization problem as shown in Equation 1.

[0277] ( )

[0278] Where obj1 represents the IQRLSTM objective function without considering the quantile constraint penalty term, y i Represents the actual observed value of photovoltaic output, Represents the photovoltaic output forecast value of the corresponding quantile, represents the quantile, and J represents the total number of quantiles.

[0279] IQRLSTM considers the inherent properties of quantiles and adds a quantile constraint penalty term to the objective function to avoid quantile crossing.

[0280] ( )

[0281] ( )

[0282] ( )

[0283] Where dif represents the sum of multi-quantile losses, penalty represents the constraint penalty term, and obj represents the final loss function.

[0284] S52 Risk-based prediction model based on statistical theory: The traditional kernel density estimation method is improved using adaptive kernel functions and local bandwidth adjustment to obtain the IKDE method. After fitting the probability density distribution of the data with the same historical timestamp, the fitting distribution of the photovoltaic output benchmark prediction error value corresponding to the timestamp is determined, and the photovoltaic output benchmark prediction value is superimposed to obtain the corresponding photovoltaic output risk-based prediction result. Adaptive dynamic bandwidth adjustment based on the nearest neighbor distance can adjust the kernel width according to the local data density. The core idea is to adjust the bandwidth according to the local density around each data point, using a larger bandwidth in sparse areas to smooth noise, and a smaller bandwidth in dense areas to retain details. The following is the specific implementation method:

[0285] 1. Divide the photovoltaic output benchmark prediction error according to the timestamp, use the time of day as the basis for division, and extract the error values at the same time as the data set.

[0286] 2. Distance calculation

[0287] ( )

[0288] Where fac1 and fac2 represent the elements to be calculated for distance.

[0289] 3. Use the binary tree search method to accelerate the nearest neighbor search and convert the photovoltaic output benchmark prediction error sequence Arrange in ascending order, n represents the sample size, and the local neighborhood size is win. For each error point err i , expand forward and backward in the sorted array points, the maximum distance between the error point and the neighborhood boundary point is .

[0290] 4. Dynamic bandwidth adjustment

[0291] ( )

[0292] ( )

[0293] Where, Represents the local density sensitivity factor (default 0.5), which controls the expansion amplitude of sparse areas. Represents the global smoothing factor (default 0.1) to prevent local overfitting, h global represents the global bandwidth, Represents the standard deviation of the data. If it is large, the data points are sparse and the bandwidth h should be increased. i To avoid overfitting, the bandwidth h should be reduced. i to increase resolution.

[0294] 5. Adaptive kernel function, using improved Gaussian kernel function:

[0295] ( )

[0296] S6. Hybrid Risk-Based Forecasting Framework Based on a Multi-Objective Tribal Competition and Member Cooperation Optimization Algorithm: This paper proposes an improved single-objective tribal competition and member cooperation optimization algorithm (ICTCM) using a triple mutation strategy of "sine mapping, spiral flight, and Lévy flight." While maintaining population diversity, this algorithm improves its global exploration capability and convergence stability. A multi-objective adaptation mechanism is employed to reconstruct the algorithm, transforming it from a single-objective to a multi-objective optimization framework (MOICTCM). By coupling the risk-based prediction results of the IKDE and IQRLSTM with the MOICTCM, a hybrid risk-based prediction of PV output is obtained.

[0297] S61 Improved single-objective tribe competition and member cooperation optimization algorithm

[0298] S611 Original single-objective tribe competition and member cooperation optimization algorithm

[0299] The concept of CTCM is inspired by the natural mechanisms of inter-tribal competition and the collaborative patterns among members within a tribe. The core of the CTCM algorithm is local development within a tribe, global search between tribes, and dynamic loyalty of members to the tribe. The CTCM algorithm optimizes local solutions through member collaboration and exploits inter-tribal competition to perform global search, thereby exploring more optimal solutions in the solution space. Finally, the algorithm regulates the behavior of tribe members during cooperation and competition and determines their dynamic loyalty, ensuring flexibility and adaptability during the optimization process. The optimization process is as follows:

[0300] In the population initialization phase, the algorithm population can be represented as the following matrix:

[0301] ( )

[0302] ( )

[0303] Where X is the initial position of the entire primitive society, pop is the population of the primitive society, Num is the number of tribes, Mem is the number of members in each tribe, and dim is the dimension of the solution space.

