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

Abstract

The invention discloses a photovoltaic output hybrid probability interval prediction method and system based on a parallel deep learning architecture. The method comprises the following steps: 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 spatial-temporal feature parallel decoupling on the original feature matrix, and inputting the decoupled time features and spatial features into a spatial-temporal feature complementary enhancement module for fusion; inputting the photovoltaic output spatial-temporal feature matrix after feature enhancement into a photovoltaic output reference type prediction model to obtain a photovoltaic output reference type prediction result and a corresponding error; inputting the photovoltaic output reference type prediction result errors into the risk type prediction model, calculating prediction error risk interval boundary values, and superposing the prediction error risk interval boundary values to the reference type prediction result to obtain respective photovoltaic output risk type prediction results; and constructing a photovoltaic output hybrid risk type prediction framework, inputting two risk type prediction model results for coupling and optimization, and obtaining a photovoltaic output hybrid risk type prediction result.

Description

Technical Field

[0001] The present invention belongs to the cross - technical 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 Technique

[0002] Current photovoltaic output prediction methods exhibit multi - dimensional and multi - model coupling characteristics. Physical models, statistical methods, machine learning, and hybrid architectures are closely related through data networks. Traditional benchmark prediction methods have significant limitations in dealing with complex meteorological conditions. When multi - dimensional spatio - temporal features are dynamically interconnected, single - point prediction models are difficult to quantify potential uncertainties, resulting in a risk misalignment between the prediction results and the actual output fluctuations. For example, in extreme weather events, the superposition of physical modeling biases in numerical weather forecasting and the static assumptions of statistical models is likely to trigger the chain - like propagation of prediction errors, highlighting the urgent need to construct risk - based predictions to enhance the robustness of power grid dispatching.

[0003] There are three main problems in the existing prediction technology system. First, photovoltaic output prediction is affected by multiple factors, but most of the existing photovoltaic output prediction studies only consider meteorological and historical sequence elements of photovoltaic output, without fully considering the influencing factors of photovoltaic output. Second, the mainstream benchmark prediction methods lack a spatio - temporal feature parallel decoupling mechanism. Although the hybrid architecture integrates multi - model features, the stacked series design results in incomplete decoupling of spatial topological associations and temporal evolution laws, restricting the improvement of prediction accuracy. Finally, risk - based prediction modeling has limitations. It usually relies on a single model for risk probability inference and is prone to problems of insufficient risk interval coverage in complex scenarios.

[0004] To break through the above bottlenecks, two - layer breakthroughs need to be achieved: First, a spatio - temporal feature parallel decoupling - complementary enhancement mechanism needs to be constructed to achieve multi - dimensional information fusion by deeply decoupling spatial topological associations and dynamic temporal evolutions. Second, a hybrid risk - based prediction framework needs to be developed, integrating the advantages of non - distribution assumptions of kernel density estimation and the ability to capture temporal dependencies of quantile regression. Although the research on hybrid models has been gradually deepened, its parallel decoupling architecture design and uncertainty quantification technology are not yet mature, making it difficult to meet the requirements of power grid systems for efficient and accurate simulation. Summary of the Invention

[0005] 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, realizing risk - based prediction of photovoltaic output through the integration of multi - source driving factor sets of regional photovoltaic output, construction of an original feature matrix, a spatio - temporal feature parallel decoupling and complementary enhancement module, a photovoltaic output benchmark prediction, a basic risk - based prediction, and a hybrid risk - based prediction framework.

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

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

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

[0009] S2: Synchronize the multi-source driving factor set for photovoltaic output under a unified spatio-temporal reference through spatial reference and time dimension synchronization and feature splicing to construct an original feature matrix;

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

[0011] S4: Construct a reference prediction model for photovoltaic output based on a sparse attention Transformer, input the spatio-temporal feature matrix of photovoltaic output after feature enhancement into the reference prediction model for photovoltaic output to obtain a reference prediction result for photovoltaic output and the corresponding error;

[0012] S5: Construct a risk prediction model based on data-driven and statistical theory, input the errors of the reference prediction results of photovoltaic output into two risk prediction models respectively, calculate the boundary values of the prediction error risk interval, and superimpose them on the reference prediction results to obtain respective risk prediction results for photovoltaic output;

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

[0014] Preferably, in S1, constructing a multi-source driving factor set for photovoltaic output includes:

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

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

[0017] S13: Construct a meteorological data adapter to convert the original 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: Construct an equipment data adapter to convert various equipment parameters into a form that meets the model input;

[0019] S15: Construct a photovoltaic output simulator based on physical mechanisms, and use the simulated values of photovoltaic output driven by physical mechanisms as the input features of the model.

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

[0021] S31: Construct a spatio-temporal feature parallel decoupling module, and use a GRU model improved based on a time-domain attention mechanism, i.e., TGRU, to extract the temporal features in the original feature matrix of photovoltaic output. The spatial features in the original feature matrix are extracted by a convolutional attention module improved based on multi-scale recursive attention and dynamic weight coupling mechanism, i.e., ICBAM;

[0022] S32: Construct a spatio-temporal feature complementary enhancement module to fuse the temporal features and spatial features decoupled by TGRU and ICBAM;

[0023] Among them, TGRU includes:

[0024] Update gate

[0025] ;

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

[0027] Reset gate

[0028] ;

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

[0030] Time-domain attention

[0031] ;

[0032] ;

[0033] In the formula, represents the time-domain 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 feature enhanced by the attention mechanism;

[0034] Candidate state

[0035] ;

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

[0037] State update

[0038] ;

[0039] where h new represents the updated hidden state, and the hidden state at the final moment is the global representation l of the time dimension feature of the photovoltaic output feature set 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 features of channel td, Row represents the height of the feature map, Col represents the width of the feature map, and row and col are the position indices respectively. l td,row,col represents the feature value at row row and col col in channel td, and l max represents the maximum response value among all the feature values of this channel;

[0047] Shared MLP structure: Two fully connected layers are adopted, and the formula is:

[0048] ;

[0049] ;

[0050] where MLP(l avg ), MLP(l max ) represent the average and max pooling features after non-linear transformation by the MLP structure respectively, They are the weight matrices of each fully connected layer, δ is the ReLU activation function, and b1 and b2 are the position offset control parameters of each fully connected layer respectively;

[0051] Weight generation:

[0052] ;

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

[0054] Feature weighting:

[0055] ;

[0056] In the formula, l channel represents the output of the channel attention module;

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

[0058] Channel compression: Perform average pooling and max pooling on the in the channel dimension to obtain , represents the average pooling value of the output of the channel attention module, represents the max pooling value of the output of the channel attention module;

[0059] Feature concatenation:

[0060] ;

[0061] In the formula, l cat represents the channel compressed feature after feature concatenation;

[0062] Convolution kernel design:

[0063] ;

[0064] In the formula, M s (l cat ) represents the concatenated feature after convolution operation;

[0065] Spatial weighting:

[0066] ;

[0067] In the formula, l spatial represents the feature output by the spatial attention module;

[0068] Dynamic weight coupling mechanism:

[0069] Design a dynamic weight controller to automatically adjust the fusion ratio of channel attention and spatial attention according to input features, and improve the adaptive ability to different scenarios:

[0070] ;

[0071] In the formula, l ICBAM represents the spatial feature extracted by ICBAM, is the coupling parameter of channel attention and spatial attention;

[0072] Construct a spatio-temporal feature complementary enhancement module to fuse the time feature and spatial feature decoupled by TGRU and ICBAM, including:

[0073] After aligning the time and spatial feature channels, input the feature after tensor splicing into a 3D convolution kernel to capture spatio-temporal local correlation:

[0074] ;

[0075] In the formula, l fused represents the fused feature after fusing the time and spatial features;

[0076] Use cross-modal residual connection to enhance the fused feature vector, and align the original time and spatial features through one-dimensional convolution and then stack them on the fused feature:

[0077] ;

[0078] In the formula, represents the spatio-temporal feature matrix of photovoltaic output after feature enhancement.

