Photovoltaic energy storage system income measuring and calculating method based on dynamic energy management

Abstract

The invention relates to the technical field of economic evaluation of a power system, and particularly discloses a photovoltaic energy storage system income measuring and calculating method based on dynamic energy management, which comprises the following steps: acquiring multi-dimensional operation state data including generated output, system load and historical price in real time; analyzing the fusion time sequence dependency and fluctuation aggregation characteristics of the data to generate a market clearing price prediction sequence; inputting the state data into a preset market causal knowledge base for reasoning to obtain a causal expected interval, and performing credibility marking on a prediction result by comparing a prediction sequence with the interval; the marked prediction sequence is input into a dynamic optimization model, the model adaptively calls different optimization strategies according to the credibility mark to carry out multi-cycle operation simulation, and a simulation revenue flow is output; and finally, integrating all the income flows, and generating a full-life-cycle income measurement report containing an income range and a risk index.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of economic evaluation of power systems, in particular to a photovoltaic energy storage system benefit calculation method based on dynamic energy management. BACKGROUND

[0002] With the deepening of energy transformation and the gradual improvement of the electricity market mechanism, photovoltaic energy storage systems have become a key way to improve project economics by participating in market transactions to achieve energy time shifting and peak-valley arbitrage. In this context, accurately calculating the benefit capacity of photovoltaic energy storage systems throughout their life cycle is the core basis for project investment decisions, risk assessment, and optimal allocation. Existing benefit calculation methods usually rely on data-driven prediction techniques, i.e., using historical data to train time series prediction models (such as long short-term memory network models) or econometric models to predict future electricity market clearing prices, and then inputting the predicted sequence into an optimization model to simulate the charging and discharging behavior of the energy storage system and calculate the expected benefit.

[0003] The existing technology has the following shortcomings: It ignores the inherent physical laws and market rules in the operation of the electricity market. When the market encounters extreme scenarios not fully covered in historical data, such as temporary price regulation by regulatory agencies due to systemic risks, different prediction models (such as those focusing on time series patterns and those focusing on volatility analysis) may produce mutually reinforcing "collusion bias" due to differences in underlying logic, i.e., both models produce a seemingly high-confidence but actually severely deviating from the market fundamentals incorrect prediction. The optimization model will generate disastrous operation strategies (such as emptying the energy storage during the regulation period) after receiving this false signal, resulting in a severely distorted benefit calculation result. The essential defect of the existing technology is that it lacks a mechanism for checking the causal logic consistency of the prediction results, cannot identify and respond to prediction uncertainty, and thus lacks robustness in critical moments, posing potential risks to investment decisions. SUMMARY

[0004] The purpose of the present application is to provide a photovoltaic energy storage system benefit calculation method based on dynamic energy management to solve the problems in the above background.

[0005] The purpose of the present application can be achieved by the following technical solutions: A photovoltaic energy storage system benefit calculation method based on dynamic energy management, comprising the following steps: S1: Real-time acquisition of multi-dimensional operating state data of a photovoltaic energy storage system connected to the electricity market; the multi-dimensional operating state data includes generator set output data, system load data, and historical price data; S2: analyze the fusion time sequence dependence and fluctuation aggregation characteristics of the multi-dimensional running state data, and generate a market clearing price prediction sequence for a future preset period; S3: input the multi-dimensional running state data into a preset market causal knowledge base for reasoning to obtain a causal expectation interval of the market price; compare the market clearing price prediction sequence with the causal expectation interval, and mark the price prediction sequence with a credibility according to the comparison result; S4: input the price prediction sequence marked with credibility into a dynamic optimization model of the photovoltaic energy storage system; the dynamic optimization model calls corresponding optimization strategies for multi-cycle running simulation according to different credibility markings, so as to output corresponding simulation income streams; S5: integrate all simulation income streams to generate a full life cycle income measurement result of the photovoltaic energy storage system.

[0006] As a further scheme of the application, the analysis of the fusion time sequence dependence and fluctuation aggregation characteristics of the multi-dimensional running state data specifically includes: The generator output data, system load data and historical price data are respectively input into a plurality of parallel time sequence analysis units, each unit uses a sliding window of different time length to convolve the input data to capture the dependence characteristics of short-term fluctuations, medium-term trends and long-term periodicity respectively; The feature vectors from different data sources at the same time point are spliced to form a fusion feature tensor, which contains the correlation information of energy supply, consumption demand and market price; The change rate of the fusion feature tensor on the continuous time step is calculated, and the persistence of the variance is identified; by monitoring the change rate variance, the time period and intensity of the fluctuation aggregation effect are determined, and the intensity is quantified as a weight.