[0304] The fitness of primitive society is:

[0305] ( )

[0306] The velocity vector of each member determines the direction of the member, and the velocity matrix of the primitive society as a whole as follows:

[0307] ( )

[0308] ( )

[0309] In a tribal structure, the chief is responsible for managing and planning the future development of the tribe, and most members tend to follow the chief's guidance and plans. However, due to the personal opinions of members, their loyalty to the chief fluctuates over time. The dynamic changes in loyalty reflect the complexity and diversity of individual behavior within the tribe. CTCM uses a sinusoidal chaos map to characterize the changes in member loyalty. The loyalty of individual members r t It can be expressed as:

[0310] ( )

[0311] Update velocity matrix:

[0312] ( )

[0313] Where, represents the speed of the mem-th member of the num-th tribe in the dim dimension in the te+1-th iteration, The inertia coefficient representing the velocity, Represents the position corresponding to the best fitness value of the member in the cycle, Represents the position of the member at the teth iteration, Represents the position corresponding to the best fitness value of the tribe in the cycle, c1 and c2 represent the experience factor and obedience factor respectively, 、 Represents the Chaos Loyalty of each member.

[0314] Random clashes may occur between tribes, during which the weaker tribe will flee while the stronger tribe will remain unaffected. Speed updates are as follows:

[0315] ( )

[0316] ( )

[0317] Where, represents the best fitness value of the num-th tribe, represents the optimal fitness value in the competition process, c3 represents the escape factor, Represents the chaotic random factor.

[0318] The position update of the competition process is shown in Equation (49). When the updated position exceeds the boundary, a mirror rebound operation of the speed is performed.

[0319] ( )

[0320] ( )

[0321] Where, Represents the position of the mem-th member of the num-th tribe in the dim dimension at the te+1-th iteration.

[0322] S612 Improvement Strategy

[0323] While CTCM offers excellent stability and high computational accuracy, it still has limitations. During population initialization, the random distribution of initial positions and the differences between individual populations can easily lead to a lack of population diversity. Furthermore, improper algorithm parameter settings can result in insufficient perturbations or overly rapid convergence, trapping the algorithm in a local optimum. Therefore, the following three improvements were made to the CTCM algorithm:

[0324] 1. Sine Mapping

[0325] Sine mapping is a chaotic system based on the sine function. It generates chaotic sequences through nonlinear transformations of the sine function, and usually has the characteristics of combining periodicity and chaos. Coupled sine mapping helps increase population diversity and enhance global search capabilities. The specific formula is as follows:

[0326] ( )

[0327] Where, X te 、X te+1 Represent the populations of the te and te+1 iterations respectively, are the parameters of the chaotic system.

[0328] 2. Spiral flight

[0329] In the basic CTCM, the performance of the algorithm depends on the information exchange between tribes and the learning strategies of individuals. Generally, tribe members tend to move closer to the optimal position of the tribe to achieve efficient resource utilization. However, the centralized search strategy may cause the algorithm to converge to the local optimal solution too early, limiting the global search ability of the algorithm. To this end, the spiral flight strategy is introduced to improve CTCM. The random walk characteristics of spiral flight can produce long-distance jumps during the search process, which can not only perform fine searches in local areas, but also explore globally to discover excellent solutions that may be overlooked. The introduction of the spiral flight strategy not only enhances the algorithm's ability to avoid premature convergence, but also improves the algorithm's search efficiency and global search ability in complex optimization problems. The specific formula is as follows:

[0330] ( )

[0331] Where, represents the best individual in the te-th iteration, represents the logarithmic spiral shape constant, Represents a random number between [-1,1].

