Method for constructing a photovoltaic power generation decommissioning path calculation model based on dynamic material flow

Abstract

This invention provides a method for constructing a photovoltaic (PV) power generation decommissioning path calculation model based on dynamic material flow, relating to the field of PV power generation technology. The method includes: collecting historical PV installation data of a target region to generate a first PV installation increase table; predicting future new installation capacity by combining industrial planning, capacity constraints, and energy demand; calculating PV installation outflow based on lifetime distribution function and cross-regional flow dynamic coupling coefficient, and calculating total inflow based on inventory difference; integrating historical and future data to generate a second PV installation increase table, and correcting model parameters by referencing regional data and operating condition data. This invention, through multi-dimensional dynamic factor fusion, sample refinement, and cross-regional error correction, solves the problems of insufficient prediction accuracy and failure to consider composite influencing factors in existing technologies, significantly improving the calculation accuracy of PV installation increase and decommissioning outflow, and providing reliable support for PV industry decommissioning path planning and resource optimization.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation technology, and in particular to a method for constructing a calculation model for the decommissioning path of photovoltaic power generation based on dynamic material flow. Background Technology

[0002] With the global energy transition accelerating and installed capacity continuing to increase rapidly, it means that the capacity of photovoltaic decommissioning will increase significantly in the future. Therefore, developing a calculation model for the future photovoltaic decommissioning capacity will help to achieve precise management of future photovoltaic product production and decommissioning.

[0003] However, photovoltaic modules have a fixed lifespan. As early photovoltaic projects gradually enter their retirement period, the scientific planning of photovoltaic power generation retirement paths has become a key issue in ensuring the sustainable development of the photovoltaic industry, reducing environmental risks, and optimizing resource allocation. In existing technologies, the calculation of photovoltaic power generation retirement paths relies on statistical analysis of historical installed capacity data, simple prediction of future new capacity, and estimation of retirement volume based on fixed module lifespan parameters. This method results in the prediction of future new photovoltaic installed capacity relying solely on industry planning or historical growth trends, without fully considering the dynamic impact of factors such as capacity supply constraints and energy demand fluctuations, ultimately leading to insufficient prediction accuracy.

[0004] Therefore, this invention proposes a method for constructing a photovoltaic power generation decommissioning path calculation model based on dynamic material flow. Summary of the Invention

[0005] This invention provides a method for constructing a photovoltaic power generation decommissioning path calculation model based on dynamic material flow, in order to solve the aforementioned technical problems.

[0006] This invention provides a method for constructing a photovoltaic power generation decommissioning path calculation model based on dynamic material flow, comprising: Step 1: Collect basic photovoltaic installation data for at least one target area corresponding to n1 historical years, and generate a first photovoltaic installation addition table, wherein the first photovoltaic installation addition table includes at least the actual newly added photovoltaic installation capacity and the actual photovoltaic installation inventory in use for each of the n1 historical years in the target area; Step 2: Based on the historical data of the first photovoltaic installation increase table, and combined with the preset conditions related to the development plan of the photovoltaic industry, capacity supply constraints, and energy demand trends of the target region, predict the new photovoltaic installation capacity of the target region in each of the n2 future years, and use it as the benchmark data for the change of photovoltaic installation inventory in future years; Step 3: Calculate the photovoltaic installed capacity outflow of the i-th target region in year t. ,in, Let this be the photovoltaic installation outflow in the i-th target region in year t. For the i-th target region in the th... Annual photovoltaic installation inflow; The lifetime distribution function represents the first... The cumulative extreme operating time of PV modules flowing in annually The failure probability in year t; For time interval parameters; This refers to the dynamic coupling coefficient for cross-regional circulation. Step 4: Calculate the total photovoltaic installation inflow for the i-th target region in year t. ,in, Let be the total inflow of photovoltaic installations in the i-th target region in year t; Let be the difference between the in-use photovoltaic installation inventory in year t and the in-use photovoltaic installation inventory in year t-1 of the i-th target region; Step 5: Integrate the actual photovoltaic installation data of the n1 historical years and the predicted and calculated data of the n2 future years to generate a second photovoltaic installation addition table covering the n1 historical years and the n2 future years. In the second photovoltaic installation addition table, each year records the new photovoltaic installation capacity, photovoltaic installation inflow, and photovoltaic installation outflow.

[0007] Preferably, before predicting the new photovoltaic installed capacity of the target region in each of the n2 future years, the following steps are included: By obtaining the actual newly installed photovoltaic capacity and the retired photovoltaic capacity in adjacent historical years, the theoretical newly installed photovoltaic capacity can be obtained. Information mining is performed on the target region in the adjacent historical years based on photovoltaic industry development plans, capacity supply constraints, and energy demand trends to construct a quantitative stimulus set. The quantitative stimulus set includes several stimulus indicators, stimulus constraint terms of other indicators on the corresponding stimulus indicators, and stimulus duration lines of each stimulus indicator. The stimulus duration lines include several high stimulus periods and several low stimulus periods. Based on the quantitative stimulus set and combined with the theoretically added photovoltaic installed capacity, a polynomial stimulus function corresponding to adjacent historical years is constructed. Calculate the optimal solution for each polynomial stimulation function, and construct several training samples based on the optimal solutions for all adjacent historical years and the regional photovoltaic development knowledge characteristics of each optimal solution. The neural network model is trained and validated based on the training samples to obtain a new prediction model, and the new photovoltaic installed capacity for each year is predicted according to the new prediction model.

[0008] Preferably, several training samples are constructed, including: Each optimal solution is labeled with a corresponding stimulus index label, and the regional photovoltaic development knowledge features associated with the optimal solution are extracted. The optimal solutions are sorted according to the hierarchical relationship of the stimulus indicator labels, and the temporal correlation weights between each node of stimulus indicator label-optimal solution-knowledge feature are calculated. The sorted optimal solution is associated with the corresponding regional photovoltaic development knowledge features based on the time-series association weight to form an initial feature link. The regional photovoltaic development trend fit is calculated to determine the fit value of each conflicting feature in the initial feature link. Features with fit values ​​higher than a preset threshold are retained to generate a feature link after conflict resolution. Based on the temporal association weights of nodes in the feature link after conflict resolution, a graph structure information containing node temporal attributes and feature confidence is constructed. Based on the relationships between nodes and feature confidence in the graph structure information, the optimal solution, stimulus index labels, and knowledge features are mapped to a two-dimensional matrix to obtain the basic matrix of optimal solution-feature-confidence. After normalizing the base matrix, and extracting the core feature dimensions based on principal component analysis and combining the temporal correlation weights corresponding to the core features, a feature space matrix containing core feature weights, cross-optimal solution correlation degree, and feature confidence is constructed. The initially constructed samples are input into the pre-trained model for trial prediction. Based on the trial prediction error, the core feature weights and feature confidence of the feature space matrix are adjusted in reverse. The temporal feature distribution of the samples is optimized by combining the temporal correlation weights of adjacent historical years. The initially constructed samples are then expanded to obtain several training samples.

[0009] Preferably, the initially constructed samples are expanded to obtain several training samples, including: The first adjustment vector is constructed by obtaining the weight adjustment amount of each initially constructed sample, and the second adjustment vector is constructed by obtaining the confidence adjustment amount of each initially constructed sample. The first adjustment vector and the second adjustment vector are aligned and divided into levels to obtain the adjustment level of each initially constructed sample, and a double matrix is ​​obtained, wherein each vector in the double matrix corresponds to an initially constructed sample; Perform two-level clustering analysis on the dual matrix to obtain several clustering results, and obtain the cluster center cluster of each clustering result. Expand the cluster center cluster according to the time-series feature distribution to obtain several double expansion levels, and obtain the new samples of each double expansion level. The newly added samples and the initially constructed samples are used as training samples.