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

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

[0081] ;

[0082] In the formula, are the query, key, and value matrices, and d k is the dimension scaling factor;

[0083] The dynamic sparse attention of the encoder:

[0084] ;

[0085] ;

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

[0087] Preferably, S5: Construct a risk-based prediction model based on data-driven and statistical theory. Input the prediction result errors of the photovoltaic output benchmark prediction into two risk-based prediction models respectively, calculate the boundary values of the prediction error risk interval, and superimpose them on the benchmark prediction result to obtain the respective risk-based prediction results of the photovoltaic output, including:

[0088] S51: Construct a risk-based prediction model based on data-driven. Relying on the quantile regression framework, construct a multi-quantile loss function, and synchronously output the error quantiles corresponding to different risk levels during the time series modeling process, and superimpose them on the benchmark prediction result to obtain the basic risk-based prediction result of the photovoltaic output driven by data;

[0089] S52: Construct a risk-based prediction model based on statistical theory. Use an adaptive kernel function and local bandwidth adjustment to improve the traditional kernel density estimation method to obtain the IKDE method, and superimpose the boundary of the prediction error risk interval on the benchmark prediction result to obtain the basic risk-based prediction result of the photovoltaic output based on statistical theory;

[0090] Among them, the S51 includes:

[0091] ;

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

[0093] The improved QRLSTM, namely IQRLSTM, considers the inherent property of the quantile and adds a quantile constraint penalty term to the objective function,

[0094] ;

[0095] ;

[0096] ;

[0097] In the formula, 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] Dividing the photovoltaic output reference prediction error according to the timestamp;

[0100] Distance calculation:

[0101]

[0102] In the formula, fac1 and fac2 respectively represent the elements for which the distance calculation is to be performed;

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

[0104] Dynamic bandwidth adjustment:

[0105] ;

[0106] ;

[0107] In the formula, represents the local density sensitive factor, controlling the expansion amplitude of the sparse area, represents the global smoothing factor, h global represents the global bandwidth, represents the data standard deviation, h i is the bandwidth;

[0108] Adaptive kernel function, adopting an improved Gaussian kernel function:

[0109] .

[0110] Preferably, in the S6, constructing a photovoltaic output hybrid risk prediction framework, inputting the results of two risk prediction models for coupling, and optimizing the coupling parameters by using a multi-objective optimization algorithm to obtain the photovoltaic output hybrid risk prediction result, including:

[0111] S61: Adopting a "sine mapping - spiral flight - Levy flight" triple mutation strategy to propose an improved single-objective tribal competition and member cooperation optimization algorithm, namely ICTCM;

[0112] S62: Using a multi-objective adaptation mechanism to reconstruct, realizing the transformation from a single-objective to a multi-objective optimization framework, namely MOICTCM;

[0113] S63: By coupling the risk prediction results of IKDE and IQRLSTM through MOICTCM, the hybrid risk prediction results of photovoltaic power output are obtained;

[0114] Among them, the S62 includes:

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

[0116] Objective vectorization:

[0117] ;

[0118] In the formula, F represents the objective vector, Ang represents the dimension of the objective function, and f Ang represents the sub-objective of the maximum dimension;

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

[0120] ;

[0121] In the formula, Ang represents the dimension of the objective function, and f ang () represents each sub-objective function;

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

[0123] Within each Pareto front, calculate the crowding degree of the solutions to maintain diversity:

[0124] ;

[0125] In the formula, CD represents the individual crowding degree value, O Now represents the current solution in the objective space, O Now−1 and O Now+1 represent the adjacent solutions of the current solution in the objective space respectively;

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

[0127] Tribal division and role definition: Divide tribes according to the non-dominated sorting results. Each tribe contains solutions from different Pareto layers, and the non-dominated solution with the largest crowding degree in the tribe is the chief;

[0128] Dynamic adjustment of member loyalty:

[0129] ;

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

[0131] Multi-objective velocity update formula:

[0132] ;

[0133] In the formula, p best is the individual's historical best position, g best is the position of the tribal chief, X random is the non-dominated solution randomly selected from other tribes, represents the velocity of the mem-th member of the num-th tribe in the dim-th dimension in the (te + 1)-th iteration, represents the inertia coefficient of velocity, represents the position of this member in the te-th iteration, c1 and c2 respectively represent the experience factor and the obedience factor, , respectively represent the chaotic loyalty of each member, c3 represents the escape factor, represents the chaotic random factor;

[0134] Integration of the spiral flight strategy: Introduce spiral exploration in position update and search near the Pareto front:

[0135] ;

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

[0137] S623: External archive management and retention:

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

[0139] Levy flight perturbation: Apply Levy perturbation to the archive solutions every Ge generations to jump out of the local optimum globally and discover potential solutions:

[0140] ;

[0141] In the formula, respectively represent the archive populations in the te-th and (te + 1)-th iterations, represents the step size parameter, , obey the random normal distribution, the value range is (0, 2], represents the gamma function;

[0142] Clash and escape mechanism:

[0143] Clash trigger condition: Randomly select two tribes Tribe every Ge generations A and Tribe B , and compare their hypervolume contributions HV:

[0144] ;

[0145] Escape strategy: The escape strategy guides the failed tribes to learn from the winning tribes to accelerate convergence. The speed of the members of the failed tribes is updated as:

[0146] ;

[0147] In the formula, represents the speed of the members of the failed tribe A in the te-th iteration process, represents the updated speed of the members of tribe A, c4 represents the escape factor, represents the updated random factor, X B,chief represents the position of the chief in the winning tribe B, represents the position of the members of tribe A in the te-th iteration;

[0148] The said S63 includes:

[0149] ;

[0150] ;

[0151] ;

[0152] In the formula, PINAW represents the predicted interval average bandwidth index, PICP represents the predicted interval coverage rate index, 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 the hybrid risk-type prediction, the upper bound of the IKDE risk-type prediction, and the upper bound of the IQRLSTM risk-type prediction, Low mix 、Low KDE 、Low QRLSTM respectively represent the lower bound of the hybrid risk-type prediction, the lower bound of the IKDE risk-type prediction, and the lower bound of the IQRLSTM risk-type prediction, 、 represent the maximum value of the upper bound and the minimum value of the lower bound of the prediction interval under the established risk level, Dis represents the maximum range of the prediction interval boundary, 、 Represent 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 true value of photovoltaic power output is within the risk-based prediction interval, Mo takes the value of 1; otherwise, it is 0. Represent the Mo value corresponding to the i-th sample point under a given risk level. The meanings of the other variables are the same as those in the previous text.

[0153] The present invention also provides a photovoltaic power output hybrid probability interval prediction system based on a parallel deep learning architecture. The system is used to implement the foregoing method. The system includes: a dataset 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 dataset construction module is used to construct a multi-source driving factor set of photovoltaic power output;

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

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

[0157] The first prediction module is used to construct a reference-type prediction model of photovoltaic power output based on a sparse attention Transformer, input the spatio-temporal feature matrix of photovoltaic power output after feature enhancement into the reference-type prediction model of photovoltaic power output, and obtain a reference-type prediction result of photovoltaic power output and the corresponding error;

[0158] The second prediction module is used to construct a risk-based prediction model based on data-driven and statistical theory, input the errors of the reference-type prediction results of photovoltaic power output into two risk-based prediction models respectively, calculate the boundary values of the prediction error risk interval, and superimpose them on the reference-type prediction results to obtain the risk-based prediction results of photovoltaic power output for each;

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

[0160] Compared with the prior art, the beneficial effects of the present invention are:

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

[0162] By constructing a multi-source data adapter and a photovoltaic output simulator based on physical mechanisms, it is possible to effectively integrate a multi-dimensional set of influencing factors including geographical location information, device parameters, physically simulated output, meteorological data, and historical photovoltaic output. Compared with the traditional factor set that only considers meteorological conditions and historical photovoltaic output, the multi-source factor set can more comprehensively reflect various key factors affecting the power generation performance of photovoltaic power plants. Geographical location information determines the conditions and methods of solar radiation received at a specific location, and device parameters incorporate the electrical characteristics and mechanical installation conditions of photovoltaic modules. The photovoltaic output simulator based on physical mechanisms can combine actual meteorological data to provide simulation results with a certain degree of accuracy.