[0007] As a further scheme of the application, the generation of the market clearing price prediction sequence for a future preset period specifically includes: The fusion feature tensor and the dynamic weight are jointly input into a recurrent neural network guided by a double attention mechanism; The first attention of the double attention mechanism is used for the key historical period most related to the prediction time in the fusion feature tensor, and the second attention is used for the period with high fluctuation aggregation effect intensity; The recurrent neural network generates the price prediction point value of each time step in the future according to the features weighted by the double attention; Based on the uncertainty propagation of the dynamic weight and the internal state of the recurrent neural network, the confidence interval of each prediction point value is calculated, and finally the market clearing price prediction sequence composed of the prediction point value and the confidence interval is output.

[0008] As a further scheme of the present application, the causal expectation interval of the market price specifically comprises: The real-time acquired generator set output data and system load data are compared with the threshold in the preset knowledge base to determine that the current market is in a specific state of supply shortage, supply-demand balance or oversupply, and the officially released mandatory intervention instruction is instantiated into an activated or inactivated state; Based on the instantiated market state, the pre-defined causal rule network in the knowledge base is traversed; if the mandatory intervention instruction is instantiated into an inactivated state, the path from the supply-demand state to the theoretical price is deduced; if the instruction is instantiated into an activated state, the path from the intervention instruction to the price control is deduced, and the supply-demand path is interrupted; According to the activated final causal path, a corresponding basic price interval is mapped; and according to the network congestion state data, the boundary of the basic price interval is widened and corrected, and finally the causal expectation interval is output.

[0009] As a further scheme of the present application, the credibility marking of the price prediction sequence according to the comparison result specifically comprises: Each prediction point in the price prediction sequence is judged one by one whether it falls within the causal expectation interval, the proportion of the number of points falling outside the interval to the total points is calculated and recorded as the interval deviation, and the maximum length of the continuous deviation points is recorded; If the interval deviation is lower than a first threshold value and there is no continuous long sequence deviation, it is determined that there is no conflict; If the interval deviation is higher than the first threshold value but lower than a higher second threshold value, it is determined that there is a mild causal conflict; if the interval deviation is higher than the second threshold value, the dominant path deduced by the knowledge base is combined to determine that the conflict is caused by a supply-demand logic conflict or an intervention rule conflict; According to the conflict mode recognition result, the whole price prediction sequence is marked with corresponding credibility: no conflict corresponds to high credibility, mild causal conflict corresponds to medium credibility, and explicit supply-demand logic conflict or intervention rule conflict corresponds to low credibility.

[0010] As a further scheme of the present application, the construction process of the dynamic optimization model is: The input of the dynamic optimization model is the market clearing price prediction sequence with credibility marking, and the output is the optimal charge-discharge power instruction sequence of the photovoltaic energy storage system in each time unit in the future multiple operation cycles; A target function is constructed, taking the total economic benefit in the whole simulation cycle as the core target and taking the health loss of the battery energy storage system as a soft constraint; wherein, the economic benefit target is calculated by the power exchange cost and income of the power grid, and the health loss constraint is embodied by limiting the charge-discharge cycle depth and average state of charge; In the objective function, when marked as high reliability, priority is given to guarantee economic benefits; when marked as low reliability, priority is given to guarantee the safety of the energy storage system and the power reserve.

[0011] As a further scheme of the present application, the output corresponding simulation income stream specifically comprises: The optimal charging and discharging power instruction sequence generated by the dynamic optimization model is substituted into a system simulation environment including photovoltaic power generation simulation, battery state of charge calculation and electric energy buying and selling accounting, and the running state of the system in the entire life cycle is simulated step by step in time sequence; In each time step of the simulation, the electricity fee expenditure or electricity sales income of the corresponding step is calculated according to the current time charging and discharging instruction, the measured photovoltaic power generation and the corresponding market clearing price, and is accumulated to form the annual net cash flow; The net cash flow of each simulation year is arranged in the order of its occurrence time to form a simulation income stream running through the entire life cycle of the photovoltaic energy storage system; the simulation income stream also records the income details generated under different strategy modes of the dynamic optimization model.

[0012] As a further scheme of the present application, the integration of all simulation income streams generates the full life cycle income estimation result of the photovoltaic energy storage system, specifically comprising: According to the reliability mark of the price prediction sequence on which each simulation income stream is based, each income stream is assigned a corresponding weight; The cash flow of each simulation income stream at each time point is multiplied by its corresponding weight and superimposed to form a weighted comprehensive income curve; the change range of all possible cash flows at different time points is counted to generate a band distribution diagram reflecting the income uncertainty; Based on the band distribution diagram, the weighted average net present value and the internal rate of return of the project are calculated; From the band distribution diagram, the lower limit value of the income under a specific confidence level and the income fluctuation range are extracted, and the value index and the risk index are jointly encapsulated into the final full life cycle income estimation result report.