[0332] 3. Levi Flight

[0333] After the iterative optimization is completed, the Levy flight strategy is added to force the adjustment of individual positions, enhance the exploration ability of the algorithm, and ensure that the local optimal solution can be jumped out during the optimization process, so as to find a better solution in the global solution space.

[0334] The random step length of Levy flight is calculated as follows:

[0335] ( )

[0336] Where, 、 It obeys the random normal distribution and can be expressed as:

[0337] ( )

[0338] Where, represents the gama function, The value range is (0, 2].

[0339] S62. Modified Multi-Objective Tribe Competition and Member Cooperation Optimization Algorithm: The key to transforming a single-objective optimization algorithm into a multi-objective one lies in introducing a multi-objective trade-off mechanism while maintaining convergence and diversity of solutions. The following are the specific steps and key methods:

[0340] S621. Multi-Objective Fitness Evaluation and Pareto Ranking

[0341] 1. Target vectorization:

[0342] ( )

[0343] Where F represents the target vector, Ang represents the dimension of the target function, and f Ang Represents the subgoal of maximum dimension.

[0344] 2. Solution A dominates solution B if and only if:

[0345] ( )

[0346] Where, f ang () represents each sub-objective function.

[0347] 3. Define the Pareto dominance relationship: nu represents the population size, traversing the population {o1,o2,…,o nu}, calculate the non-dominated relations of all solutions according to formula (56), all non-dominated solutions constitute the first-level Pareto frontier PF1, remove the first-level Pareto frontier solutions from the population, and recursively extract the subsequent Pareto frontiers from the remaining solutions, and finally obtain ku Pareto frontier layers PF1, PF2, ..., PF ku , where PF1 is the optimal solution set, PF2 is the second best, and so on.

[0348] 4. Within each layer of the Pareto front, calculate the crowding degree of the solution to maintain diversity:

[0349] ( )

[0350] Where CD represents the individual crowding value, O Now represents the current solution in the target space, O Now−1 and O Now+1 They represent the adjacent solutions of the current solution in the target space.

[0351] S622. Tribal Structure and Multi-Objective Extension of Competition Rules

[0352] 1. Tribe Division and Role Definition: Tribes are divided based on the non-dominated sorting results. Each tribe contains solutions at different Pareto levels. The non-dominated solution with the highest crowding degree in the tribe becomes the chief.

[0353] 2. Dynamic adjustment of member loyalty:

[0354] ( )

[0355] In the formula, te represents the current number of iterations, r te represents the loyalty of members in the teth iteration, Represents the maximum number of iterations.

[0356] 3. Multi-objective update formula:

[0357] ( );

[0358] Where p best is the best historical position of the individual, g best For the tribal chief position, X random is a randomly selected non-dominated solution from other tribes, represents the speed of the mem-th member of the num-th tribe in the dim dimension in the te+1-th iteration, The inertia coefficient representing the velocity, Represents the position of the member in the teth iteration, c1 and c2 represent the experience factor and compliance factor respectively, 、 Represents the chaos loyalty of each member, c3 represents the escape factor, represents the chaotic random factor;

[0359] 4. Integration of spiral flight strategy: Introducing spiral exploration into position updates, performing refined searches near the Pareto frontier, and enhancing local development capabilities:

[0360] ;

[0361] Where, represents the logarithmic spiral shape constant, Represents a random number between [-1, 1].

[0362] S623. External Archive Management and Elite Retention

[0363] 1. Archive update rule: merge the parent and child populations, filter non-dominated solutions and add them to the temporary archive. If the temporary archive exceeds the capacity SI archive , eliminate solutions from low to high according to the congestion degree.

[0364] 2. Levy flight perturbation: Apply Levy perturbation to the archived solution every Ge generation to jump out of the local optimum on a global scale, discover potential solutions, and avoid being trapped in the local Pareto frontier:

[0365] ( )

[0366] Where, Represent the archive groups in the te and te+1 iterations respectively, Represents the step size parameter.