[0010] Preferably, the newly added samples for each double expansion level include: The temporal feature distribution corresponding to the cluster center cluster is hierarchically decomposed to obtain temporal sub-distributions in the dimensions of year, season, and month, and the fluctuation correlation degree between each temporal sub-distribution is calculated; Based on the fluctuation correlation, a temporal expansion weight is assigned to each cluster center cluster, and the temporal constraint boundary of the dual expansion level is determined by combining the adjustment level corresponding to the dual matrix. According to the temporal expansion weight and temporal constraint boundary, the cluster center cluster is expanded bidirectionally in both the forward and reverse temporal directions to generate several double expansion levels. Calculate the temporal fit between the new samples and the original temporal feature distribution under each double expansion level, and filter the new samples whose temporal fit is higher than the preset fit.

[0011] Preferably, after generating the second photovoltaic installation addition table, it also includes: Based on the second photovoltaic installation addition table, the predicted photovoltaic installation inflow and predicted photovoltaic installation outflow for each of the n1 historical years are obtained, and compared with the actual photovoltaic installation inflow and actual photovoltaic installation outflow for the corresponding historical years to obtain the basic error vector for each historical year. At the same time, the actual operating condition data and standard operating condition data of each batch of PV modules put into production in the target area in the n1 historical years are collected, along with the extreme vector based on the extreme occurrence probability of each operating condition item. Each batch of production corresponds to one extreme vector. Obtain the reference error matrix of photovoltaic installation data for n1 historical years of a reference region that has the same operating conditions as the corresponding target region. At the same time, obtain the reference extreme probability of each operating condition item for n1 historical years of the reference region and construct the reference probability matrix of the corresponding reference region. Determine a first discrete relationship between each basic error vector and the reference error matrix, and simultaneously determine a second discrete relationship between each extreme vector and the reference probability matrix; Based on the first discrete relationship and the second discrete relationship, the operating error information of each production batch is fused and decomposed in multiple dimensions to obtain the error distribution characteristics, error contribution characteristics and nonlinear characteristics of operating parameters and failure probability of the target area based on different production batches. Based on the decomposition results Make corrections and re-obtain the updated second photovoltaic installation table.

[0012] Preferably, determining the second discrete relationship between each extreme vector and the reference probability matrix includes: Perform time-series alignment processing on each working condition item in the extreme vector and the working condition item corresponding to the reference probability matrix to obtain the time-series aligned extreme sub-vector and reference probability sub-matrix; For each time-aligned operating condition item, calculate the weighted dispersion of the extreme occurrence probability of the corresponding operating condition item in the extreme sub-vector and the reference extreme probability of the corresponding operating condition item in the reference probability sub-matrix; Calculate the correlation weight of each production batch in the target area with the batch of the same operating condition in the reference area. At the same time, calculate the correlation coefficient between the corresponding operating condition item and the failure probability of PV modules. Based on the weighted discreteness of each working condition item, the corresponding working condition correlation weight, and the correlation coefficient, the second discrete relationship between each extreme vector and the reference probability matrix is ​​obtained.

[0013] Preferably, the calculation of weighted dispersion includes: Calculate the absolute discrete values ​​of the extreme occurrence probability of the corresponding working condition item in the extreme sub-vector and the reference extreme probability of the corresponding working condition item in the reference probability sub-matrix; The weighted dispersion is obtained by multiplying the absolute discrete value with the corresponding extreme weight coefficient and the working condition area fit.

[0014] Compared with the prior art, the beneficial effects of this application are as follows: 1. By integrating historical and future data, a photovoltaic installation data system covering three dimensions—new installations, inflows, and outflows—is constructed. By combining the lifetime distribution function and cross-regional flow coupling coefficient, the outflow is accurately calculated, realizing the systematic calculation of photovoltaic power generation retirement paths. Compared with existing technologies, the data coverage is more comprehensive and the calculation logic is more realistic, providing basic data support for retirement path planning, significantly improving the reliability of calculation results, and ensuring prediction accuracy.

[0015] 2. A combined approach based on quantitative stimulus sets, multinomial functions, neural networks, sample refinement, and reference region correction yields the following results: When each module is improved individually, the prediction error is reduced by 1.2%-1.8% respectively, and the combined error is reduced by 4.3% (far exceeding the sum of the effects of individual improvements).

[0016] The combination of cross-regional flow coefficient and extreme operating condition duration reduced the outflow calculation error from 15.7% to 2.4%, solving the problem that existing technologies cannot quantify the combined effects of regional flow and environmental impact.

[0017] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating a method for constructing a photovoltaic power generation decommissioning path calculation model based on dynamic material flow, as described in an embodiment of the present invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] This invention provides a method for constructing a photovoltaic power generation decommissioning path calculation model based on dynamic material flow, such as... Figure 1 As shown, it includes: Step 1: Collect basic photovoltaic installation data for at least one target area corresponding to n1 historical years, and generate a first photovoltaic installation addition table, wherein the first photovoltaic installation addition table includes at least the actual newly added photovoltaic installation capacity and the actual photovoltaic installation inventory in use for each of the n1 historical years in the target area; Step 2: Based on the historical data of the first photovoltaic installation increase table, and combined with the preset conditions related to the development plan of the photovoltaic industry, capacity supply constraints, and energy demand trends of the target region, predict the new photovoltaic installation capacity of the target region in each of the n2 future years, and use it as the benchmark data for the change of photovoltaic installation inventory in future years; Step 3: Calculate the photovoltaic installed capacity outflow of the i-th target region in year t. ,in, Let this be the photovoltaic installation outflow in the i-th target region in year t. For the i-th target region in the th... Annual photovoltaic installation inflow; The lifetime distribution function represents the first... The cumulative extreme operating time of PV modules flowing in annually The failure probability in year t; For time interval parameters; This refers to the dynamic coupling coefficient for cross-regional circulation. Step 4: Calculate the total photovoltaic installation inflow for the i-th target region in year t. ,in, Let be the total inflow of photovoltaic installations in the i-th target region in year t; Let be the difference between the in-use photovoltaic installation inventory in year t and the in-use photovoltaic installation inventory in year t-1 of the i-th target region; Step 5: Integrate the actual photovoltaic installation data of the n1 historical years and the predicted and calculated data of the n2 future years to generate a second photovoltaic installation addition table covering the n1 historical years and the n2 future years. In the second photovoltaic installation addition table, each year records the new photovoltaic installation capacity, photovoltaic installation inflow, and photovoltaic installation outflow.

[0022] In this embodiment: Where T is the component lifespan under standard operating conditions; This is the standard operating condition duration threshold. For shape factor; ,in, For the first The cross-regional transfer weight of components flowing into the market each year, with positive values ​​indicating inflows and negative values ​​indicating outflows.

[0023] This includes collecting full lifecycle operating condition data for each batch of components put into production in the target area, and calculating the cumulative duration of extreme operating conditions. Extreme operating conditions are defined as temperatures > 45°C and irradiance > Or humidity > 90% for a continuous period > 24 hours; the actual lifespan of the components corresponding to different C values ​​was obtained through accelerated aging experiments, and the results were fitted to obtain... The lifespan attenuation factor under extreme conditions is 0.02 / 100 hours (standard operating conditions). The shape factor of the Weibull curve is based on accelerated aging test data from 1000 sets of photovoltaic modules from different brands (temperature 45-85℃, humidity 30%-95%, irradiance 800-1500 ppm). The value obtained by fitting the data using the maximum likelihood estimation method is 5.3759. It should be noted that... Adjust according to the scenario, that is Where k is the scene correction coefficient, and k=0.1 for high temperature and high humidity areas. In areas with normal temperature and humidity, k=0.

[0024] In this embodiment, The value ranges from -0.5 to 0.5, and is based on the statistical data of circulation in 31 provincial regions across the country from 2010 to 2022.

[0025] In this embodiment, based on the IEC61215 photovoltaic module standard test conditions and combined with 100,000 hours of outdoor operation data statistics, This refers to the cumulative operating time under normal conditions over 5 years.