[0013] The present application has the following beneficial effects: (1) The present application constructs a double analysis mechanism, which not only uses a deep learning model to capture the complex time sequence characteristics of market prices, but also introduces a rule-based causal knowledge base to verify the logical consistency of the prediction results and generate a reliability mark. This mechanism effectively identifies and corrects the prediction bias caused by the model misjudging the supply and demand relationship or ignoring policy intervention, enhances the robustness of price prediction in extreme or mutation scenarios, and provides a higher reliability input basis for subsequent optimization decisions.

[0014] (2) The application dynamically adjusts the optimization strategy of the energy storage system according to the credibility of the price prediction, pursues maximum profit when the credibility is high, and prioritizes system safety and power storage when the credibility is low, thereby achieving intelligent trade-off between aggressive arbitrage and conservative operation. By weighting and fusing the simulated income streams under multiple risk scenarios, the final output includes the comprehensive measurement results of the expected income curve and the fluctuation range, which not only quantifies the economic value (such as net present value, internal rate of return) of the project, but also clearly reveals its potential risks, supporting investors to make more scientific and comprehensive investment decisions in an uncertain environment. BRIEF DESCRIPTION OF DRAWINGS

[0015] The application will be further described below in conjunction with the accompanying drawings.

[0016] Figure 1 is a flow chart of the method of the application. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the application will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the application. Obviously, the described embodiments are only a part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the application.

[0018] Please refer to Figure 1 The application is a photovoltaic energy storage system income measurement method based on dynamic energy management, comprising the following steps: S1: Real-time acquisition of multi-dimensional operation state data of the photovoltaic energy storage system connected to the power market; the multi-dimensional operation state data includes generator set output data, system load data and historical price data; S2: Analysis of the fusion time sequence dependence and fluctuation aggregation characteristics of the multi-dimensional operation state data, to generate a market clearing price prediction sequence for a future preset period; S3: Inputting the multi-dimensional operation state data into a preset market causal knowledge base for reasoning to obtain a causal expectation interval of the market price; comparing the market clearing price prediction sequence with the causal expectation interval, and marking the credibility of the price prediction sequence according to the comparison result; S4: Inputting the price prediction sequence marked by the credibility into a dynamic optimization model of the photovoltaic energy storage system; the dynamic optimization model calls corresponding optimization strategies for multi-cycle operation simulation according to different credibility markings, thereby outputting corresponding simulated income streams; S5: Integrating all simulated income streams to generate a full life cycle income measurement result of the photovoltaic energy storage system.

[0019] In S1, real-time multi-dimensional operation state data of the photovoltaic energy storage system accessing the electricity market is acquired; the multi-dimensional operation state data includes generator unit output data, system load data, and historical price data, specifically including: The acquisition process of the generator unit output data is as follows: the data is used to represent the real-time power supply capacity of the entire power system. The data acquisition source is the real-time market information data interface published by the power dispatch center or the independent system operator (ISO) of the region where the photovoltaic energy storage system is connected. Through a pre-set application programming interface (API) or file data service (FTP / SFTP), the total power generation output data of the entire network or key nodes and the composition ratio of various types of energy (such as thermal power, gas power, hydroelectric power, wind power, etc.) are automatically acquired at specific time intervals (for example, every 5 minutes or every 15 minutes). The acquired raw data is usually in a structured data format, such as XML or JSON. After data acquisition, validity check is required, for example, checking whether the data is within a reasonable physical range, and using linear interpolation method of adjacent time point data to complete the short-term data loss caused by network transmission, etc. The data verified and cleaned is assigned a uniform time stamp and stored in the designated data table of the local or cloud database for subsequent analysis and call.

[0020] The acquisition process of the system load data is as follows: the data is used to represent the real-time power demand of the entire power system. Its data source is also the system total load forecast or actual value published by the power dispatch center or system operator. The data acquisition method is similar to that of the generator unit output data, which is achieved by regular API calls. In addition to acquiring the actual load value at the current time, load prediction data for a specific period in the future (such as the next 24 hours) is also required. These prediction data are crucial for early judgment of market supply and demand tension trends. The acquired load data also needs to be checked for range and missing values. In addition, since load data usually has obvious periodicity (daily cycle, weekly cycle), simple smoothing filtering can be performed on it during data preprocessing to eliminate obvious measurement noise without changing its overall trend. The processed load data and the generator output data are aligned in time stamp to ensure the consistency of the data for subsequent analysis.

[0021] The process of obtaining historical price data is as follows: this data is the basis for training price prediction models and conducting market analysis. Historical price data refers to the settlement price sequence of the power market in the past period (such as the past year), usually including day-ahead market prices and real-time market prices. The data source is the public historical data release platform of the power trading center. Data acquisition is usually completed in the form of batch download, such as automatic download of all price data of the previous trading day every day. The preprocessing of historical price data includes outlier rejection, such as a small number of price points that are obviously outside the normal fluctuation range (such as three times the standard deviation outside the average value) are considered as outliers caused by special events and are removed to avoid interference with model training. At the same time, the price data of different time granularity need to be standardized and aligned, such as aggregating 5-minute interval real-time price data into 1-hour average value, so as to keep consistent with the common time scale of load and output data. The historical price data cleaned and standardized is constructed into a time series database for subsequent model training and analysis.