[0367] 3. Clash of Clans and Escape Mechanism

[0368] (1) Conflict triggering conditions

[0369] Every generation, two tribes are randomly selected. A and Tribe B , and compare their hypervolume contributions HV:

[0370] ( )

[0371] (2) Escape strategy

[0372] The escape strategy guides the losing tribe to learn from the winning tribe and accelerates convergence. The speed of the losing tribe members is updated as follows:

[0373] ( )

[0374] Where, represents the speed of the failed tribe A member during the te-th iteration, represents the update speed of tribe A members, c4 represents the escape factor, Represents the update random factor, X B,chief Represents the position of the chief in the winning tribe B, represents the position of the tribe A member at the te-th iteration.

[0375] S63 Hybrid Risk-Based Prediction Framework

[0376] IKDE generates statistically significant risk intervals based on the benchmark PV output forecast error, suitable for scenarios with a stable global error distribution. IQRLSTM can directly model time-dependent error quantiles and adapt to local fluctuations. To leverage the complementary strengths of IKDE and IQRLSTM, a hybrid risk-based forecasting framework based on MOICTCM is proposed, using the prediction interval average bandwidth (PINAW) and prediction interval coverage (PICP) metrics as objective functions. By synergistically integrating the advantages of statistical methods and data-driven models, this framework ensures the statistical reliability of risk-based forecast intervals while improving adaptability to the non-steady-state fluctuations of PV output.

[0377] ( )

[0378] ( )

[0379] ( )

[0380] In the formula, PINAW represents the average bandwidth index of the prediction interval, PICP represents the coverage index of the prediction interval, represents the established risk level, ω represents the risk type prediction result coupling parameter of IKDE and IQRLSTM, Up mix 、Up IKDE 、Up IQRLSTM Respectively represent the upper bound of mixed risk type prediction, the upper bound of IKDE risk type prediction, the upper bound of IQRLSTM risk type prediction, Low mix 、Low KDE 、Low QRLSTM They represent the lower bounds of mixed risk type prediction, IKDE risk type prediction, and IQRLSTM risk type prediction, respectively. 、 Represents the maximum upper bound and minimum lower bound of the prediction interval under a given risk level, Dis represents the maximum range of the prediction interval boundary, 、 Represents the upper and lower bounds of the prediction interval at the i-th sample point under a given risk level. Mo is a Boolean value. When the actual photovoltaic output value is within the risk-type prediction interval, Mo takes the value of 1, otherwise it takes the value of 0. Represents the Mo value corresponding to the i-th sample point under a given risk level.

[0381] Example 2

[0382] The present invention also provides a photovoltaic output hybrid probability interval prediction system based on a parallel deep learning architecture, the system is used to implement the method of embodiment 1, the system includes: a data set construction module, a matrix construction module, a feature fusion module, a first prediction module, a second prediction module and a third prediction module;

[0383] Dataset construction module, used to construct a set of multi-source driving factors of photovoltaic output;

[0384] A matrix construction module is used to construct an original feature matrix by synchronizing the set of multi-source driving factors of photovoltaic output under a unified spatiotemporal reference through spatial reference and time dimension and feature splicing;

[0385] A feature fusion module is used to construct a feature extraction module based on a parallel deep learning architecture, perform parallel decoupling of the spatiotemporal features of the original feature matrix, and input the decoupled temporal and spatial features into the spatiotemporal feature complementary enhancement module for fusion;

[0386] The first prediction module is used to build a photovoltaic output benchmark prediction model based on sparse attention Transformer, input the photovoltaic output spatiotemporal feature matrix after feature enhancement into the photovoltaic output benchmark prediction model, and obtain the photovoltaic output benchmark prediction result and corresponding error;

[0387] The second prediction module is used to build a risk-based prediction model based on data-driven and statistical theory. The error of the PV output benchmark prediction result is input into the two risk-based prediction models respectively, and the boundary value of the prediction error risk interval is calculated. The boundary value is superimposed on the benchmark prediction result to obtain the respective PV output risk-based prediction results.