[0163] 2. Achieve parallel decoupling and complementary enhancement of the spatio-temporal characteristics of photovoltaic output:

[0164] By synergistically integrating the complementary advantages of TGRU and ICBAM to construct a spatio-temporal characteristic parallel decoupling architecture, it is possible to effectively separate the dynamic change laws of time series and the spatial heterogeneity characteristics in the input of photovoltaic output prediction. Compared with the traditional stacked hybrid architecture with a single feature transfer path and difficulty in achieving dynamic mutual feedback, this architecture can more accurately capture the complex laws of photovoltaic output affected by time-frequency mutation characteristics and spatial heterogeneity characteristics.

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

[0166] The sparse attention Transformer quickly locks key features through a sparse pattern, realizes optimization of computational efficiency and dynamic feature focusing, strengthens the response to mutation scenarios, and significantly improves the timeliness of benchmark prediction. The global self-attention of the traditional Transformer has a high computational complexity and is difficult to handle high-dimensional long-time series data (such as meteorological sequences and device status time series) in photovoltaic output prediction. Sparse attention significantly reduces the computational complexity through local window constraints.

[0167] 4. Enhance the reliability of the risk-based prediction of photovoltaic output:

[0168] IKDE fits the probability density function to the benchmark prediction error of photovoltaic output at the same historical timestamp and inversely solves the cumulative distribution function to obtain the upper and lower boundaries of the error value. IQRLSTM constructs a multi-quantile loss function and synchronously outputs the error quantiles corresponding to different risk levels during the time series modeling process. The risk interval boundaries of the photovoltaic output prediction error obtained by the two risk-based prediction methods are respectively superimposed on the benchmark prediction results to generate their respective risk-based prediction ranges of photovoltaic output. The coupling parameters of the two basic risk-based prediction results are optimized by the MOICTCM method to balance the dual requirements of risk interval coverage rate and risk interval width. Description of the Drawings

[0169] To more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0170] Figure 1 Schematic flow chart of the photovoltaic output hybrid probability interval prediction method based on a parallel deep learning architecture in an embodiment of the present invention;

[0171] Figure 2 Schematic flow chart of constructing a multi-source driving factor set for photovoltaic output in an embodiment of the present invention;

[0172] Figure 3 Schematic flow chart of spatio-temporal feature parallel decoupling and complementary enhancement in an embodiment of the present invention;

[0173] Figure 4 Schematic flow chart of constructing a risk-based prediction model library based on data-driven and statistical theories in an embodiment of the present invention. Detailed implementation manners

[0174] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0175] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the drawings and specific implementation manners.

[0176] Embodiment 1

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

[0178] Insufficient consideration of influencing factors for PV output: Although existing methods are based on meteorological condition parameters and historical output data for modeling, there is a lack of systematic integration of multi-dimensional correlation factors. Specifically, it is manifested as the absence of three aspects of mechanisms. First, geographical feature parameters (latitude and longitude, altitude) directly affect the spatio-temporal distribution characteristics of solar radiation flux by changing the atmospheric mass coefficient and surface reflectivity. Second, equipment technical parameters (component photoelectric conversion efficiency, array tilt / azimuth angle) determine the effective capture rate of the system for incident radiant energy, but no dynamic coupling relationship has been established in the existing models. Third, the irradiance distribution simulation based on physical simulation (including shadow occlusion, diffuse reflection component) can provide input source terms with spatial resolution, and such high-precision data has not been effectively incorporated into the prediction system. The disconnection of the above multi-dimensional influence mechanisms makes it difficult for the model to accurately characterize the output response characteristics of PV power plants, ultimately resulting in a significant weakening of the prediction accuracy and reliability under complex working conditions.

[0179] Decoupling and disharmony of spatio-temporal characteristics: Due to the single feature transfer path in the traditional stacked hybrid architecture, it is 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 accumulation and 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, and this spatial heterogeneity characteristic has an important impact on output prediction. 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 leads to the difficulty for the prediction model to accurately capture the changes in the spatial heterogeneity characteristics of the meteorological field under complex meteorological conditions, thereby causing the accumulation and amplification of prediction errors.

[0180] Fragmentation of risk-based prediction methods: In the current research on PV output risk interval prediction, the separate application of probability distribution modeling and time series quantile prediction leads to the difficulty in effectively coupling statistical inference and dynamic evolution characteristics. Probability distribution modeling captures the stochastic characteristics of output fluctuations through non-parametric density estimation, but its static modeling paradigm cannot represent the time-lag effect of meteorological factors. Although time series quantile prediction can directly construct a dynamic risk interval based on historical sequences, it lacks explicit constraints on the probability distribution form of output (such as multi-modality, heteroscedasticity). The two are endogenously related in risk interval construction. Probability distribution constraints can improve the statistical rigor of interval confidence, and time series dynamic modeling ensures the tracking ability of the interval for the evolution of weather processes. However, the isolated application framework of existing methods leads to the decoupling of two types of information flows, either over-relying on historical distributions and causing conservative decisions in power plant regulation, or one-sidedly tracking time series fluctuations and amplifying the failure risk of extreme weather. This lack of system coupling makes interval prediction under complex climate scenarios face the dual dilemmas of coverage bias and reliability attenuation.

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

[0182] S1. Construct a multi-source driving factor set for photovoltaic output: First, systematically collect the research data required within the photovoltaic power station area, preprocess the multi-source heterogeneous data, and construct a set of influencing factors for photovoltaic output prediction. Using a photovoltaic output simulator based on physical mechanisms, obtain the numerical solution of photovoltaic output driven by numerical weather forecast data, and the calculation results are incorporated into the photovoltaic output prediction feature set for constructing an original feature matrix containing spatial topological association information and time dynamic evolution characteristics. The specific process is as Figure 2 shown.

[0183] S11. Define the research area, and collect multi-source heterogeneous data such as historical output data of photovoltaic power stations within the area, numerical weather forecast data, meteorological data, terrain data, and physical parameters of station equipment.

[0184] S12. Geographic data adapter: To quantify the coupling mechanism between the geographical attributes of photovoltaic power stations and the solar radiation energy field, this method constructs a spatial feature fusion architecture based on a digital elevation model (DEM) and an irradiance distribution matrix. First, align the elevation raster data and the irradiance sampling matrix through a spatial interpolation algorithm to establish a grid system with a unified geographical coordinate reference. Perform matrix operations on the elevation matrix AL (representing the terrain undulation at the site) and the irradiance matrix RA (reflecting the radiation flux at the site) to obtain an elevation-irradiance correlation matrix, where each element represents the elevation-irradiance correlation at the photovoltaic site. The elevation-irradiance correlation matrix introduces spatial topological constraints, and each element contains information about surrounding grid points, enhancing 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 the irradiance matrix to the same scale

[0186] ( )

[0187] ( )

[0188] In the formula, AL scaled , RA scaled respectively represent the normalized elevation and irradiance matrices.

[0189] 2. Weighted average processing

[0190] Assign weights to each element according to the Kendall correlation between elevation, irradiance, and photovoltaic output, and then perform weighted averaging.

[0191] ( )

[0192] In the formula, RAL represents the matrix that initially fuses elevation and irradiance information, and w AL and w RA respectively represent the weight parameters of the elevation and irradiance matrix assigned based on correlation.