[0022] In S2, the fusion time series dependence and fluctuation aggregation characteristics of multi-dimensional operating state data are analyzed to generate a market settlement price prediction sequence for a future preset period, which specifically includes: The three types of time series data, generator unit output data, system load data and historical price data from step S1, are input into three independent data processing flows. Each flow uses a sliding window of a specific time length to perform convolution operation on the input data. Specifically, the first flow uses a sliding window of length 6 time units, aiming to capture the short-term fluctuation characteristics in the data, such as rapid changes within a few hours; the second flow uses a sliding window of length 24 time units, aiming to extract the medium-term trend characteristics of the data, such as daily periodic variation rules; the third flow uses a sliding window of length 168 time units, aiming to capture the long-term periodic characteristics contained in the data, such as periodic variation patterns. The output of each convolution operation is the weighted average result of the data within the window, which represents the smoothed shape of the original data at the corresponding time scale. For each type of original data (output, load, price), three feature values representing short-term, medium-term and long-term characteristics will be generated in parallel at each current time.

[0023] For the same time point, short-term, medium-term, and long-term feature values from the generator set output data, short-term, medium-term, and long-term feature values from the system load data, and short-term, medium-term, and long-term feature values from the historical price data are connected in a predetermined order to form a comprehensive feature vector. This feature vector contains information on energy supply, consumption demand, and historical prices at multiple time scales at the same time. These comprehensive feature vectors at consecutive time points are arranged in chronological order, i.e., a three-dimensional fusion feature tensor is formed, and the three dimensions of the fusion feature tensor correspond to time points, data source categories, and time scales, respectively.

[0024] The first-order difference of the fusion feature tensor at consecutive time steps is calculated, i.e., the change in each dimension of the feature vector at adjacent time points is calculated. The variance characteristics of these change sequences are analyzed. By calculating the variance of the change sequence within a rolling time window (e.g., 24 time units) and observing whether the variance value itself is persistent, i.e., high variance stages tend to be followed by high variance stages, and low variance stages tend to be followed by low variance stages, the volatility clustering effect is determined. When the rolling variance is detected to be consistently higher than its long-term average, it is considered that there is a volatility clustering phenomenon in this period. The ratio of the rolling variance value to the long-term average variance, or its standardized value, is quantified as a dynamic weight, which reflects the relative intensity of the current market volatility.

[0025] The fusion feature tensor and the dynamic weight are used as inputs to the prediction model. The core of the prediction model is a recurrent neural network assisted by a dual attention mechanism.

[0026] The dual attention mechanism is applied. The first attention mechanism acts on the time dimension, and its purpose is to assign appropriate importance weights to the feature vectors at different historical time points in the fusion feature tensor. This mechanism achieves this by calculating the correlation score between the prediction target time and each historical time feature vector. The historical time with high correlation obtains a higher attention weight, so that the model can focus on the most relevant historical patterns. The second attention mechanism focuses specifically on the volatility clustering effect, and its input is the dynamic weight sequence representing the volatility intensity. This mechanism automatically increases the influence of the historical feature vector at the corresponding time when the dynamic weight is high, forcing the model to pay more attention to the patterns in the historical high volatility period, so as to learn how to make predictions in new volatility periods.

[0027] Sequence prediction is performed. The recurrent neural network receives the historical feature sequence adjusted by the double attention weight. The network processes the input sequence step by step by time step, and passes the internal state backward. After processing all the historical time steps, the final state of the network contains a summary of the historical information. Then, the network enters the decoding prediction phase, taking the prediction output and internal state of the previous time as input, and recursively generates the market clearing price prediction point value of each future preset time step.

[0028] Prediction uncertainty is evaluated. While generating point predictions, the uncertainty information contained in the internal state of the recurrent neural network is utilized, and the dynamic weight of the external input (as a volatility prior) is combined to estimate the possible distribution range of the prediction value at each future time point through the forward propagation process inside the model. Specifically, the model outputs the conditional variance or quantile of each prediction point value to determine a confidence interval. Finally, the output is a complete market clearing price prediction sequence composed of a series of point prediction values and their corresponding confidence intervals.