[0026] In this embodiment, the target area refers to the specific geographical area where photovoltaic power generation decommissioning path calculations need to be performed, and is defined according to standards such as administrative divisions and energy planning areas.

[0027] n1 historical years refer to consecutive or discontinuous years in the past for which complete photovoltaic installation data has been accumulated. n1 is a positive integer greater than or equal to 10 to ensure that the data has statistical significance.

[0028] Basic data on photovoltaic installations refers to core data reflecting changes in the scale of photovoltaic installations in a target area, including the capacity of newly connected photovoltaic projects each year and the total capacity of photovoltaic modules still in operation at the end of the year.

[0029] The actual newly installed photovoltaic capacity refers to the total capacity of photovoltaic modules newly connected to the grid and put into operation in the target area within a certain year.

[0030] Actual installed photovoltaic capacity refers to the total capacity of photovoltaic modules that are still operating normally in the target area at the end of a certain year.

[0031] In this embodiment, n2 future years refer to consecutive future years in which the new photovoltaic installed capacity needs to be predicted, and n2 is a positive integer greater than or equal to 3.

[0032] Preset conditions refer to the set of key factors affecting future new photovoltaic (PV) installed capacity, including three core conditions: PV industry development plan, capacity supply constraints, and energy demand trends. For example, if the energy plan clearly states the target of 500GW of PV installed capacity nationwide by 2025, and Province A, as a major PV province, plans to account for 12% of that, then this plan is one of the preset conditions; if Province A's local PV module production capacity in 2023 is 15GW / year, and after deducting 3GW for external supply, the local supply capacity is 12GW, this is a capacity supply constraint; if Province A expects its total electricity consumption to increase by 6% in 2023, with the proportion of new energy power generation needing to increase to 20%, then the estimated new PV demand is 9-11GW, this is an energy demand trend condition.

[0033] In this embodiment, the benchmark data for future changes in photovoltaic installed capacity inventory refers to the new photovoltaic installed capacity in future years as the core reference for inventory changes, that is, the predicted new capacity is used as the basic input data for inventory changes.

[0034] In this embodiment, the second photovoltaic installation addition table refers to a structured table covering historical years and future years, recording three types of data: additions, inflows, and outflows, as shown in Table 1, including some years: Table 1. Second Table of New Photovoltaic Installations The beneficial effects of the above technical solution are as follows: by integrating historical and future data, a photovoltaic installation data system covering three dimensions of new installations, inflows, and outflows is constructed. By combining the lifetime distribution function and the cross-regional flow coupling coefficient, the outflow is accurately calculated, realizing the systematic calculation of the photovoltaic power generation retirement path. Compared with the existing technology, the data coverage is more comprehensive and the calculation logic is more in line with reality, providing basic data support for retirement path planning and significantly improving the reliability of the calculation results.

[0035] This invention provides a method for constructing a photovoltaic power generation decommissioning path calculation model based on dynamic material flow, which predicts the new photovoltaic installed capacity in the target region for each of the n² future years, including: By obtaining the actual newly installed photovoltaic capacity and the retired photovoltaic capacity in adjacent historical years, the theoretical newly installed photovoltaic capacity can be obtained. Information mining is performed on the target region in the adjacent historical years based on photovoltaic industry development plans, capacity supply constraints, and energy demand trends to construct a quantitative stimulus set. The quantitative stimulus set includes several stimulus indicators, stimulus constraint terms of other indicators on the corresponding stimulus indicators, and stimulus duration lines of each stimulus indicator. The stimulus duration lines include several high stimulus periods and several low stimulus periods. Based on the quantitative stimulus set and combined with the theoretically added photovoltaic installed capacity, a polynomial stimulus function corresponding to adjacent historical years is constructed. Calculate the optimal solution for each polynomial stimulation function, and construct several training samples based on the optimal solutions for all adjacent historical years and the regional photovoltaic development knowledge characteristics of each optimal solution. The neural network model is trained and validated based on the training samples to obtain a new prediction model, and the new photovoltaic installed capacity for each year is predicted according to the new prediction model.

[0036] In this embodiment, adjacent historical years refer to two consecutive historical years. A combination of consecutive years is selected from n1 historical years, such as 2020 and 2021.

[0037] In this embodiment, the retired photovoltaic installed capacity refers to the total capacity of photovoltaic modules that are completely withdrawn from operation in a target area within a certain year, which is statistically analyzed through retirement filing data from the energy authority and power station dismantling records.

[0038] The theoretical new photovoltaic installed capacity refers to the new capacity required to maintain the growth of existing inventory. The theoretical new capacity is calculated as the actual new capacity plus the decommissioned capacity.

[0039] In this embodiment, information mining refers to extracting key information affecting new installations from materials such as policy documents, industry reports, and statistical data, and obtaining it by combining text mining algorithms with manual screening.

[0040] In this embodiment, the quantitative stimulus set refers to a set that includes stimuli that affect new installations, as well as related constraints and duration characteristics. The implementation method is to structure and organize the mined information to form the set.

[0041] Stimulus indicators refer to quantitative factors that have a direct impact on new installed capacity. They are determined based on information mining results. For example, stimulus indicators include subsidy intensity (yuan / kWh), capacity utilization rate (%), and power gap (GW).

[0042] Stimulus constraints refer to the restrictions imposed by other indicators on a certain stimulus indicator. Specifically, they are determined by analyzing the interaction between indicators. For example, the stimulus constraint for subsidy intensity is the upper limit of the fiscal budget (in hundreds of millions of yuan), meaning that the subsidy intensity cannot cause the total subsidy amount to exceed the fiscal budget.

[0043] Incentive constraints: Subsidy intensity constraints: ,in, For the subsidy intensity, B represents the annual fiscal budget for photovoltaic subsidies; E represents the projected annual photovoltaic power generation.

[0044] Capacity utilization constraints: ,in, For capacity utilization rate; Minimum effective capacity; For total capacity, and .

[0045] The duration of stimulus measures refers to the time distribution characteristics of the effects of stimulus indicators, which are determined based on the policy implementation cycle and industry cycle. High-stimulation period: The stimulation index value is ≥75th percentile and the duration is ≥6 months.

[0046] Low-stimulation period: Stimulation index value ≤ 25th percentile, duration ≥ 6 months.

[0047] For example, the statistical quantiles of subsidy intensity from 2010 to 2020 (25% = 0.03 yuan / kWh, 75% = 0.08 yuan / kWh) indicate that the period from 2010 to 2018 (subsidy ≥ 0.08 yuan / kWh, lasting 9 years) was a period of high stimulus, while the period from 2019 to 2022 (subsidy ≤ 0.03 yuan / kWh, lasting 4 years) was a period of low stimulus.

[0048] In this embodiment, the polynomial stimulus function refers to a polynomial function that describes the quantitative relationship between stimulus indicators and theoretical new installed capacity. It is derived by fitting the indicators from the quantitative stimulus set with the theoretical new installed capacity using the least squares method. For example, using stimulus indicators x1 (subsidy intensity), x2 (capacity utilization rate), and x3 (electricity gap) as independent variables, and theoretical new installed capacity y as the dependent variable, a bivariate cubic polynomial stimulus function is constructed: , where a0 to a9 are coefficients to be determined.

[0049] In this embodiment, the optimal solution refers to the polynomial stimulus function that minimizes the deviation between the theoretical and actual increase within the stimulus constraints. It is obtained using an optimization algorithm (such as gradient descent). For example, for the polynomial stimulus function in 2021, the optimal solution is (x1 = 0.07 yuan / degree, x2 = 85%, x3 = 5.2GW).

[0050] In this embodiment, the regional photovoltaic development knowledge characteristics refer to the unique attributes of the photovoltaic industry development in the target region. For example, the regional photovoltaic development knowledge characteristics of Jiangsu Province include having the largest module manufacturing capacity in the country, a high proportion of distributed photovoltaics, and strong absorption capacity.