[0193] 3. Introduce spatial topological constraints

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

[0195] ( )

[0196] Among them, YS ma represents the element of the elevation-irradiance correlation matrix corresponding to the target photovoltaic power station at position (m, a), Qu ma represents the neighborhood of the target photovoltaic power station, represents the positions of other power stations within the neighborhood of the target photovoltaic power station, and RAL ma , represent the elements corresponding to the target photovoltaic power station and the power stations within its neighborhood in the matrix that initially fuses elevation and irradiance information, is the spatial weight coefficient between each grid point in the neighborhood and the photovoltaic power station.

[0197] S13. Meteorological data adapter: Convert the original meteorological data (such as irradiance, temperature, etc.) into a format suitable for the photovoltaic power output prediction model and ensure that the data is aligned with the elevation grid of the geographical data. The input data includes the original meteorological data, target grid parameters, geographical data containing elevation information, etc. The format of the original meteorological data may be data in a latitude-longitude grid or in a certain projection coordinate system. First, determine the coordinate system of the original data and perform necessary projection conversions to integrate the data into a unified coordinate system. Subsequently, use the bilinear interpolation method to grid the meteorological data to the target resolution to ensure that the meteorological data at the grid points is aligned with the elevation grid of the geographical data. Bilinear interpolation uses the values of the four nearest known data points around to estimate the value of the unknown point. For each meteorological variable, find the four nearest data points around the photovoltaic site in the original data, and then obtain the meteorological data value of the photovoltaic site by averaging the values of the four points.

[0198] S14. Device Data Adapter: The time series of photovoltaic power station device parameters is segmented by a dynamic sliding window. Among them, time-varying data retains the original time series record (such as photovoltaic module temperature, panel received irradiance), and time-invariant data broadcasts to align the time window (such as module tilt angle, inverter model), explicitly representing the coupling state of the device's dynamic operation and static attributes, and providing input features that integrate physical constraints for the prediction model.

[0199] S15. Physical Mechanism-based Photovoltaic Output Simulator: The simulated photovoltaic output value driven by the physical mechanism is used as the input feature of the model. Essentially, it constructs a dual-channel collaborative modeling paradigm of physical constraint and data-driven. The benchmark simulated photovoltaic output value output by the physical simulator contains strong prior knowledge of the solar radiation transfer equation and the thermoelectric conversion mechanism. Under extreme weather conditions, the physical constraint deduction mechanism based on the atmospheric motion equation in numerical weather prediction (NWP) can effectively compensate for the extrapolation limitations of the pure data-driven model in the super-sample space scenario. The calculation process is as follows:

[0200] ( )

[0201] ( )

[0202] ( )

[0203] ( )

[0204] In the formula, SP g,t represents the physical simulated photovoltaic output of the gth photovoltaic power station at the tth moment, α is the solar radiation intensity rate, is the solar radiation intensity of the numerical weather prediction at the tth moment, G stc is the solar radiation intensity under standard test conditions, 1000 W / m 2 , is the temperature of the solar panel at the tth moment, is the air temperature of the numerical weather prediction at the tth moment, C noc is the temperature of the normally operating solar panel, usually taken as , C stc is the temperature under standard test conditions, 25 °C, P stc is the photovoltaic output under standard test conditions, β is the thermal loss efficiency corresponding to the photovoltaic cell, is the area of the photovoltaic panel of the gth photovoltaic power station, SP t is the sum of the outputs of each photovoltaic power station, NM PV is the number of photovoltaic power stations.

[0205] S2 Original Feature Matrix Construction: Under the unified spatio-temporal benchmark, through the synchronization of the spatial benchmark and the time dimension, and the feature splicing operation, multi-source heterogeneous data such as the geography, meteorology, physical simulation output, and historical output of the photovoltaic power station are linked to construct the original feature matrix.

[0206] S21 Spatial Alignment: Map the geographical data, meteorological data, physical simulation output, and historical output to the same grid to ensure that the resolution and range of each element grid are consistent. First, identify the original coordinate systems of each data source, which may include geographical coordinate systems or various projection coordinate systems, and use coordinate transformation algorithms to convert all data to the coordinate system that matches the geographical data. Then, through grid processing, bilinear interpolation is used to obtain the values of each element at each photovoltaic site.

[0207] S22 Time Alignment: When the time resolutions of the geographical data, meteorological data, physical simulation output, and historical output are different, all time series data need to be aligned with the smallest time granularity. For data with a higher time resolution, linear interpolation is used for downsampling. For data with a lower time resolution, linear interpolation is also used for interpolation to fill in the missing time points to ensure that all time series data are aligned in the time dimension.

[0208] S23 Feature Splicing: Through spatial alignment and time alignment processing, dynamic features such as irradiance, temperature, humidity, and historical output form time series data with a unified grid and resolution. Static feature parameters (such as component tilt angle, inverter model, etc.) do not change with time and are extended to all time steps through a broadcast operation to ensure that the static features at each time step correspond to the dynamic features. Finally, using tensor operation techniques, the static feature tensor and the dynamic feature tensor are spliced along the feature dimension to form the final original feature matrix. This tensor not only contains the dynamic changes in the time series but also covers the distribution characteristics in space, providing comprehensive data input for the photovoltaic output prediction model.

[0209] S3. Spatio-Temporal Feature Parallel Decoupling and Complementary Enhancement: Based on the original feature matrix of photovoltaic output, a feature extraction module based on a parallel deep learning architecture is constructed using the GRU model with a time-domain attention mechanism (TGRU) and an improved convolutional attention module (ICBAM) respectively to perform spatio-temporal feature parallel decoupling. The decoupled time features and space features are input into the spatio-temporal feature complementary enhancement module for fusion to enhance the expression ability of the features. The content of each part is as Figure 3 shown.

[0210] S31. Spatiotemporal Feature Parallel Decoupling: Introduce a temporal attention mechanism into the GRU to construct a TGRU model, extract the temporal dimension features in the original feature matrix, and enhance the weights of key time steps. Based on multi-scale recursive attention and dynamic weight coupling mechanisms, improve the convolutional attention module to obtain the ICBAM module and enhance the spatial feature extraction ability.

[0211] S311. Temporal Feature Decoupling: Use the TGRU model to separate the temporal dimension features in the original feature matrix to better capture the changing patterns of time series. The TGRU synthesizes temporal dependencies and the current state, locates key time points by encoding the temporal patterns in historical states, can capture features such as irradiance and temperature at the current moment, and extract temporal patterns such as "a sharp drop in power usually follows cloud movement". The key steps of the TGRU are as follows:

[0212] 1. Update Gate

[0213] ( )

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

[0215] 2. Reset Gate

[0216] ( )

[0217] In the formula, re represents the output of the reset gate, and W r represents the weight parameter of the reset gate.

[0218] 3. Temporal Attention

[0219] ( )

[0220] ( )

[0221] In the formula, 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, and l attn represents the feature strengthened by the attention mechanism.

[0222] 4. Candidate State

[0223] ( )

[0224] In the formula, h tilde represents the candidate hidden state, W tilde represents the weight for calculating the candidate hidden state, and tanh is the hyperbolic tangent function.

[0225] 5. State update

[0226] ( )

[0227] In the formula, h new represents the updated hidden state, and the hidden state at the final moment is the global representation l of the time dimension features of the PV output feature set TGRU .