[0029] In S3, the multi-dimensional running state data is input into the preset market causal knowledge base for reasoning to obtain the causal expectation interval of the market price; the market clearing price prediction sequence is compared with the causal expectation interval, and the confidence of the price prediction sequence is marked according to the comparison result, specifically including: Market state and intervention instruction instantiation is performed. The generator output data and system load data obtained in real time in step S1 are calculated to obtain the net supply and demand difference at the current time (i.e., the system load data minus the generator output data). The calculated net supply and demand difference is compared with the threshold value stored in the preset knowledge base, which is usually determined based on historical statistical quantiles, for example, the 70% quantile of the historical net supply and demand difference can be set as the judgment threshold of the shortage state, and the 30% quantile can be set as the judgment threshold of the oversupply state. If the current net supply and demand difference is higher than the 70% quantile threshold, it is determined that the current market is in a specific state of "shortage"; if it is lower than the 30% quantile threshold, it is determined to be in a "surplus" state; if it is between the two, it is determined to be in a "balanced supply and demand" state. At the same time, the mandatory intervention instruction text obtained from the official channel is parsed, for example, whether the text contains keywords such as "price ceiling", "suspend market pricing", "implement control" and the like. If such keywords are identified and the instruction is within the effective period, the intervention instruction is instantiated as an "activated" state; otherwise, it is instantiated as an "unactivated" state. The purpose of this step is to convert continuous real-time data into discrete logical states that can be processed by the knowledge base.

[0030] Logical path deduction based on the causal rule network is performed. The pre-set knowledge base contains a predefined causal rule network, which is composed of nodes and directed edges, where the nodes represent different market states or events (e.g. "shortage" "intervention activated"), and the directed edges represent the causal relationship between the nodes (e.g. "shortage" can lead to "price up"). Based on the instantiated states (e.g. "shortage" and "intervention not activated") obtained in the first step, the system starts from the initial state node and traverses the directed edges in the causal rule network. When a node with branches is encountered, the path is selected according to the instantiated states. Specifically, if the "mandatory intervention instruction" is instantiated as "not activated", the reasoning path will follow the edge from "supply and demand state" to "theoretical price direction" and deduce, for example, from the "shortage" state to the "price theoretically up" state. Conversely, if the "mandatory intervention instruction" is instantiated as "activated", the reasoning path will follow the edge from "intervention instruction" to "price control", which will interrupt or override the aforementioned path based on supply and demand relationship, and directly conclude that "price is controlled". This step determines the causal chain that plays a dominant role in the current market environment.

[0031] Mapping and correction of price interval is performed. According to the final causal path activated in the second step of deduction, the basic price interval pre-set for this path in the knowledge base is queried. For example, for the "shortage and no intervention" path, the basic price interval may be mapped to the price range of the historical high price period (e.g. the top 10% of the price period); for the "intervention activated" path, the basic price interval may be mapped to the price upper limit specified by the instruction or a very narrow regulated price range. Then, the network congestion state data obtained in step S1 is introduced to correct the basic price interval. If there is severe network congestion, it indicates that local supply and demand imbalance may intensify, and price volatility will increase. Therefore, the correction process usually involves widening the upper and lower boundaries of the basic price interval by a certain percentage (e.g. 15% outward expansion of the upper and lower boundaries when the congestion level is high) to reflect additional uncertainty. The final output is the corrected causal expectation interval reflecting the current causal logic and physical constraints.

[0032] The deviation of the prediction sequence from the causal interval is quantified. It is determined whether each predicted point value in the price prediction sequence at each future time point falls within the causal expected interval at the corresponding time obtained in the above step. The total number of predicted points that fall outside the causal expected interval is counted within the entire preset prediction period (e.g., 24 hours in the future), and the percentage of the total number of points in the prediction period is calculated. This percentage is defined as the interval deviation. At the same time, the entire prediction sequence is scanned to find a sequence segment consisting of consecutive predicted points that fall outside the causal expected interval, and the length of the longest segment, i.e., the maximum length of consecutive deviated points, is recorded. These two indicators respectively depict the degree of deviation between the prediction sequence and the causal logic expectation from the whole and the local.

[0033] The conflict pattern is identified and classified. Two deviation threshold values are set, for example, the first threshold value is set to 10%, and the second threshold value is set to 30%. The calculated interval deviation is compared with the two threshold values. If the interval deviation is lower than the first threshold value (10%) and the maximum length of consecutive deviated points does not exceed a set value (e.g., 5 consecutive time points), it is determined that there is "no conflict" between the prediction sequence and the causal expectation. If the interval deviation is higher than the first threshold value (10%) but lower than the second threshold value (30%), it is determined that there is "mild causal conflict". If the interval deviation is higher than the second threshold value (30%), it is determined that there is a significant conflict, and further classification is needed in combination with the dominant path inferred from the causal knowledge base: if the dominant path is based on supply and demand state deduction (such as "shortage of supply"), it is determined that the conflict is caused by "supply and demand logic conflict", i.e., the price signal identified by the prediction model is significantly inconsistent with the conclusion derived from the current supply and demand fundamentals; if the dominant path is based on intervention instruction deduction (such as "intervention activation"), it is determined that the conflict is caused by "intervention rule conflict", i.e., the prediction model fails to fully respond to the constraints of price regulation, and its prediction result is still based on free market logic.