[0051] Training samples refer to the sample data used to train the neural network model, which is formed by associating the optimal solution, stimulus indicator labels, and regional photovoltaic development knowledge characteristics. For example, the training sample format is (optimal solution: (0.07, 85%, 5.2), stimulus indicator labels: subsidy-capacity-gap, knowledge characteristics: leading capacity + high proportion of distributed generation).

[0052] In this embodiment, the neural network model refers to a machine learning model used to predict the new photovoltaic installed capacity, and model architectures such as BP neural network and LSTM are selected. For example, a 3-layer BP neural network is selected, with the number of neurons in the input layer being the number of stimulus indicators (3), the number of neurons in the hidden layer being 10, and the number of neurons in the output layer being 1 (the predicted value of the new capacity).

[0053] Training and validation refer to optimizing model parameters and verifying accuracy using training samples. This involves dividing the training samples into a training set and a validation set in a 7:3 ratio, using mean squared error (MSE) as the loss function. For example, the training set is used to adjust the neural network weights, and the validation set is used to test the model's prediction error. Model training is considered complete when the validation set MSE < 0.01.

[0054] The newly added prediction model refers to the trained and validated neural network model, whose parameters are saved for future predictions. For example, a trained BP neural network model can take the predicted values ​​of stimulus indicators for future years as input and output the corresponding new photovoltaic installed capacity.

[0055] The beneficial effects of the above technical solution are: by constructing a quantitative stimulus set to quantify abstract influencing factors, and combining a multinomial stimulus function with a neural network model, accurate prediction of future new photovoltaic installed capacity is achieved. Compared with traditional prediction methods, it fully considers the dynamic stimulus characteristics of influencing factors and regional development differences, significantly improves prediction accuracy, and provides reliable benchmark data for subsequent decommissioning path calculation.

[0056] This invention provides a method for constructing a photovoltaic power generation decommissioning path calculation model based on dynamic material flow, which constructs several training samples, including: Each optimal solution is labeled with a corresponding stimulus index label, and the regional photovoltaic development knowledge features associated with the optimal solution are extracted. The optimal solutions are sorted according to the hierarchical relationship of the stimulus indicator labels, and the temporal correlation weights between each node of stimulus indicator label-optimal solution-knowledge feature are calculated. The sorted optimal solution is associated with the corresponding regional photovoltaic development knowledge features based on the time-series association weight to form an initial feature link. The regional photovoltaic development trend fit is calculated to determine the fit value of each conflicting feature in the initial feature link. Features with fit values ​​higher than a preset threshold are retained to generate a feature link after conflict resolution. Based on the temporal association weights of nodes in the feature link after conflict resolution, a graph structure information containing node temporal attributes and feature confidence is constructed. Based on the relationships between nodes and feature confidence in the graph structure information, the optimal solution, stimulus index labels, and knowledge features are mapped to a two-dimensional matrix to obtain the basic matrix of optimal solution-feature-confidence. After normalizing the base matrix, and extracting the core feature dimensions based on principal component analysis and combining the temporal correlation weights corresponding to the core features, a feature space matrix containing core feature weights, cross-optimal solution correlation degree, and feature confidence is constructed. The initially constructed samples are input into the pre-trained model for trial prediction. Based on the trial prediction error, the core feature weights and feature confidence of the feature space matrix are adjusted in reverse. The temporal feature distribution of the samples is optimized by combining the temporal correlation weights of adjacent historical years. The initially constructed samples are then expanded to obtain several training samples.

[0057] In this embodiment, the stimulus indicator label refers to the classification identifier assigned to the stimulus indicator corresponding to the optimal solution. For example, the stimulus indicator subsidy intensity corresponds to the label policy category - subsidy, the capacity utilization rate corresponds to the label industry category - capacity, and the electricity gap corresponds to the label demand category - gap.

[0058] In this embodiment, the hierarchical relationship of the stimulus indicator labels refers to the classification hierarchy between labels (such as first-level labels: policy, industry, demand; second-level labels: subsidies, capacity, gap), with the hierarchy being policy > subsidies > industry > capacity and demand > gap.

[0059] Sorting refers to classifying and sorting the optimal solutions according to hierarchical relationships. This is achieved using a hierarchical clustering algorithm. For example, all optimal solutions are divided into three categories according to the first-level label, and then sorted according to the second-level label within each category.

[0060] Temporal association weight refers to the strength of association between different nodes (stimulus label - optimal solution - knowledge feature) over time, calculated using the mutual information method. ,in, For joint probability density, This represents the marginal probability density.

[0061] Normalized weights: ,in, For all The maximum value in.

[0062] In this embodiment, the initial feature chain refers to the chain formed by connecting the sorted optimal solution and knowledge features according to the temporal association weights. The relevant nodes are connected using a graph structure. For example, the initial feature chain is: policy category - subsidy (tag) → optimal solution (0.07 yuan / degree) → knowledge feature (subsidy policy continues).

[0063] The regional photovoltaic development trend adaptability is obtained based on a pre-constructed regional photovoltaic development regional adaptability model, and the model structure is as follows: Input parameters: annual growth rate of newly installed photovoltaic capacity, adjustment range of policy subsidies, capacity expansion rate, and change rate of power shortage (4 dimensions).

[0064] Output parameter: Feature fit (0-1).

[0065] Model formula: Fit = ,in: Growth rate fit is the growth rate corresponding to the feature, and rtrend is the average growth rate of the region.

[0066] Policy adaptability and the trend of subsidy reduction Where x1base is the baseline subsidy intensity, and x1 is the current subsidy intensity.

[0067] Production capacity suitability =min(x2 / x2target,1), where x2target is the target capacity utilization rate and x2 is the current capacity utilization rate; : Gap fit is the gap value corresponding to the feature, and gavg is the average gap value.

[0068] In this embodiment, conflict features refer to features in the initial feature chain whose adaptability is lower than a preset threshold. The preset threshold is 0.5. The adaptability value of 0.3 for a subsidy intensity of 0.08 yuan / degree is <0.5, which is a conflict feature.

[0069] The feature chain after conflict resolution refers to the chain formed after retaining features with adaptation values ​​higher than a preset threshold. For example, after deleting the node with a subsidy intensity of 0.08 yuan / degree, the feature chain is: policy category - subsidy → optimal solution (0.03 yuan / degree) → knowledge feature (subsidy reduction).

[0070] Graph structure information refers to a graph data structure that includes the temporal attributes and feature confidence of nodes, and uses a graph database (such as Neo4j) to store nodes and their relationships.

[0071] The node time sequence attribute refers to the time characteristics corresponding to the node, that is, the year corresponding to the optimal solution. For example, the node time sequence attribute of the optimal solution (0.03 yuan / degree) is 2022.

[0072] Feature confidence refers to the reliability of a feature, calculated based on the credibility of the data source and statistical significance, with a value ranging from 0 to 1. In this embodiment, the base matrix refers to mapping the optimal solution, stimulus indicator labels, and knowledge features into a two-dimensional matrix. For example, the optimal solution is quantified into numerical values, the labels are quantified into codes (policy category = 1, industry category = 2, demand category = 3), and the knowledge features are quantified into confidence scores, constructing a 3×N base matrix (N is the number of samples), as shown in Table 2: Table 2 Basic Matrix In this embodiment, normalization refers to mapping the basic matrix data to the [0,1] interval and using the min-max normalization algorithm.

[0073] Principal component analysis (PCA) is a dimensionality reduction method that extracts core feature dimensions. It uses the PCA algorithm to retain principal components with a cumulative variance contribution rate ≥ 85%. For example, performing PCA analysis on a 3-dimensional fundamental matrix would retain two principal components (with a cumulative variance contribution rate of 88%).