[0228] S312. Spatial feature decoupling: Based on the multi-scale recursive attention and dynamic weight coupling mechanism, the convolutional attention module is improved to construct the ICBAM module to separate the spatial dimension features in the original feature matrix. The single-scale attention in the traditional CBAM is difficult to capture the details and global correlations in complex scenarios. Pooling operations are introduced in the channel attention to generate multi-scale channel weights. The spatial attention constructs the receptive field through dilated convolution to enhance the cross-scale feature interaction. The ICBAM module is suitable for tasks with drastic illumination changes and variable target scales. Specifically as follows:

[0229] 1. Multi-scale recursive attention:

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

[0231] ① Multi-scale pooling layer

[0232] Global average pooling: ( )

[0233] Global max pooling: ( )

[0234] In the formula, td represents the channel index, l avg represents the global average value of the features 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 indices respectively, and l td,row,col represents the feature value at row row and col col in channel td, and l max represents the maximum response value among all the feature values of this channel.

[0235] ② Shared MLP structure: Two fully connected layers are adopted, and the formula is:

[0236] ( )

[0237] ( )

[0238] In the formula, MLP(l avg ), MLP(l max ) represent the average and max-pooling features after the non-linear transformation of the MLP structure respectively. are the weight matrices of each fully connected layer respectively, δ is the ReLU activation function, and b1, b2 are the position offset control parameters of each fully connected layer respectively.

[0239] ③ Weight generation:

[0240] ( )

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

[0242] ④ Feature weighting:

[0243] ( )

[0244] In the formula, 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 max pooling on in the channel dimension. , represents the average pooling value of the output of the channel attention module, and represents the max pooling value of the output of the channel attention module.

[0247] ② Feature concatenation:

[0248] ( )

[0249] In the formula, l cat represents the channel compressed feature after feature concatenation.

[0250] ③ Convolution kernel design:

[0251] ( )

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

[0253] ④ Spatial weighting:

[0254] ( )

[0255] where l spatial represents the features output by the spatial attention module.

[0256] 2. Dynamic weight coupling mechanism

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

[0258] ( )

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

[0260] S32. Spatiotemporal feature complementary enhancement module: Fuse the time features and spatial features decoupled by TGRU and ICBAM to enhance the feature expression ability and improve the accuracy of photovoltaic power output prediction.

[0261] S321. Feature fusion: After aligning the time and spatial feature channels, input the features after tensor concatenation into a 3D convolution kernel to capture spatiotemporal local correlations:

[0262] ( )

[0263] where l fused represents the fused features after fusing the time and spatial features.

[0264] S322. Feature enhancement: Use cross-modal residual connections to enhance the fused feature vectors, align the original time and spatial features through one-dimensional convolution and then stack them on the fused features to avoid information loss:

[0265] ( )

[0266] where represents the spatiotemporal feature matrix of photovoltaic power output after feature enhancement.

[0267] S4. Photovoltaic Output Benchmark Prediction Model: Since the photovoltaic output is jointly affected by local meteorological mutations (such as cloud occlusion) and long-term seasonal trends, a fixed window size is difficult to adapt to the changes in photovoltaic output. A dynamic sparse window is introduced to construct a photovoltaic output benchmark prediction model based on a sparse attention Transformer. By using the sparse attention mechanism, high-correlation positions are retained for calculation. Only the encoder uses sparse attention to process long input sequences, and the decoder retains the standard attention, reducing the computational cost and enhancing the long-sequence modeling ability. The key steps are as follows:

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

[0269] ( )

[0270] In the formula, are the query, key, and value matrices, and d k is the dimension scaling factor.

[0271] 2. Dynamic sparse attention of the encoder

[0272] ( )

[0273] ( )

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

[0275] S5. Risk-based Prediction Model Library Based on Data-driven and Statistical Theory: The improved QRLSTM (IQRLSTM) relies on the quantile regression framework and directly outputs the quantiles of photovoltaic output errors corresponding to different risk levels during the time series modeling process. Finally, by superimposing the predicted quantile results of the errors output by it on the benchmark prediction value, the risk-based prediction result can be obtained. The traditional kernel density estimation method is improved using an adaptive kernel function and local bandwidth adjustment to establish the IKDE method. IKDE fits the probability density function of the benchmark prediction error data of photovoltaic output at the same timestamp, and then derives the cumulative probability distribution curve, thereby extracting the error threshold based on the specified risk level. The upper and lower bounds of the error value are inversely solved through the cumulative distribution function and superimposed on the benchmark prediction result to generate the risk-based prediction range of photovoltaic output. 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 requires the independent and identical distribution assumption of the benchmark prediction error. The content of each part is as Figure 4 shown.

[0276] S51 Risk-based Prediction Model Based on Data-driven: IQRLSTM relies on the quantile regression framework to construct a multi-quantile loss function and synchronously outputs the quantiles of errors corresponding to different risk levels during the time series modeling process. By optimizing the quantile loss function, this model dynamically captures the time series dependence of photovoltaic output prediction errors and realizes the end-to-end prediction of the upper and lower quantiles of errors. Finally, by superimposing the predicted quantile results of the error values output by IQRLSTM on the benchmark prediction value, the risk-based prediction result of photovoltaic output can be obtained. The model can be transformed into an optimization problem as shown in the formula.

[0277] ( )

[0278] In the formula, 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 predicted value of photovoltaic output corresponding to the quantile, represents the quantile, and J represents the total number of quantiles.

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

[0280] ( )

[0281] ( )

[0282] ( )

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

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

[0285] 1. Divide the photovoltaic output benchmark - type prediction error according to the timestamp. Taking the intra - day time as the division basis, the error values at the same time are extracted as a data set.

[0286] 2. Distance calculation

[0287] ( )

[0288] In the formula, fac1 and fac2 respectively represent the elements for which the distance calculation is to be performed.

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

[0290] 4. Dynamic bandwidth adjustment

[0291] ( )

[0292] ( )

[0293] In the formula, represents the local density - sensitive factor (default 0.5), which controls the expansion amplitude in the sparse region, 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 is large, the data points are sparse around, and the bandwidth h should be increased i to avoid overfitting. Conversely, the bandwidth h should be reduced i to improve the resolution.

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

[0295] ( )

[0296] S6. Hybrid risk-based prediction framework based on the multi-objective tribal competition and member cooperation optimization algorithm: Adopting the "sine mapping - spiral flight - Levy flight" triple mutation strategy, an improved single-objective tribal competition and member cooperation optimization algorithm (ICTCM) is proposed. While maintaining the population diversity, the global exploration ability and convergence stability of the single-objective tribal competition and member cooperation optimization algorithm are enhanced. Using the multi-objective adaptation mechanism reconstruction, the transformation from single-objective to multi-objective optimization framework is realized (MOICTCM). By coupling the risk-based prediction results of IKDE and IQRLSTM through MOICTCM, the hybrid risk-based prediction result of photovoltaic output is obtained.

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

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

[0299] The concept of CTCM originates from the inter-tribal competition mechanism in nature and the cooperation mode among members within the tribe. The core of the CTCM algorithm is the local development within the tribe, the global search among tribes, and the dynamic loyalty of members to the tribe. The CTCM algorithm optimizes the local solution through member cooperation and performs global search using the competitive relationship among tribes, thereby exploring better solutions in the solution space. Finally, the algorithm adjusts the behavior of tribal members in the process of cooperation and competition, discriminates the dynamic loyalty of members, and ensures the flexibility and adaptability of the algorithm in the optimization process. The optimization process is mainly as follows:

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

[0301] ( )

[0302] ( )

[0303] Wherein, 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 Fit of the primitive society is as follows:

[0305] ( )

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

[0307] ( )

[0308] ( )

[0309] In the tribal structure, the chief is responsible for managing and planning the future development of the tribe, and most members tend to follow the guidance and plans of the chief. However, the personal opinions of the members lead to the volatility of the members' loyalty to the chief over time, and the dynamic changes in loyalty reflect the complexity and diversity of individual behaviors within the tribe. CTCM uses a sine chaotic map to characterize the changes in members' loyalty, and the loyalty r of an individual member t can be expressed as:

[0310] ( )

[0311] Update the velocity matrix:

[0312] ( )

[0313] Wherein, represents the velocity of the mem-th member of the num-th tribe in the dim dimension in the (te + 1)-th iteration, represents the inertia coefficient of the velocity, represents the position corresponding to the best fitness value of this member in the cycle, represents the position of this member in the te-th iteration, represents the position corresponding to the best fitness value of this tribe in the cycle, c1 and c2 respectively represent the experience factor and the obedience factor, , respectively represent the chaotic loyalties of each member.