[0034] The credibility label is assigned. According to the conflict pattern identification result of the second step, a comprehensive credibility label is assigned to the entire price prediction sequence. When it is determined to be "no conflict", a "high credibility" label is assigned, indicating that the prediction sequence is highly consistent with the causal logic. When it is determined to be "mild causal conflict", a "medium credibility" label is assigned, indicating that the prediction sequence has some deviation, but has not completely deviated from the basic logic. When it is determined to be a clear "supply and demand logic conflict" or "intervention rule conflict", a "low credibility" label is assigned, indicating that the prediction sequence has a serious divergence from the key causal expectation, and its reliability is questionable. This credibility label will be passed as a key parameter to the subsequent optimization model to guide it to adopt different risk preference decision strategies.

[0035] In S4, the credibility-labeled price prediction sequence is input into the dynamic optimization model of the photovoltaic energy storage system; the dynamic optimization model calls the corresponding optimization strategy for multi-cycle operation simulation according to different credibility labels, thereby outputting the corresponding simulation income stream, specifically including: The main input of the dynamic optimization model is the market clearing price prediction sequence checked in step S3 and attached with a credibility label. This sequence contains the predicted price point value, uncertainty interval and a comprehensive credibility label (such as high, medium and low) of each time unit (such as 1 hour) in the future period (such as the next 24 hours or longer). The output of the model is the optimal charge and discharge power instruction sequence required by the photovoltaic energy storage system in each time unit within the same future period. The instruction sequence specifically specifies that in each time unit, the battery should be charged (absorbing power from the grid or photovoltaic system, represented by a negative value) or discharged (releasing power to the grid, represented by a positive value) at a certain power, and the remaining power of photovoltaic power generation after meeting local load is put on the grid.

[0036] The core of the dynamic optimization model is an objective function that needs to be optimized in the entire simulation period. The objective function takes maximizing the total economic benefit of the system as the core goal. The calculation of economic benefit is based on the exchange of electric energy with the grid: the difference between the cost of charging in the period with low electricity price and the income obtained by discharging (or surplus power on the grid) in the period with high electricity price, and the net income accumulated in the whole period. At the same time, the objective function contains a soft constraint on the health loss of the battery energy storage system. The soft constraint is not an absolute restriction, but is embodied by introducing a penalty term. Specifically, it is realized in two ways: one is to limit the depth of a single charge and discharge cycle, that is, to avoid the state of charge of the battery from changing too much in a single cycle, for example, to limit the depth of each cycle to no more than 60% of the total capacity of the battery; the second is to control the average state of charge of the battery to maintain it within an ideal range (for example, 40% to 80%), avoiding the battery being in a fully charged or discharged state for a long time, thereby slowing down the battery aging. These constraint conditions are integrated into the construction of the objective function in the form of inequalities, and when the optimization result deviates from these ideal states, the corresponding virtual loss cost will be deducted from the total income.

[0037] The objective function contains adjustable weight parameters, whose values are dynamically controlled by the input confidence label. When the confidence label is "high confidence", the model sets the weight of the economic return objective to the highest, while setting the weight of the penalty term representing battery degradation to a relatively lower value. This means that the optimization process will prioritize capturing market arbitrage opportunities, allowing the battery to undergo deeper and more frequent charge-discharge cycles to maximize returns, and the strategy is thus aggressive. When the confidence label is "low confidence", the model performs the opposite operation, significantly increasing the weight of the battery health degradation penalty term while reducing the weight of the economic return objective. This makes the optimization process prioritize the safety and power reserve of the energy storage system, and the strategy becomes conservative, manifested as limiting the charge-discharge power, avoiding deep discharge, and maintaining a higher level of standby power. For a "medium confidence" label, a weight setting between the two is adopted. In this way, the behavior of the optimization model can adaptively respond to the risk brought by prediction uncertainty.

[0038] The specific process of outputting the simulated return stream is as follows: The optimal charge-discharge power instruction sequence generated by the dynamic optimization model is input into a high-precision system simulation environment. This simulation environment contains the following several interrelated calculation parts: a photovoltaic power generation simulation part, which calculates the actual power generation at each time step based on the rated power of the photovoltaic panel, historical meteorological data (such as solar irradiance, ambient temperature), and the efficiency of the photovoltaic system. A battery energy storage state calculation part, which updates the state of charge of the battery in real time based on the initial state of charge, the charge-discharge power instruction at each time step, and the charge-discharge efficiency of the battery, and ensures that it is always between the minimum and maximum allowed values. An electricity buying and selling accounting part, which is responsible for recording the energy flow between the system and the grid. The simulation is carried out in chronological order, starting from the initial state, simulating the operation of the system in each year of the entire predetermined life period (e.g. 25 years), and considering the annual degradation of device performance such as the annual depreciation of photovoltaic panel efficiency year by year.