[0388] The third prediction module is used to build a photovoltaic output hybrid risk-type prediction framework, input the results of two risk-type prediction models for coupling, and use a multi-objective optimization algorithm to optimize the coupling parameters to obtain the photovoltaic output hybrid risk-type prediction results.

[0389] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A photovoltaic output hybrid probability interval prediction method based on a parallel deep learning architecture is characterized by: The method comprises: S1: Construct a set of multi-source driving factors for photovoltaic output; S2: Synchronize the multi-source driving factor set of photovoltaic output under a unified spatiotemporal reference by spatial reference and time dimension, and construct an original feature matrix; S3: Construct a feature extraction module based on a parallel deep learning architecture, perform parallel decoupling of the temporal and spatial features of the original feature matrix, and input the decoupled temporal and spatial features into the temporal and spatial feature complementary enhancement module for fusion; S4: Construct a photovoltaic output benchmark prediction model based on sparse attention Transformer, input the photovoltaic output spatiotemporal feature matrix after feature enhancement into the photovoltaic output benchmark prediction model, and obtain the photovoltaic output benchmark prediction result and corresponding error; S5: Construct a risk-based prediction model based on data-driven and statistical theory. Input the error of the PV output benchmark prediction result into the two risk-based prediction models respectively, calculate the boundary value of the prediction error risk interval, and superimpose it on the benchmark prediction result to obtain the respective PV output risk-based prediction results. S6: Construct a hybrid risk-based prediction framework for photovoltaic output, input the results of two risk-based prediction models for coupling, and use a multi-objective optimization algorithm to optimize the coupling parameters to obtain the hybrid risk-based prediction results for photovoltaic output.

2. The method according to claim 1, characterized in that S1, constructing a set of multi-source driving factors for photovoltaic output, includes: S11: Define the research area and systematically collect multi-source heterogeneous data including historical output data of photovoltaic stations, numerical weather forecast data, meteorological data, topographic data, and physical parameters of station equipment; S12: constructing a geographic data adapter to obtain the elevation-irradiance correlation matrix of regional photovoltaic stations; S13: Build a meteorological data adapter to convert the raw meteorological data into a format that meets the model input and align the meteorological data with the elevation grid of the geographic data; S14: Build a device data adapter to convert various device parameters into a form that meets the model input; S15: Construct a photovoltaic output simulator based on physical mechanisms, and use the photovoltaic output simulation value driven by the physical mechanism as the input feature of the model.