[0314] Random conflicts may occur between tribes. At this time, the weaker tribes will flee, while the stronger tribes will be unaffected. The velocity update is as follows:

[0315] ( )

[0316] ( )

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

[0318] For the position update in the competition process, see Equation (49). When the updated position exceeds the boundary, a velocity mirror bounce operation is performed.

[0319] ( )

[0320] ( )

[0321] In the formula, represents the position of the mem-th member of the num-th tribe in the dim dimension in the (te + 1)-th iteration.

[0322] S612 Improvement Strategy

[0323] CTCM has good stability and high computational accuracy, but still has limitations. On the one hand, in the population initialization process, the random distribution of the initial positions and the differences in the individual populations easily lead to the lack of population diversity. On the other hand, inappropriate algorithm parameter settings result in insufficient perturbation or too fast convergence, causing the algorithm to fall into a local optimum. Therefore, the CTCM algorithm is improved in the following three aspects:

[0324] 1. Sine Mapping

[0325] The sine mapping is a chaotic system based on the sine function, which generates a chaotic sequence through the nonlinear transformation of the sine function and usually has the characteristics of combining periodicity and chaos. Coupling the sine mapping helps to increase population diversity and enhance the global search ability. The specific formula is as follows:

[0326] ( )

[0327] In the formula, X te , X te+1 represent the populations in the te-th and (te + 1)-th iterations respectively, is the chaotic system parameter.

[0328] 2. Spiral Flight

[0329] In the basic CTCM, the performance of the algorithm depends on the information exchange among tribes and the learning strategies of individuals. Usually, tribe members tend to move closer to the optimal position of the tribe to achieve efficient resource utilization. However, the centralized search strategy may lead to the premature convergence of the algorithm to the local optimal solution, restricting the global search ability of the algorithm. Therefore, a spiral flight strategy is introduced to improve the CTCM. The random walk characteristic of the spiral flight can generate long-distance jumps during the search process, which can not only perform fine search within the local area 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 search efficiency and global search ability of the algorithm in complex optimization problems. The specific formula is as follows:

[0330] ( )

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

[0332] 3. Levy flight

[0333] After the iterative optimization is completed, the Levy flight strategy is added to forcibly adjust the individual positions, enhancing the exploration ability of the algorithm and ensuring that it can jump out of the local optimal solution during the optimization process, so as to find a better solution in the global solution space.

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

[0335] ( )

[0336] In the formula, 、 follow a random normal distribution and can be expressed as:

[0337] ( )

[0338] In the formula, represents the gamma function, and the value range is (0, 2].

[0339] S62. Improved multi-objective tribal competition and member cooperation optimization algorithm: The core of transforming a single-objective optimization algorithm into a multi-objective optimization algorithm lies in introducing a multi-objective trade-off mechanism while maintaining the 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. Objective Vectorization:

[0342] ( )

[0343] In the formula, F represents the objective vector, Ang represents the dimension of the objective function, and f Ang represents the sub-objective of the maximum dimension.

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

[0345] ( )

[0346] In the formula, f ang () represents each sub-objective function.

[0347] 3. Define the Pareto dominance relationship: nu represents the population size. Traverse all solutions in the population {o1, o2, …, o nu}, calculate the non-dominance relationship of all solutions according to formula (56). All non-dominated solutions form the first-layer Pareto front PF1. Remove the first-layer Pareto front solutions from the population, and recursively extract subsequent Pareto fronts from the remaining solutions. Finally, obtain ku Pareto front layers PF1, PF2, …, PF ku , where PF1 is the optimal solution set, PF2 is the sub-optimal, and so on.

[0348] 4. Calculate the crowding degree of solutions within each layer of the Pareto front to maintain diversity:

[0349] ( )

[0350] In the formula, CD represents the individual crowding degree value, O Now represents the current solution in the objective space, and O Now−1 and O Now+1 represent the adjacent solutions of the current solution in the objective space respectively.

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

[0352] 1. Tribal Division and Role Definition: Divide tribes according to the non-dominated sorting result. Each tribe contains solutions from different Pareto layers. The non-dominated solution with the largest crowding degree in the tribe is the chief.

[0353] 2. Dynamic Adjustment of Member Loyalty:

[0354] ( )

[0355] In the formula, te represents the current iteration number, and r te represents the membership loyalty during the te-th iteration, and represents the maximum number of iterations.

[0356] 3. Multi-objective update formula:

[0357] ( );

[0358] In the formula, p best is the individual's historical best position, g best is the position of the tribal chief, X random is a non-dominated solution randomly selected from other tribes, represents the velocity of the mem-th member of the num-th tribe in the dim dimension during the (te + 1)-th iteration, represents the inertia coefficient of the velocity, represents the position of this member during the te-th iteration, c1 and c2 represent the experience factor and the compliance factor respectively, , represent the chaotic loyalties of each member respectively, c3 represents the escape factor, represents the chaotic random factor;

[0359] 4. Integration of the spiral flight strategy: Introduce spiral exploration in position update to perform refined search near the Pareto front and enhance the local development ability:

[0360] ;

[0361] In the formula, 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: Combine the parent and offspring populations, screen out the non-dominated solutions and add them to the temporary archive. If the temporary archive exceeds the capacity SI archive , eliminate the solutions from low to high according to the crowding degree.

[0364] 2. Levy flight perturbation: Apply Levy perturbation to the archived solutions every Ge generations to jump out of the local optimum in the global scope, discover potential solutions, and avoid falling into the local Pareto front:

[0365] ( )

[0366] In the formula, respectively represent the archive populations in the \(t_e\)-th and \((t_e + 1)\)-th iteration processes, represents the step size parameter.

[0367] 3. Tribe Conflict and Escape Mechanism

[0368] (1)Conflict Trigger Conditions

[0369] Randomly select two tribes, Tribe A and Tribe B every \(G_e\) generations, and compare their hypervolume contributions HV:

[0370] ( )

[0371] (2)Escape Strategy

[0372] The escape strategy guides the failed tribe to learn from the winning tribe to accelerate convergence. The velocity of the members of the failed tribe is updated as:

[0373] ( )

[0374] In the formula, represents the velocity of the members of the failed tribe A in the \(t_e\)-th iteration process, represents the updated velocity of the members of tribe A, \(c_4\) represents the escape factor, represents the updated random factor, \(X\) B,chief represents the position of the chief in the winning tribe B, represents the position of the members of tribe A in the \(t_e\)-th iteration.

[0375] S63 Hybrid Risk-based Prediction Framework

[0376] IKDE generates a risk interval in a statistical sense based on the prediction error of the photovoltaic output benchmark type, and is applicable to scenarios where the global error distribution is stable. IQRLSTM can directly model the quantiles of the time series-dependent error and adapt to local fluctuation characteristics. To integrate the complementary advantages of IKDE and IQRLSTM, a hybrid risk-based prediction framework based on MOICTCM is proposed, with the prediction interval average bandwidth (PINAW) and prediction interval coverage probability (PICP) indicators as the objective functions. This framework synergistically integrates the advantages of statistical methods and data-driven models, while ensuring the statistical reliability of the risk-based prediction interval, improving the adaptability to the non-steady state fluctuation characteristics of photovoltaic output.