[0039] At each time step (e.g. 1 hour) of the simulation run, financial calculations are made based on the current system state. Specifically, first determine the net grid interaction power of the system at this time, which is the photovoltaic power generation minus the local load consumption, plus (or minus) the discharge (or charge) power of the battery. If the net interaction power is positive, indicating that the system sells electricity to the grid, the income is equal to the power value multiplied by the market clearing price at that time (using the value in the price prediction sequence relied upon during optimization). If the net interaction power is negative, indicating that the system purchases electricity from the grid, the cost is equal to the absolute value of the power multiplied by the corresponding electricity price. The net cash flow of this time step is the difference between the electricity sales income and the electricity purchase cost. All the net cash flows of the time steps in a simulation year are accumulated, and the annual operating and maintenance costs of that year are deducted, to obtain the annual net cash flow of that year.

[0040] After the whole life cycle simulation is completed, the calculated annual net cash flow of each year is arranged in a time series according to the chronological order of the years. This time series is the simulation income stream throughout the life cycle of the photovoltaic energy storage system. It needs to be specially pointed out that, since the dynamic optimization model will adopt different strategies according to the reliability label, for the same set of input data, it may run a simulation under "high reliability" to generate a "standard income stream" and run another simulation under "low reliability" to generate a "conservative income stream". The final output of the simulation income stream set will include these income details generated under different risk strategies, providing a comprehensive data basis for subsequent comprehensive income evaluation. The income stream clearly reflects the expected cash inflows and outflows of the system throughout the life cycle.

[0041] In S5, all simulation income streams are integrated to generate the life cycle income estimation results of the photovoltaic energy storage system, including: The qualitative reliability assessment generated in step S3 is quantitatively integrated into the final comprehensive analysis. The system will read the reliability label attached to each price prediction sequence used in the simulation run in step S4. According to the label, a weight is assigned to the corresponding simulation income stream. Specifically, if the reliability label is "high reliability", a higher weight is assigned to the corresponding income stream, for example 0.7; if the reliability label is "medium reliability", a medium weight is assigned, for example 0.2; if the reliability label is "low reliability", a lower weight is assigned, for example 0.1. The sum of all weights should be 1 to ensure the effectiveness of the weighted calculation. This allocation reflects the decision maker's preference for the corresponding income scenarios under different prediction reliabilities, considering that the income results based on high reliability predictions are more likely to approach the actual future situation.

[0042] The cash flow value of each of the weighted simulated revenue streams at each identical time point (e.g. each year of the operational life) is multiplied by its corresponding weight. Then, all these weighted cash flow values at the same time point are algebraically summed to obtain the weighted average cash flow at that time point. Connecting all the weighted average cash flows at the time points in chronological order, a weighted aggregate revenue curve is formed, which represents the expected revenue path considering the confidence of the forecast. To visually demonstrate the risk, a revenue band distribution chart also needs to be generated. At each time point, the maximum and minimum values of all the simulated revenue streams (without weighting) at that point are found to determine the range. Further, the mean and standard deviation of all the cash flows at that time point can be calculated. With the time points as the horizontal axis and the cash flows as the vertical axis, connecting the range of "mean plus or minus two times the standard deviation" at each time point, or connecting the minimum and maximum values, a band distribution chart is formed, which reflects the uncertainty of the revenue. This chart clearly shows the range of the possible fluctuations of the revenue at each year.

[0043] Based on the obtained weighted aggregate revenue curve, the traditional capital budgeting metrics are calculated. The weighted average cash flow at each year of the curve is discounted to the current time point at a selected discount rate (e.g. 8%), and all the discounted values are added to obtain the weighted average net present value of the project. The weighted internal rate of return of the project is obtained by iteratively calculating the discount rate at which the net present value is equal to zero. In addition, risk metrics are extracted from the revenue band distribution chart. For example, at a 95% confidence level, the lower bound of the revenue at each time point (which can be approximated as the mean minus one and a half times the standard deviation) is determined to assess the worst-case revenue scenario. Meanwhile, the standard deviation or coefficient of variation of the revenue over the entire life can be calculated as a quantitative indicator of the range of fluctuations of the revenue. Finally, the calculated value metrics of the weighted average net present value and the weighted internal rate of return, together with the risk metrics of the lower bound of the revenue and the range of fluctuations of the revenue, are encapsulated in a structured electronic report (such as a PDF document or a database record) along with the weighted aggregate revenue curve and the revenue band distribution chart, which constitute the final full-life-cycle revenue estimation results.