3. The method according to claim 1, characterized in that In S3, a feature extraction module based on a parallel deep learning architecture is constructed to perform parallel decoupling of the spatiotemporal features of the original feature matrix, and the decoupled temporal and spatial features are input into a spatiotemporal feature complementary enhancement module for fusion, including: S31: Construct a parallel decoupling module for spatiotemporal features. The GRU model (TGRU) based on the improved temporal attention mechanism is used to extract the temporal features from the original PV output feature matrix. The spatial features in the original feature matrix are extracted by the convolutional attention module (ICBAM) based on the improved multi-scale recursive attention and dynamic weight coupling mechanism. S32: Construct a spatiotemporal feature complementary enhancement module to fuse the temporal and spatial features after decoupling TGRU and ICBAM; Among them, TGRU includes: Update Gate ； In the formula, z represents the update gate output, l represents the original feature matrix of the input, σ represents the Sigmoid function, and h prev Represents the hidden state of the previous moment, W z Represents the weight parameter of the update gate; Reset Gate ； Where re represents the reset gate output, W r Represents the weight parameter of the reset gate; Temporal Attention ； ； Where, Represents the temporal attention score, U a represents the hidden state projection matrix, V a Represents the attention score matrix, W a Represents the weight parameter of the attention mechanism, l attn Represents the characteristics after the attention mechanism is strengthened; Candidate status ； Where h tilde represents the candidate hidden state, tanh is the hyperbolic tangent function; Status Update ； Where h new Represents the updated hidden state. The hidden state at the final moment is the global representation of the time dimension feature of the photovoltaic output feature set l TGRU ; ICBAM includes: multi-scale recursive attention and dynamic weight coupling mechanism; Among them, the multi-scale recursive attention includes: multi-scale channel attention module and multi-receptive field spatial attention module; The multi-scale channel attention module includes: Multi-scale pooling layer: Global average pooling: ; Global max pooling: ; Where td represents the channel index, l avg Represents the global average value of the feature of channel td, Row represents the height of the feature map, Col represents the width of the feature map, row and col are the position index l td,row,col Represents the characteristic value of row row col column in channel td, l max Represents the maximum response value among all eigenvalues of the channel; Shared MLP structure: using two layers of full connection, the formula is: ； ； Where, MLP(l avg )、MLP(l max ) represent the average and maximum pooling features after nonlinear transformation of MLP structure, are the weight matrices of each fully connected layer, δ is the ReLU activation function, b1 and b2 are the position offset control parameters of each fully connected layer; Weight generation: ； Where M td (l) is the output channel weight matrix, TD is the channel dimension capacity; Feature weighting: ； Where, l channel Represents the output of the channel attention module; The multi-receptive field spatial attention module includes: Channel Compression: Yes Perform average pooling and maximum pooling in the channel dimension to obtain , represents the average pooled value output by the channel attention module, Represents the maximum pooling value of the channel attention module output; Feature stitching: ； Where, l cat Represents the channel compression feature after feature splicing; Convolution kernel design: ； Where M s (l cat ) represents the concatenated features after convolution operation; Spatial weighting: ； Where, l spatial Features representing the output of the spatial attention module; Dynamic weight coupling mechanism: A dynamic weight controller is designed to automatically adjust the fusion ratio of channel attention and spatial attention according to input features, improving the adaptability to different scenarios: ； Where, l ICBAM represents the spatial features extracted by ICBAM, is the coupling parameter of channel attention and spatial attention; Construct a spatiotemporal feature complementary enhancement module to fuse the temporal and spatial features after decoupling TGRU and ICBAM, including: After aligning the temporal and spatial feature channels, the concatenated features are fed into the 3D convolution kernel to capture the temporal and spatial local correlations: ； Where, l fused Represents the fusion feature after fusing time and space features; The fused feature vector is enhanced using cross-modal residual connections, and the original temporal and spatial features are aligned through one-dimensional convolution and then superimposed on the fused features: ； Where, Represents the spatiotemporal feature matrix of photovoltaic output after feature enhancement.

4. The method according to claim 1, wherein In S4, the method for constructing a photovoltaic output benchmark prediction model based on sparse attention Transformer includes: The standard attention mechanism retained by the decoder: ； Where, is the query, key, and value matrix, d k is the dimension scaling factor; Dynamic Sparse Attention for Encoder: ； ； Where LocalWindow(P) represents the dynamic window centered at position P, GlobalAnchor(P) represents the global anchor point selected according to the importance score, M represents the attention mask matrix, and q represents the candidate key position within the dynamic window or within the range of the anchor point.