[0377] ( )

[0378] ( )

[0379] ( )

[0380] In the formula, PINAW represents the predicted interval average bandwidth index, PICP represents the predicted interval coverage rate index, represents the established risk level, ω represents the risk-based prediction result coupling parameter of IKDE and IQRLSTM, Up mix 、Up IKDE 、Up IQRLSTM respectively represent the upper bound of the hybrid risk-based prediction, the upper bound of the IKDE risk-based prediction, and the upper bound of the IQRLSTM risk-based prediction. Low mix 、Low KDE 、Low QRLSTM respectively represent the lower bound of the hybrid risk-based prediction, the lower bound of the IKDE risk-based prediction, and the lower bound of the IQRLSTM risk-based prediction. 、 represent the maximum value of the upper bound and the minimum value of the lower bound of the prediction interval under the established risk level. Dis represents the maximum range of the prediction interval boundary. 、 represent the upper and lower bounds of the prediction interval at the i-th sample point under the established risk level. Mo is a boolean value. When the true value of the photovoltaic output is within the risk-based 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 the established risk level.

[0381] Embodiment 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, and the system includes: a dataset construction module, a matrix construction module, a feature fusion module, a first prediction module, a second prediction module, and a third prediction module;

[0383] The dataset construction module is used to construct a multi-source driving factor set of photovoltaic output;

[0384] The matrix construction module is used to synchronize the multi-source driving factor set of photovoltaic output under a unified space-time reference through space reference and time dimension synchronization and feature splicing to construct an original feature matrix;

[0385] The feature fusion module is used to construct a feature extraction module based on a parallel deep learning architecture, perform spatio-temporal feature parallel decoupling on the original feature matrix, and input the decoupled time features and space features into a spatio-temporal feature complementary enhancement module for fusion;

[0386] The first prediction module is used to construct a photovoltaic output benchmark prediction model based on the sparse attention Transformer, input the spatio-temporal feature matrix of the enhanced photovoltaic output into the photovoltaic output benchmark prediction model, and obtain the photovoltaic output benchmark prediction result and the corresponding error;

[0387] The second prediction module is used to construct a risk prediction model based on data-driven and statistical theory, input the prediction result errors of the photovoltaic output benchmark into two risk prediction models respectively, calculate the boundary values of the prediction error risk interval, and superimpose them on the benchmark prediction result to obtain the respective photovoltaic output risk prediction results;

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

[0389] The embodiments described above are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A photovoltaic output hybrid probability interval prediction method based on a parallel deep learning architecture, characterized in that, The method includes: S1: Construct a multi-source driving factor set for photovoltaic output; S2: Synchronize the multi-source driving factor set for photovoltaic output in a unified spatio-temporal reference through spatial reference and time dimension synchronization, and perform feature splicing to construct an original feature matrix; S3: Construct a feature extraction module based on a parallel deep learning architecture, perform spatio-temporal feature parallel decoupling on the original feature matrix, and input the decoupled time features and spatial features into a spatio-temporal feature complementary enhancement module for fusion; S4: Construct a reference-type prediction model for photovoltaic output based on a sparse attention Transformer, input the spatio-temporal feature matrix of photovoltaic output with enhanced features into the reference-type prediction model for photovoltaic output, and obtain the reference-type prediction result of photovoltaic output and the corresponding error; S5: Construct a risk-type prediction model based on data-driven and statistical theory, input the prediction result errors of the reference-type prediction of photovoltaic output into two risk-type prediction models respectively, calculate the boundary values of the prediction error risk interval, and superimpose them on the reference-type prediction result to obtain the respective risk-type prediction results of photovoltaic output; S6: Construct a hybrid risk-type prediction framework for photovoltaic output, 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 hybrid risk-type prediction result of photovoltaic output.

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

3. The method according to claim 1, wherein In the above S3, constructing a feature extraction module based on a parallel deep learning architecture, performing spatio-temporal feature parallel decoupling on the original feature matrix, and inputting the decoupled time features and spatial features into a spatio-temporal feature complementary enhancement module for fusion includes: S31: Construct a spatio-temporal feature parallel decoupling module, use a GRU model improved based on a time-domain attention mechanism, namely TGRU, to extract time features from the original feature matrix of photovoltaic output, and use a convolutional attention module improved based on a multi-scale recursive attention and dynamic weight coupling mechanism, namely ICBAM, to extract spatial features from the original feature matrix; S32: Construct a spatio-temporal feature complementary enhancement module to fuse the time features and spatial features decoupled by TGRU and ICBAM; Among them, TGRU includes: Update gate ； where z represents the output of the update gate, l represents the original feature matrix of the input, σ represents the Sigmoid function, h prev represents the hidden state at the previous moment, W z represents the weight parameter of the update gate; Reset gate ； where re represents the reset gate output, and W r represents the weight parameter of the reset gate; Time-domain attention ； ； In the formula, represents the time-domain 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 feature after the attention mechanism is strengthened; Candidate state ； where h tilde represents the candidate hidden state, and tanh is the hyperbolic tangent function; State update ； where h new represents the updated hidden state, and the hidden state at the final moment is the global representation l of the time dimension feature of the photovoltaic output feature set TGRU ; ICBAM includes: multi-scale recursive attention and dynamic weight coupling mechanism; Among them, multi-scale recursive attention includes: multi-scale channel attention module and multi-scale receptive field spatial attention module; The multi-scale channel attention module includes: Multi-scale pooling layer: Global average pooling: ; Global maximum pooling: ; where td represents the channel index, l avg represents the global average value of the features of channel td, Row represents the height of the feature map, Col represents the width of the feature map, and row and col are the position indices l td,row,col represents the feature value at row row and column col in channel td, l max represents the maximum response value among all the feature values of this channel; Shared MLP structure: Two fully connected layers are adopted, and the formula is: ； ； where MLP(l avg ) and MLP(l max ) respectively represent the average and max pooling features after non-linear transformation by the MLP structure, are the weight matrices of each fully connected layer, δ is the ReLU activation function, and b1 and b2 are the position offset control parameters of each fully connected layer respectively; Weight generation: ； where M td (l) is the output channel weight matrix, and 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: For perform average pooling and max pooling on the channel dimension to obtain , representing the average pooling value of the output of the channel attention module, representing the max pooling value of the output of the channel attention module; Feature concatenation: ； where l cat represents the channel compression feature after feature splicing; Convolution kernel design: ； Where M s (l cat ) represents the concatenated features after the convolution operation; Spatial weighting: ； where l spatial represents the feature output by the spatial attention module; Dynamic weight coupling mechanism: Design a dynamic weight controller to automatically adjust the fusion ratio of channel attention and spatial attention according to the input features, and improve the adaptive ability 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 spatio-temporal feature complementary enhancement module to fuse the time features and spatial features decoupled by TGRU and ICBAM, including: After aligning the time and spatial feature channels, input the features after tensor concatenation into a 3D convolution kernel to capture spatio-temporal local correlations: ； where l fused represents the fused feature after fusing time and space features; Use cross-modal residual connections to enhance the fused feature vectors, and align the original time and spatial features through one-dimensional convolution and then stack them to the fused features: ； In the formula, represents the spatio-temporal 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: ； Wherein, is a query, key, value matrix, and d k is a dimension scaling factor; The dynamic sparse attention of the encoder: ； ； In the formula, LocalWindow(P) represents the dynamic window centered on position P, GlobalAnchor(P) represents the global anchor 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 anchor range.