[0074] The core feature dimension, determined by the PCA algorithm, refers to the feature dimension that has the greatest impact on the prediction results. For example, the core feature dimensions could be the normalized value of subsidy intensity or the normalized value of the power gap.

[0075] The feature space matrix is ​​a matrix that contains the weights of core features, the correlation degree across optimal solutions, and the feature confidence. For example, the feature space matrix is ​​a 3×2 matrix, where the rows represent the weights of core features, the correlation degree across optimal solutions, and the feature confidence, and the columns represent the two core feature dimensions.

[0076] In this embodiment, the pre-trained model refers to the initial model used for trial prediction. A linear regression model is selected as the pre-trained model, the basic matrix data is input, and the trial prediction value is output.

[0077] Trial prediction error refers to the deviation between the predicted value and the actual value of the pre-trained model. The trial prediction error MSE = 0.03.

[0078] Reverse adjustment refers to correcting the parameters of the feature space matrix based on the trial prediction error, using the gradient descent method to adjust the weights and confidence levels of the core features. For example, if the prediction error of the power shortage feature is large, its core feature weight is adjusted from 0.4 to 0.6.

[0079] Temporal feature distribution optimization refers to adjusting the temporal distribution of samples to better reflect actual trends, adjusting the sample size based on the temporal correlation weights of historical years. For example, if the temporal correlation weights for 2020-2022 are 0.7, 0.8, and 0.9, the sample size for each year should be adjusted accordingly.

[0080] Augmentation refers to increasing the number of samples to improve the model's generalization ability, such as expanding the initial 100 samples to 300 samples.

[0081] The beneficial effects of the above technical solution are as follows: through a series of refined processes such as label annotation, conflict resolution, matrix construction and sample expansion, high-quality training samples are constructed. Compared with conventional sample construction methods, it fully considers the temporal correlation characteristics, feature confidence and conflict resolution, significantly improves the adaptability and reliability of training samples, and provides a solid foundation for high-precision training of new prediction models.

[0082] This invention provides a method for constructing a photovoltaic power generation decommissioning path calculation model based on dynamic material flow. The method involves expanding the initially constructed sample to obtain several training samples, including: The first adjustment vector is constructed by obtaining the weight adjustment amount of each initially constructed sample, and the second adjustment vector is constructed by obtaining the confidence adjustment amount of each initially constructed sample. The first adjustment vector and the second adjustment vector are aligned and divided into levels to obtain the adjustment level of each initially constructed sample, and a double matrix is ​​obtained, wherein each vector in the double matrix corresponds to an initially constructed sample; Perform two-level clustering analysis on the dual matrix to obtain several clustering results, and obtain the cluster center cluster of each clustering result. Expand the cluster center cluster according to the time-series feature distribution to obtain several double expansion levels, and obtain the new samples of each double expansion level. The newly added samples and the initially constructed samples are used as training samples.

[0083] In this embodiment, the weight adjustment amount refers to the adjustment range of the weights of the core features in the feature space matrix, that is, the difference between the weights before and after the adjustment.

[0084] The first adjustment vector is a vector composed of the weight adjustments of all core features, that is, the weight adjustments are arranged in the order of the core features. For example, if the core feature is the subsidy intensity power gap, and the weight adjustments are 0.1 and 0.05 respectively, then the first adjustment vector is [0.1, 0.05].

[0085] The confidence level adjustment refers to the adjustment range of the feature confidence level, that is, the difference between the confidence levels before and after the adjustment. For example, if the confidence level of the subsidy intensity feature is adjusted from 0.9 to 0.95, the confidence level adjustment is 0.05.

[0086] The second adjustment vector is a vector composed of the confidence adjustment values ​​of all features, that is, the confidence adjustment values ​​are arranged in the order of features. For example, if the confidence adjustment values ​​of the feature subsidy intensity power gap are 0.05 and 0.03 respectively, then the second adjustment vector is [0.05, 0.03].

[0087] In this embodiment, vector elements are arranged in the same characteristic order to achieve position alignment, such as the first adjustment vector [0.1 (subsidy), 0.05 (gap)] being aligned with the second adjustment vector [0.05 (subsidy), 0.03 (gap)].

[0088] Level classification processing refers to dividing the adjustment amount into different levels (such as low, medium, and high). For example, the threshold for weight adjustment amount level is: low (0-0.05), medium (0.05-0.1), and high (>0.1); the threshold for confidence adjustment amount level is: low (0-0.02), medium (0.02-0.05), and high (>0.05).

[0089] The adjustment level refers to the combination of the weight adjustment level and the confidence adjustment level for each sample. For example, if the first adjustment vector elements of a sample are 0.1 (medium) and 0.05 (medium), and the second adjustment vector elements are 0.05 (medium) and 0.03 (medium), then the adjustment level is (medium, medium).

[0090] The bimatrix refers to the matrix composed of the first adjustment vector and the second adjustment vector of all samples, as shown in Table 3: Table 3 Two matrices In this embodiment, the two-level clustering analysis refers to using the K-means clustering algorithm and clustering the two matrices based on the adjusted levels. The elbow rule is used to calculate the sum of squared errors (SSE) within clusters for different K values ​​(1-10), and the K value corresponding to the point where the SSE decrease rate abruptly changes is taken. Since the SSE of a certain sample set abruptly changes when K=2, K=2 is chosen.

[0091] In this embodiment, the clustering result refers to the sample cluster formed after clustering. For example, clustering result 1 contains 2 samples and clustering result 2 contains 1 sample.

[0092] The cluster center cluster refers to the center vector of each cluster result. For example, if the average value of the first adjustment vector of cluster result 1 is [0.09, 0.055] and the average value of the second adjustment vector is [0.045, 0.025], then the cluster center cluster is ([0.09, 0.055], [0.045, 0.025]).

[0093] In this embodiment, temporal feature distribution expansion refers to expanding the cluster center cluster according to the historical temporal feature distribution, generating new vectors according to the year, quarter, and month temporal distribution. For example, according to the temporal distribution from 2020 to 2030, 10 temporal expansion vectors are generated for the cluster center cluster.

[0094] In this embodiment, the dual expansion level refers to the combination of the expanded adjustment levels, and the implementation method is to determine the level based on the adjustment amount of the expansion vector. For example, if the weight adjustment amount of the expansion vector is 0.09 (medium) and the confidence adjustment amount is 0.045 (medium), then the dual expansion level is (medium, medium).

[0095] New samples refer to new samples generated based on dual expansion levels, such as generating new samples based on expansion vectors [0.09, 0.055] and [0.045, 0.025], associating feature labels and knowledge features.

[0096] The beneficial effects of the above technical solution are: by constructing a dual matrix and adopting dual-level clustering and temporal expansion, the training samples are accurately expanded. Compared with the traditional random sampling expansion method, it fully considers the level characteristics and temporal distribution of the adjustment amount, ensures the consistency and adaptability of the new samples and the original samples, effectively improves the scale and quality of the sample set, and enhances the generalization ability of the model.

[0097] This invention provides a method for constructing a photovoltaic power generation decommissioning path calculation model based on dynamic material flow, obtaining new samples for each double-expansion level, including: The temporal feature distribution corresponding to the cluster center cluster is hierarchically decomposed to obtain temporal sub-distributions in the dimensions of year, season, and month, and the fluctuation correlation degree between each temporal sub-distribution is calculated; Based on the fluctuation correlation, a temporal expansion weight is assigned to each cluster center cluster, and the temporal constraint boundary of the dual expansion level is determined by combining the adjustment level corresponding to the dual matrix. According to the temporal expansion weight and temporal constraint boundary, the cluster center cluster is expanded bidirectionally in both the forward and reverse temporal directions to generate several double expansion levels. Calculate the temporal fit between the new samples and the original temporal feature distribution under each double expansion level, and filter the new samples whose temporal fit is higher than the preset fit.