5. The method according to claim 1, wherein S5: Construct a risk-based prediction model based on data-driven and statistical theory. Input the error of the PV output benchmark prediction result into two risk-based prediction models respectively, calculate the boundary value of the prediction error risk interval, and superimpose it on the benchmark prediction result to obtain the respective PV output risk-based prediction results, including: S51: Build a data-driven risk-based prediction model. Based on the quantile regression framework, a multi-quantile loss function is constructed. During the time series modeling process, error quantiles corresponding to different risk levels are simultaneously output and superimposed on the baseline prediction results to obtain a data-driven risk-based prediction result for PV output. S52: Construct a risk-based prediction model based on statistical theory. Use adaptive kernel functions and local bandwidth adjustment to improve the traditional kernel density estimation method, resulting in the IKDE method. Superimpose the prediction error risk interval boundaries on the baseline prediction results to obtain a statistically based risk-based prediction result for photovoltaic output. Wherein, the S51 includes: ； Where obj1 represents the IQRLSTM objective function without considering the quantile constraint penalty term, y i Represents the actual observed value of photovoltaic output, Represents the photovoltaic output forecast value of the corresponding quantile, represents the quantile, and J represents the total number of quantiles; The improved QRLSTM, i.e. IQRLSTM, takes into account the inherent properties of quantiles and adds a quantile constraint penalty term to the objective function. ； ； ； Where dif represents the sum of multi-quantile losses, penalty represents the constraint penalty term, and obj represents the final loss function; The S52 includes: The photovoltaic output benchmark prediction error is divided according to the timestamp; Distance calculation: ； In the formula, fac1 and fac2 represent the elements to be calculated; Use the binary tree search method to accelerate the nearest neighbor search and convert the photovoltaic output benchmark prediction error sequence Arrange in ascending order, n represents the sample size, the local neighborhood size is win, and for each error point err i , expand forward and backward in the sorted array points, the maximum distance between the error point and the neighborhood boundary point is ; Dynamic bandwidth adjustment: ； ； Where, Represents the local density sensitivity factor, which controls the expansion amplitude of the sparse area. represents the global smoothing factor, h global represents the global bandwidth, represents the standard deviation of the data, h i is bandwidth; Adaptive kernel function, using improved Gaussian kernel function: 。 6. The method according to claim 5, characterized in that In S6, a photovoltaic output hybrid risk-based prediction framework is constructed, the results of the two risk-based prediction models are input for coupling, and a multi-objective optimization algorithm is used to optimize the coupling parameters to obtain photovoltaic output hybrid risk-based prediction results, including: S61: Using the triple mutation strategy of "sine mapping-spiral flight-Lévy flight", an improved single-objective tribal competition and member cooperation optimization algorithm, namely ICTCM, is proposed; S62: Use multi-objective adaptation mechanism reconstruction to achieve the transformation from single-objective to multi-objective optimization framework, namely MOICTCM; S63: The risk-based prediction results of PV output are obtained by coupling the risk-based prediction results of IKDE and IQRLSTM through MOICTCM; Wherein, the S62 includes: S621: Multi-objective fitness evaluation and Pareto ranking: Target vectorization: ； Where F represents the target vector, Ang represents the dimension of the target function, and f Ang represents the subgoal with the largest dimension; Solution A dominates solution B if and only if: ； Where Ang represents the objective function dimension, f ang () represents each sub-objective function; Define the Pareto dominance relationship: calculate the non-dominated relationship of all solutions, extract the first layer of non-dominated solutions, recursively extract the next layer of non-dominated solutions from the remaining solutions, and divide the population into multiple Pareto frontier layers; Within each level of the Pareto front, the crowding of solutions is calculated to maintain diversity: ； Where CD represents the individual crowding value, O Now represents the current solution in the target space, O Now−1 and O Now+1 Represent the adjacent solutions of the current solution in the target space; S622: Multi-objective extension of tribal structure and competition rules: Tribe division and role definition: Tribes are divided according to the non-dominated sorting results. Each tribe contains solutions at different Pareto levels. The non-dominated solution with the highest crowding degree in the tribe is the chief. Dynamic adjustment of member loyalty: ； In the formula, te represents the current number of iterations, r te represents the loyalty of members during the te-th iteration, Represents the maximum number of iterations; Multi-target velocity update formula: ； Where p best is the best historical position of the individual, g best For the tribal chief position, X random is a randomly selected non-dominated solution from other tribes, represents the speed of the mem-th member of the num-th tribe in the dim dimension in the te+1-th iteration, The inertia coefficient representing the velocity, Represents the position of the member in the teth iteration, c1 and c2 represent the experience factor and compliance factor respectively, 、 Represents the chaos loyalty of each member, c3 represents the escape factor, represents the chaotic random factor; Integration of spiral flight strategy: Introducing spiral exploration into position update and searching near the Pareto frontier: ； Where, represents the logarithmic spiral shape constant, Represents a random number between [-1, 1]; S623: External Archive Management and Retention: Archive update rule: merge the parent and child populations, filter non-dominated solutions and add them to the temporary archive. If the temporary archive exceeds the capacity SI archive , eliminate solutions from low to high according to the congestion degree; Levy flight perturbation: Apply Levy perturbation to the archived solution every Ge generation, jump out of the local optimum on a global scale, and discover potential solutions: ； Where, Represent the archive groups in the te and te+1 iterations respectively, represents the step size parameter, 、 Obeying a random normal distribution, The value range is (0, 2], Represents the gama function; Clash of Clans and Escape Mechanics: Conflict triggering conditions: Every generation, two tribes are randomly selected. A and Tribe B , and compare their hypervolume contributions HV: ； Escape strategy: The escape strategy guides the losing tribe to learn from the winning tribe and accelerates convergence. The speed of the losing tribe members is updated as follows: ； Where, represents the speed of the failed tribe A member during the te-th iteration, represents the update speed of tribe A members, c4 represents the escape factor, Represents the update random factor, X B,chief Represents the position of the chief in the winning tribe B, represents the position of the tribe A member in the teth iteration; The S63 includes: ； ； ； In the formula, PINAW represents the average bandwidth index of the prediction interval, PICP represents the coverage index of the prediction interval, represents the established risk level, ω represents the risk type prediction result coupling parameter of IKDE and IQRLSTM, Up mix 、Up IKDE 、Up IQRLSTM Respectively represent the upper bound of mixed risk type prediction, the upper bound of IKDE risk type prediction, the upper bound of IQRLSTM risk type prediction, Low mix 、Low KDE 、Low QRLSTM They represent the lower bounds of mixed risk type prediction, IKDE risk type prediction, and IQRLSTM risk type prediction, respectively. 、 Represents the maximum upper bound and minimum lower bound of the prediction interval under a given risk level, Dis represents the maximum range of the prediction interval boundary, 、 Represents the upper and lower bounds of the prediction interval at the i-th sample point under a given risk level. Mo is a Boolean value. When the actual photovoltaic output value is within the risk-type prediction interval, Mo takes the value of 1, otherwise it takes the value of 0. represents the Mo value corresponding to the i-th sample point under a given risk level. The meanings of other variables remain consistent with the previous text.