5. The method according to claim 1, characterized in that, S5: Construct a risk prediction model based on data-driven and statistical theory. Input the photovoltaic output benchmark prediction result errors into two risk prediction models respectively, calculate the boundary values of the prediction error risk interval, and stack them to the benchmark prediction result to obtain the respective photovoltaic output risk prediction results, including: S51: Construct a data-driven risk prediction model. Relying on the quantile regression framework, construct a multi-quantile loss function, and synchronously output the error quantiles corresponding to different risk levels during the time series modeling process, and stack them to the benchmark prediction result to obtain the data-driven basic risk prediction result of photovoltaic output; S52: Construct a risk prediction model based on statistical theory. Use an adaptive kernel function and local bandwidth adjustment to improve the traditional kernel density estimation method to obtain the IKDE method, and stack the boundary of the prediction error risk interval to the benchmark prediction result to obtain the basic risk prediction result of photovoltaic output based on statistical theory; Among them, S51 includes: ； where obj1 represents the IQRLSTM objective function without considering the quantile constraint penalty term, y i represents the actual observed value of PV output, represents the predicted value of PV output corresponding to the quantile, represents the quantile point, and J represents the total number of quantile points; The improved QRLSTM, namely IQRLSTM, considers the inherent properties of quantiles and adds a quantile constraint penalty term to the objective function, ； ； ； In the formula, dif represents the sum of multi-quantile losses, penalty represents the constraint penalty term, and obj represents the final loss function; S52 includes: Divide the photovoltaic output benchmark prediction error according to the timestamp; Distance calculation: ； In the formula, fac1 and fac2 respectively represent the elements to be calculated for the distance; Accelerate the nearest neighbor search using the binary tree search method, and arrange the photovoltaic output reference prediction error sequence in ascending order. Let n represent the sample size and the local neighborhood size be win. For each error point err i , expand points forward and backward in the sorted array. The maximum distance between the error point and the neighborhood boundary points is ; Dynamic bandwidth adjustment: ； ； In the formula, represents the local density sensitive factor, which controls the expansion amplitude of the sparse region, represents the global smoothing factor, h global represents the global bandwidth, represents the data standard deviation, h i is the bandwidth; Adaptive kernel function, using an improved Gaussian kernel function: 。 6. The method according to claim 5, wherein In S6, construct a photovoltaic output hybrid risk prediction framework, input the results of two risk prediction models for coupling, and use a multi-objective optimization algorithm to optimize the coupling parameters to obtain the photovoltaic output hybrid risk prediction result, including: S61: An improved single-objective tribal competition and member cooperation optimization algorithm, namely ICTCM, is proposed by adopting a triple mutation strategy of "sine mapping - spiral flight - Levy flight". S62: The single-objective to multi-objective optimization framework transformation, namely MOICTCM, is realized by using a multi-objective adaptation mechanism reconstruction. S63: The hybrid risk prediction result of photovoltaic power output is obtained by coupling the risk prediction results of IKDE and IQRLSTM through MOICTCM. Among them, the S62 includes: S621: Multi-objective fitness evaluation and Pareto ranking: Objective vectorization: ； In the formula, F represents the target vector, Ang represents the dimension of the objective function, and f Ang represents the sub-objective of the maximum dimension; Solution A dominates solution B if and only if: ； where Ang represents the dimension of the objective function, and f ang () represents each sub-objective function; Define the Pareto dominance relationship: Calculate the non-dominance relationship of all solutions, extract the first-layer non-dominated solutions, recursively extract the next-layer non-dominated solutions from the remaining solutions, and divide the population into multiple Pareto front layers. Within each Pareto front layer, calculate the crowding degree of the solutions to maintain diversity: ； Wherein, CD represents the individual crowding degree value, O Now represents the current solution in the target space, O Now−1 and O Now+1 respectively represent the adjacent solutions of the current solution in the target space; S622: Multi-objective extension of tribal structure and competition rules: Tribal division and role definition: Divide tribes according to the non-dominated sorting result. Each tribe contains solutions from different Pareto layers. The non-dominated solution with the largest crowding degree in the tribe is the chief. Dynamic adjustment of member loyalty: ； where te represents the current iteration number, and r te represents the membership loyalty during the te-th iteration, and represents the maximum number of iterations; Multi-objective speed update formula: ； Where p best is the individual's historical best position, g best is the position of the tribal chief, X random is a non-dominated solution randomly selected from other tribes, represents the velocity of the mem-th member of the num-th tribe in the dim-th dimension at the (te + 1)-th iteration, represents the inertia coefficient of the velocity, represents the position of this member at the te-th iteration, c1 and c2 represent the experience factor and the obedience factor respectively, 、 represent the chaotic loyalties of each member respectively, c3 represents the escape factor, represents the chaotic random factor; Integration of spiral flight strategy: Introduce spiral exploration in position update and search near the Pareto front. ； In the formula, represents the logarithmic spiral shape constant, represents a random number between [-1, 1]; S623: External archive management and retention: Archive update rule: Combine the parent and offspring populations, select non-dominated solutions to add to the temporary archive, and if the temporary archive exceeds the capacity SI archive , eliminate solutions from lowest to highest crowding distance; Levy flight perturbation: Apply Levy perturbation to the archived solutions every Ge generations to jump out of the local optimum globally and discover potential solutions. ； In the formula, respectively represent the archive populations in the te-th and te+1-th iteration processes, represents the step size parameter, , obeys the random normal distribution, and its value range is (0, 2], represents the gamma function; Tribal conflict and escape mechanism: Conflict triggering condition: Randomly select two tribes every Ge generations A and Tribe B , and compare their hypervolume contributions HV: ； Escape strategy: The escape strategy guides the failed tribes to learn from the winning tribes to accelerate convergence. The speed update of the members of the failed tribes is: ； In the formula, represents the velocity of the members of the failed tribe A in the te-th iteration process, represents the updated velocity of the members of tribe A, and c4 represents the escape factor, represents the updated random factor, X B,chief represents the position of the chief in the winning tribe B, represents the position of the members of tribe A in the te-th iteration; The S63 includes: ； ； ； Wherein, PINAW represents the predicted interval average bandwidth index, and PICP represents the predicted interval coverage rate index. represents the established risk level, ω represents the coupling parameter of the risk-based prediction results of IKDE and IQRLSTM, Up mix 、Up IKDE 、Up IQRLSTM represent the upper bounds of the hybrid risk-based prediction, the upper bounds of the IKDE risk-based prediction, and the upper bounds of the IQRLSTM risk-based prediction respectively. Low mix 、Low KDE 、Low QRLSTM represent the lower bounds of the hybrid risk-based prediction, the lower bounds of the IKDE risk-based prediction, and the lower bounds of the IQRLSTM risk-based prediction respectively. 、 represent the maximum value of the upper bound and the minimum value of the lower bound of the predicted interval under the established risk level. Dis represents the maximum range of the predicted interval boundary. 、 represent the upper and lower bounds of the predicted interval at the i-th sample point under the established risk level. Mo is a Boolean value. When the true value of the photovoltaic output is within the risk-based 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 the established risk level. The meanings of the remaining variables are the same as those in 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 according to any one of claims 1-6, characterized in that, The system includes: a dataset construction module, a matrix construction module, a feature fusion module, a first prediction module, a second prediction module, and a third prediction module. The dataset construction module is used to construct a multi-source driving factor set of photovoltaic power output. The matrix construction module is used to synchronize the spatial reference and time dimension and splice features for the multi-source driving factor set of photovoltaic power output under a unified spatio-temporal reference to construct an original feature matrix. The feature fusion module is used to construct a feature extraction module based on a parallel deep learning architecture, decouple the spatio-temporal features of the original feature matrix in parallel, and input the decoupled time features and spatial features into a spatio-temporal feature complementary enhancement module for fusion. The first prediction module is used to construct a benchmark prediction model of photovoltaic power output based on a sparse attention Transformer, input the spatio-temporal feature matrix of photovoltaic power output after feature enhancement into the benchmark prediction model of photovoltaic power output, and obtain the benchmark prediction result of photovoltaic power output and the corresponding error. The second prediction module is used to construct a risk prediction model based on data-driven and statistical theory, input the prediction result errors of the photovoltaic power output benchmark prediction into two risk prediction models respectively, calculate the boundary values of the prediction error risk interval, and superimpose them on the benchmark prediction result to obtain the respective risk prediction results of photovoltaic power output. The third prediction module is used to 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 of photovoltaic output.

Images