[0098] In this embodiment, the temporal feature distribution refers to the distribution pattern of samples in the time dimension (such as annual distribution, quarterly distribution). For example, the temporal feature distribution of the original samples is 30 samples each year from 2020 to 2022.

[0099] Hierarchical decomposition refers to splitting the temporal feature distribution into sub-distributions of different dimensions such as year, quarter, and month. For example, the 30 samples in 2022 can be split into 4 quarters (7-8 samples per quarter) and 12 months (2-3 samples per month) to obtain temporal sub-distributions in the dimensions of year, quarter, and month.

[0100] A time series sub-distribution refers to the distribution of samples at a certain time level. For example, the time series sub-distribution for the second quarter of 2022 has 8 samples, and the time series sub-distribution for June 2022 has 3 samples.

[0101] The volatility correlation degree refers to the degree of consistency of volatility among different time series sub-distributions. It is calculated using the Pearson correlation coefficient and ranges from -1 to 1.

[0102] In this embodiment, the temporal expansion weight refers to the expansion strength assigned to each cluster center cluster, which is determined according to the volatility correlation (the higher the correlation, the greater the weight). For example, the volatility correlation corresponding to cluster center cluster A is 0.9, and the temporal expansion weight is 0.6; the volatility correlation corresponding to cluster center cluster B is 0.7, and the temporal expansion weight is 0.4.

[0103] The time series constraint boundary refers to the time range limit of the time series extension, which is determined according to the range of future years (n² future years). For example, if the future years are 2023-2030, then the time series constraint boundary is January 2023 to December 2030.

[0104] In this embodiment, forward temporal expansion refers to expanding the sample from historical years to future years, generating samples of future time nodes in chronological order, such as expanding forward from December 2022 to generate samples of time nodes such as January and February 2023.

[0105] Reverse time extension refers to extending the sample from historical years to earlier years (supplementing earlier data), generating samples of past time nodes in reverse time order, such as extending from January 2020 to generate samples of time nodes such as December and November 2019.

[0106] Dual expansion levels refer to sample sets at different time levels formed by forward expansion and reverse expansion. For example, dual expansion levels include quarterly samples in 2019 and monthly samples in 2023.

[0107] In this embodiment, temporal fit refers to the degree of fit between the newly added samples and the original temporal feature distribution, and temporal fit = ,in, For newly added sample time series values; is the original sample time series value; n is the number of time nodes. It should be noted that, based on the verification of 100 groups of expanded samples, the sample time series consistency meets the requirements when the fitness is ≥0.7. At this time, the newly added samples with a time series fitness higher than the preset fitness are retained.

[0108] The beneficial effects of the above technical solution are: through temporal hierarchical decomposition, bidirectional expansion and adaptability screening, the temporal expansion of samples is accurately achieved. Compared with the single-direction expansion method, it fully considers the fluctuation correlation and adaptability of different time dimensions, ensures the temporal rationality of the newly added samples, further improves the quality of the training sample set, and provides more comprehensive temporal feature support for model training.

[0109] This invention provides a method for constructing a photovoltaic power generation decommissioning path calculation model based on dynamic material flow. After generating the second photovoltaic installed capacity addition table, it also includes: Based on the second photovoltaic installation addition table, the predicted photovoltaic installation inflow and predicted photovoltaic installation outflow for each of the n1 historical years are obtained, and compared with the actual photovoltaic installation inflow and actual photovoltaic installation outflow for the corresponding historical years to obtain the basic error vector for each historical year. At the same time, the actual operating condition data and standard operating condition data of each batch of PV modules put into production in the target area in the n1 historical years are collected, along with the extreme vector based on the extreme occurrence probability of each operating condition item. Each batch of production corresponds to one extreme vector. Obtain the reference error matrix of photovoltaic installation data for n1 historical years of a reference region that has the same operating conditions as the corresponding target region. At the same time, obtain the reference extreme probability of each operating condition item for n1 historical years of the reference region and construct the reference probability matrix of the corresponding reference region. Determine a first discrete relationship between each basic error vector and the reference error matrix, and simultaneously determine a second discrete relationship between each extreme vector and the reference probability matrix; Based on the first discrete relationship and the second discrete relationship, the operating error information of each production batch is fused and decomposed in multiple dimensions to obtain the error distribution characteristics, error contribution characteristics and nonlinear characteristics of operating parameters and failure probability of the target area based on different production batches. Based on the decomposition results Make corrections and re-obtain the updated second photovoltaic installation table.

[0110] In this embodiment, the predicted photovoltaic installation inflow and predicted photovoltaic installation outflow refer to the predicted values ​​of the inflow and outflow for historical years calculated in steps 3-4, that is, the historical year data is substituted into the formula for calculation.

[0111] In this embodiment, the basic error vector refers to the vector composed of the error between the predicted value and the actual value for each historical year. That is, the predicted value minus the actual value is calculated and arranged by year. For example, the inflow errors for 2020-2022 are 0.05GW, 0.03GW, and 0.05GW, respectively, and the outflow errors are 0.0GW, 0.0GW, and 0.05GW, respectively. Then the basic error vector is [(0.05,0.0),(0.03,0.0),(0.05,0.05)].

[0112] A batch of PV modules put into production refers to a collection of photovoltaic modules that were put into production at the same time and have the same technical parameters. They are divided into batches according to the production time and module model.

[0113] Actual operating data refers to the actual environment and operating parameters during the operation of PV modules, such as the actual temperature, humidity, and irradiance data of Batch 1 modules from 2020 to 2022.

[0114] Standard operating condition data refers to the set rated operating parameters of the component, which are determined with reference to the component's technical specifications, such as standard temperature 25℃, standard humidity 60%, and standard irradiance. .

[0115] Operating error information refers to the deviation between actual operating condition data and standard operating condition data, i.e., the difference between the actual value and the standard value. For example, the average temperature error of component 1 in batch 1 is 5℃, and the average irradiance error is 1℃. .

[0116] The probability of an extreme condition occurring refers to the probability of an extreme condition occurring under a certain operating condition, which is the proportion of the duration of extreme operating conditions to the total operating time.

[0117] The extreme vector refers to the vector composed of the extreme occurrence probabilities of each operating condition item in each production batch, that is, the extreme occurrence probabilities are arranged in the order of the operating conditions items. For example, the operating conditions items of batch 1 are temperature, humidity and irradiance, and the extreme occurrence probabilities are 0.03, 0.01 and 0.02 respectively, then the extreme vector is [0.03, 0.01, 0.02].

[0118] Criteria for selecting reference areas: Operating condition similarity: The annual average difference in temperature, humidity, and radiation intensity is ≤10% (based on meteorological data statistics from the past 10 years).

[0119] Industry scale: The difference in photovoltaic installed capacity is ≤20%, and the similarity in production capacity structure is ≥80%.

[0120] The photovoltaic installation data for the reference region refers to the data on new installations, inflows, outflows, and inventory for n1 historical years in the reference region, which are obtained from the energy authorities of the reference region.

[0121] The reference error matrix is ​​a matrix composed of the basic error vectors for each historical year in the reference area, arranged by year and error type. As shown in Table 4, the reference error matrix is: Table 4 Reference Error Matrix The reference probability matrix refers to the matrix composed of the reference extreme probabilities of each operating condition item in each historical year of the reference area, arranged by year and operating condition item, as shown in Table 5: Table 5 Reference Probability Matrix In this embodiment, multi-dimensional fusion decomposition refers to decomposing the operating error information according to dimensions such as error distribution, error contribution, and nonlinear characteristics, and using factor analysis algorithm. For example, the operating error information of batch 1 is decomposed into error distribution characteristics (normal distribution, mean 0.02), error contribution characteristics (temperature error contributes 60%), and nonlinear characteristics (irradiation intensity error and failure probability are exponentially related).

[0122] Error distribution characteristics refer to the statistical distribution law of errors (such as normal distribution, uniform distribution), which is determined by fitting a probability density function. For example, the operating error of a batch of components follows a normal distribution with a mean of 0.02 and a variance of 0.001.