7. A photovoltaic output hybrid probability interval prediction system based on a parallel deep learning architecture, the system being used to implement the method of any one of claims 1 to 6, characterized in that: The system includes: a data set construction module, a matrix construction module, a feature fusion module, a first prediction module, a second prediction module and a third prediction module; The data set construction module is used to construct a set of multi-source driving factors for photovoltaic output; The matrix construction module is used to construct an original feature matrix by synchronizing the set of photovoltaic output multi-source driving factors under a unified time and space reference through spatial reference and time dimension and feature splicing; The feature fusion module is used to construct a feature extraction module based on a parallel deep learning architecture, perform parallel decoupling of the spatiotemporal features of the original feature matrix, and input the decoupled temporal and spatial features into the spatiotemporal feature complementary enhancement module for fusion; The first prediction module is used to construct a photovoltaic output benchmark prediction model based on sparse attention Transformer, input the photovoltaic output spatiotemporal feature matrix after feature enhancement into the photovoltaic output benchmark prediction model, and obtain the photovoltaic output benchmark prediction result and corresponding error; The second prediction module is used to build a risk-based prediction model based on data-driven and statistical theory, input the photovoltaic output benchmark prediction result error into the two risk-based prediction models respectively, calculate the prediction error risk interval boundary value, and superimpose it on the benchmark prediction result to obtain the respective photovoltaic output risk-based prediction results; The third prediction module is used to construct a photovoltaic output hybrid risk-type prediction framework, input the results of two risk-type prediction models for coupling, and use a multi-objective optimization algorithm to optimize the coupling parameters to obtain a photovoltaic output hybrid risk-type prediction result.

Images