[0123] Error contribution characteristics refer to the proportion of each operating condition error to the total error, which is calculated using the variance decomposition method. For example, the contribution proportion of temperature error is 60%, humidity error is 20%, and irradiance intensity error is 20%.

[0124] The nonlinear characteristics of operating parameters and failure probability refer to the nonlinear correlation between operating parameter errors and component failure probability, which is determined through nonlinear regression fitting. For example, for every 1 increase in irradiance intensity error... The failure probability increases exponentially.

[0125] In this embodiment, the coefficients in the error characteristic adjustment formula, such as the lifetime decay coefficient, are adjusted from 0.02 / 100 hours to 0.022 / 100 hours based on the error decomposition results.

[0126] The updated second photovoltaic installation addition table refers to a revised and recalculated table of photovoltaic installation data covering both historical and future years.

[0127] In this embodiment, the data before correction (2022) is as follows: New photovoltaic installed capacity: 10.3GW (actual value); Predicted inflow: 0.45GW (calculated based on the original model); Actual inflow: 0.4GW (historical statistical value); Predicted outflow: 0.06GW (calculated based on the original model); Actual outflow: 0.05GW (historical statistical value).

[0128] Corrected inflow: 0.41GW (recalculated based on the corrected model), corrected outflow: 0.052GW, and the updated second photovoltaic installation additions are shown in Table 6: Table 6: Partially Updated Table The beneficial effects of the above technical solution are: by introducing reference area data to construct a multi-dimensional error analysis system, the key parameters of the model are accurately corrected. Compared with the correction method that only relies on the data of the target area itself, it makes full use of the reference value of similar working conditions, decomposes error characteristics in multiple dimensions, significantly improves the accuracy of model calculation, and makes the decommissioning path calculation results more in line with the actual situation.

[0129] This invention provides a method for constructing a photovoltaic power generation decommissioning path calculation model based on dynamic material flow, determining a second discrete relationship between each extreme vector and the reference probability matrix, including: Perform time-series alignment processing on each working condition item in the extreme vector and the working condition item corresponding to the reference probability matrix to obtain the time-series aligned extreme sub-vector and reference probability sub-matrix; For each time-aligned operating condition item, calculate the weighted dispersion of the extreme occurrence probability of the corresponding operating condition item in the extreme sub-vector and the reference extreme probability of the corresponding operating condition item in the reference probability sub-matrix; Calculate the correlation weight of each production batch in the target area with the batch of the same operating condition in the reference area. At the same time, calculate the correlation coefficient between the corresponding operating condition item and the failure probability of PV modules. Based on the weighted discreteness of each working condition item, the corresponding working condition correlation weight, and the correlation coefficient, the second discrete relationship between each extreme vector and the reference probability matrix is ​​obtained.

[0130] Preferably, the calculation of weighted dispersion includes: Calculate the absolute discrete values ​​of the extreme occurrence probability of the corresponding working condition item in the extreme sub-vector and the reference extreme probability of the corresponding working condition item in the reference probability sub-matrix; The weighted dispersion is obtained by multiplying the absolute discrete value with the corresponding extreme weight coefficient and the working condition area fit.

[0131] In this embodiment, time alignment means that the time nodes of the extreme vectors of the target region correspond one-to-one with the time nodes of the reference probability matrix of the reference region, and the time dimension is adjusted by year, quarter, or month. For example, the extreme vector of batch 1 of the target region corresponds to the cumulative value from 2020 to 2022, and the reference probability matrix of the reference region is recalculated according to the cumulative value from 2020 to 2022 to achieve time alignment.

[0132] The extreme sub-vector refers to the extreme vector of a certain time unit in the target region after time alignment. The extreme vector is split according to the aligned time unit. For example, if the time alignment is by year, the extreme sub-vector of batch 12020 is [0.01, 0.005, 0.008].

[0133] The reference probability submatrix refers to the reference probability matrix of a certain time unit in the reference region after time alignment. The reference probability matrix is ​​split according to the aligned time unit. For example, the reference probability submatrix of the reference region in 2020 is [0.025, 0.012, 0.018].

[0134] In this embodiment, the working condition correlation weight refers to the correlation strength between the target area and the reference area of ​​the same type of working condition batches. It is calculated based on the similarity of working condition data and has a value range of 0-1.

[0135] The correlation coefficient refers to the strength of the correlation between a certain operating condition and the failure probability of a PV module. It is calculated using the Pearson correlation coefficient and ranges from -1 to 1.

[0136] The second discrete relation refers to the total degree of difference after comprehensively considering the weighted dispersion, the working condition correlation weight, and the correlation coefficient. The second discrete relation is calculated as Σ(weighted dispersion × working condition correlation weight × correlation coefficient) / Σ(working condition correlation weight × correlation coefficient). The principle of the first discrete relation is similar to that of the second discrete relation, and will not be elaborated here.

[0137] First discrete relation = .

[0138] In this embodiment, the absolute discrete value = |extreme occurrence probability - reference extreme probability|.

[0139] In this embodiment, the extreme weight coefficient refers to the weight of the impact of an extreme operating condition on the component lifespan, which is determined based on accelerated aging experiments. For example, the extreme temperature condition has the greatest impact on lifespan, with a weight coefficient of 0.5; the irradiance intensity is the second most significant at 0.3; and the humidity is the least significant at 0.2.

[0140] In this embodiment, the working condition region adaptability refers to the degree of adaptability of a certain working condition item in the target region and the reference region. It is calculated based on the similarity of working condition data and has a value range of 0-1. Where m is the number of operating conditions. For the target area operating parameters, For reference area operating parameters, areas with a fit degree ≥ 0.7 are considered to be of the same type of operating condition area.

[0141] Weighted dispersion = absolute dispersion value × extreme weight coefficient × working condition area fit.

[0142] The beneficial effects of the above technical solution are as follows: by time alignment, weighted discreteness calculation and multi-dimensional correlation parameter fusion, the second discrete relationship between the extreme vector and the reference probability matrix is ​​accurately determined. Compared with the simple discreteness calculation method, it fully considers the consistency of time sequence, the correlation strength of working conditions and the degree of correlation with failure probability, which significantly improves the accuracy of discrete relationship calculation and provides a reliable basis for subsequent error decomposition and model correction.

[0143] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for constructing a photovoltaic power generation decommissioning path calculation model based on dynamic material flow, characterized in that, include: Step 1: Collect basic photovoltaic installation data for at least one target area corresponding to n1 historical years, and generate a first photovoltaic installation addition table, wherein the first photovoltaic installation addition table includes at least the actual newly added photovoltaic installation capacity and the actual photovoltaic installation inventory in use for each of the n1 historical years in the target area; Step 2: Based on the historical data of the first photovoltaic installation increase table, and combined with the preset conditions related to the development plan of the photovoltaic industry, capacity supply constraints, and energy demand trends of the target region, predict the new photovoltaic installation capacity of the target region in each of the n2 future years, and use it as the benchmark data for the change of photovoltaic installation inventory in future years; Step 3: Calculate the photovoltaic installed capacity outflow of the i-th target region in year t. ,in, Let this be the photovoltaic installation outflow in the i-th target region in year t. For the i-th target region in the th... Annual photovoltaic installation inflow; The lifetime distribution function represents the first... The cumulative extreme operating time of PV modules flowing in annually The failure probability in year t; For time interval parameters; This refers to the dynamic coupling coefficient for cross-regional circulation. Step 4: Calculate the total photovoltaic installation inflow for the i-th target region in year t. ,in, Let be the total inflow of photovoltaic installations in the i-th target region in year t; Let be the difference between the in-use photovoltaic installation inventory in year t and the in-use photovoltaic installation inventory in year t-1 of the i-th target region; Step 5: Integrate the actual photovoltaic installation data of the n1 historical years and the predicted and calculated data of the n2 future years to generate a second photovoltaic installation addition table covering the n1 historical years and the n2 future years. In the second photovoltaic installation addition table, each year records the new photovoltaic installation capacity, photovoltaic installation inflow, and photovoltaic installation outflow.

2. The method for constructing a photovoltaic power generation decommissioning path calculation model based on dynamic material flow according to claim 1, characterized in that, Before predicting the new photovoltaic installed capacity of the target region in each of the n² future years, the following are included: By obtaining the actual newly installed photovoltaic capacity and the retired photovoltaic capacity in adjacent historical years, the theoretical newly installed photovoltaic capacity can be obtained. Information mining is performed on the target region in the adjacent historical years based on photovoltaic industry development plans, capacity supply constraints, and energy demand trends to construct a quantitative stimulus set. The quantitative stimulus set includes several stimulus indicators, stimulus constraint terms of other indicators on the corresponding stimulus indicators, and stimulus duration lines of each stimulus indicator. The stimulus duration lines include several high stimulus periods and several low stimulus periods. Based on the quantitative stimulus set and combined with the theoretically added photovoltaic installed capacity, a polynomial stimulus function corresponding to adjacent historical years is constructed. Calculate the optimal solution for each polynomial stimulation function, and construct several training samples based on the optimal solutions for all adjacent historical years and the regional photovoltaic development knowledge characteristics of each optimal solution. The neural network model is trained and validated based on the training samples to obtain a new prediction model, and the new photovoltaic installed capacity for each year is predicted according to the new prediction model.

3. The method for constructing a photovoltaic power generation decommissioning path calculation model based on dynamic material flow according to claim 2, characterized in that, Construct several training samples, including: Each optimal solution is labeled with a corresponding stimulus index tag, and the regional photovoltaic development knowledge features associated with the optimal solution are extracted. The optimal solutions are sorted according to the hierarchical relationship of the stimulus indicator labels, and the temporal correlation weights between each node of stimulus indicator label-optimal solution-knowledge feature are calculated. The sorted optimal solution is associated with the corresponding regional photovoltaic development knowledge features based on the time-series association weight to form an initial feature link. The regional photovoltaic development trend fit is calculated to determine the fit value of each conflicting feature in the initial feature link. Features with fit values ​​higher than a preset threshold are retained to generate a feature link after conflict resolution. Based on the temporal association weights of nodes in the feature link after conflict resolution, a graph structure information containing node temporal attributes and feature confidence is constructed. Based on the relationships between nodes and feature confidence in the graph structure information, the optimal solution, stimulus index labels, and knowledge features are mapped to a two-dimensional matrix to obtain the basic matrix of optimal solution-feature-confidence. After normalizing the base matrix, and extracting the core feature dimensions based on principal component analysis and combining the temporal correlation weights corresponding to the core features, a feature space matrix containing core feature weights, cross-optimal solution correlation degree, and feature confidence is constructed. The initially constructed samples are input into the pre-trained model for trial prediction. Based on the trial prediction error, the core feature weights and feature confidence of the feature space matrix are adjusted in reverse. The temporal feature distribution of the samples is optimized by combining the temporal correlation weights of adjacent historical years. The initially constructed samples are then expanded to obtain several training samples.

4. The method for constructing a photovoltaic power generation decommissioning path calculation model based on dynamic material flow according to claim 3, characterized in that, The initially constructed samples are augmented to obtain several training samples, including: The first adjustment vector is constructed by obtaining the weight adjustment amount of each initially constructed sample, and the second adjustment vector is constructed by obtaining the confidence adjustment amount of each initially constructed sample. The first adjustment vector and the second adjustment vector are aligned and divided into levels to obtain the adjustment level of each initially constructed sample, and a double matrix is ​​obtained, wherein each vector in the double matrix corresponds to an initially constructed sample; Perform two-level clustering analysis on the dual matrix to obtain several clustering results, and obtain the cluster center cluster of each clustering result. Expand the cluster center cluster according to the time-series feature distribution to obtain several double expansion levels, and obtain the new samples of each double expansion level. The newly added samples and the initially constructed samples are used as training samples.

5. The method for constructing a photovoltaic power generation decommissioning path calculation model based on dynamic material flow according to claim 4, characterized in that, The new samples obtained for each double expansion level include: The temporal feature distribution corresponding to the cluster center cluster is hierarchically decomposed to obtain temporal sub-distributions in the dimensions of year, season, and month, and the fluctuation correlation degree between each temporal sub-distribution is calculated; Based on the fluctuation correlation, a temporal expansion weight is assigned to each cluster center cluster, and the temporal constraint boundary of the dual expansion level is determined by combining the adjustment level corresponding to the dual matrix. According to the temporal expansion weight and temporal constraint boundary, the cluster center cluster is expanded bidirectionally in both the forward and reverse temporal directions to generate several double expansion levels. Calculate the temporal fit between the new samples and the original temporal feature distribution under each double expansion level, and filter the new samples whose temporal fit is higher than the preset fit.

6. The method for constructing a photovoltaic power generation decommissioning path calculation model based on dynamic material flow according to claim 1, characterized in that, After generating the second photovoltaic installation addition table, it also includes: Based on the second photovoltaic installation addition table, the predicted photovoltaic installation inflow and predicted photovoltaic installation outflow for each of the n1 historical years are obtained, and compared with the actual photovoltaic installation inflow and actual photovoltaic installation outflow for the corresponding historical years to obtain the basic error vector for each historical year. At the same time, the actual operating condition data and standard operating condition data of each batch of PV modules put into production in the target area in the n1 historical years are collected, along with the extreme vector based on the extreme occurrence probability of each operating condition item. Each batch of production corresponds to one extreme vector. Obtain the reference error matrix of photovoltaic installation data for n1 historical years of a reference region that has the same operating conditions as the corresponding target region. At the same time, obtain the reference extreme probability of each operating condition item for n1 historical years of the reference region and construct the reference probability matrix of the corresponding reference region. Determine a first discrete relationship between each basic error vector and the reference error matrix, and simultaneously determine a second discrete relationship between each extreme vector and the reference probability matrix; Based on the first discrete relationship and the second discrete relationship, the operating error information of each production batch is fused and decomposed in multiple dimensions to obtain the error distribution characteristics, error contribution characteristics and nonlinear characteristics of operating parameters and failure probability of the target area based on different production batches. Based on the decomposition results Make corrections and re-obtain the updated second photovoltaic installation table.

7. The method for constructing a photovoltaic power generation decommissioning path calculation model based on dynamic material flow according to claim 6, characterized in that, Determining the second discrete relationship between each extreme vector and the reference probability matrix includes: Perform time-series alignment processing on each working condition item in the extreme vector and the working condition item corresponding to the reference probability matrix to obtain the time-series aligned extreme sub-vector and reference probability sub-matrix; For each time-aligned operating condition item, calculate the weighted dispersion of the extreme occurrence probability of the corresponding operating condition item in the extreme sub-vector and the reference extreme probability of the corresponding operating condition item in the reference probability sub-matrix; Calculate the correlation weight of each production batch in the target area with the batch of the same operating condition in the reference area. At the same time, calculate the correlation coefficient between the corresponding operating condition item and the failure probability of PV modules. Based on the weighted discreteness of each working condition item, the corresponding working condition correlation weight, and the correlation coefficient, the second discrete relationship between each extreme vector and the reference probability matrix is ​​obtained.

8. The method for constructing a photovoltaic power generation decommissioning path calculation model based on dynamic material flow according to claim 7, characterized in that, Calculating the weighted dispersion includes: Calculate the absolute discrete values ​​of the extreme occurrence probability of the corresponding working condition item in the extreme sub-vector and the reference extreme probability of the corresponding working condition item in the reference probability sub-matrix; The weighted dispersion is obtained by multiplying the absolute discrete value with the corresponding extreme weight coefficient and the working condition area fit.

Images