[0044] The working principle of the present application is: firstly, real-time collection of power market generation output, load demand and historical price multi-dimensional operation data, use of recursive neural network with multiple time scale convolution and double attention mechanism to mine the time sequence dependence and volatility aggregation characteristics between data, generate future market price prediction sequence containing uncertainty interval; input multi-dimensional operation data into the preset market causal knowledge base, combine net supply and demand state recognition and forced intervention instruction analysis, form causal expected price interval reflecting current market fundamentals and policy constraints through causal rule network deduction, and through comparison of the deviation degree of the prediction sequence and the causal interval, quantify the conflict mode and give "high / medium / low" three-level credibility label; input the labeled price prediction sequence into the dynamic optimization model of the photovoltaic energy storage system, dynamically adjust the weight ratio of economic benefit and battery loss penalty in the objective function according to the credibility level, respectively execute aggressive, balanced or conservative charging and discharging strategy, carry out multi-cycle operation simulation to output simulated income stream under different risk scenarios; in the income integration stage, according to the credibility weight corresponding to each income stream, weighted fusion is carried out to generate expected income curve and risk distribution band, and according to this, calculate the key indicators such as weighted net present value, internal rate of return and income volatility, form a comprehensive measurement report covering income level and risk characteristics, so as to improve the scientificity and robustness of photovoltaic energy storage system investment decision in uncertain market environment.

[0045] The above describes one embodiment of the present application in detail, but the content described is only the preferred embodiment of the present application, and cannot be considered as limiting the scope of the present application. Any equivalent changes and improvements made according to the scope of the present application should still belong to the scope of the present application.

Claims

1. A method for calculating the revenue of a photovoltaic energy storage system based on dynamic energy management, characterized in that, Includes the following steps: S1: Real-time acquisition of multi-dimensional operational status data of photovoltaic energy storage systems connected to the electricity market; The multi-dimensional operating status data includes: generator set output data, system load data, and historical price data; S2: Analyze the fusion time-series dependence and volatility clustering characteristics of multi-dimensional operational status data to generate a market liquidation price prediction sequence for a future preset time period; S3: Input multi-dimensional operational status data into a preset market causal knowledge base for reasoning to obtain the causal expectation range of market prices; compare the market liquidation price prediction sequence with the causal expectation range, and mark the credibility of the price prediction sequence based on the comparison results; S4: Input the price prediction sequence after confidence labeling into the dynamic optimization model of the photovoltaic energy storage system; the dynamic optimization model calls the corresponding optimization strategy to perform multi-cycle operation simulation according to different confidence labels, thereby outputting the corresponding simulated revenue stream; S5: Integrate all simulated revenue streams to generate full lifecycle revenue calculation results for the photovoltaic energy storage system.

2. The method for calculating the revenue of a photovoltaic energy storage system based on dynamic energy management according to claim 1, characterized in that, The analysis of the time-series dependency and fluctuation clustering characteristics of the fusion of multi-dimensional operational status data specifically includes: Generator output data, system load data, and historical price data are input into multiple parallel time series analysis units. Each unit uses a sliding window of different time lengths to convolve the input data to capture the dependence features of short-term fluctuations, medium-term trends, and long-term cycles, respectively. Feature vectors from different data sources at the same point in time are concatenated to form a fused feature tensor, which contains information about the relationship between energy supply, consumption demand and market prices. Calculate the rate of change of the fusion feature tensor over a continuous time step and identify the persistence of its variance; by monitoring the variance of the rate of change, determine the timing and intensity of the fluctuation clustering effect, and quantify this intensity as a weight.

3. The method for calculating the revenue of a photovoltaic energy storage system based on dynamic energy management according to claim 1, characterized in that, The generation of the market liquidation price prediction sequence for a future preset time period specifically includes: The fused feature tensor and dynamic weights are input together into a recurrent neural network guided by a dual attention mechanism; The first layer of attention in the dual attention mechanism is used to fuse the key historical periods most relevant to the prediction time in the feature tensor, while the second layer of attention is used in periods with high intensity of fluctuation clustering effect. The recurrent neural network generates the price prediction point value for each future time step in turn based on the features weighted by double attention. Based on the dynamic weights and the uncertainty propagation of the internal state of the recurrent neural network, the confidence interval of each prediction point value is calculated, and the final output is a market clearing price prediction sequence composed of the prediction point values ​​and their confidence intervals.

4. The method for calculating the revenue of a photovoltaic energy storage system based on dynamic energy management according to claim 1, characterized in that, The causal expectation range of the market price specifically includes: The real-time generator output data and system load data are compared with the thresholds in the preset knowledge base to determine whether the current market is in a specific state of supply shortage, supply and demand balance or supply over demand, and the official mandatory intervention instructions are instantiated as active or inactive. Based on the instantiated market state, the predefined causal rule network in the knowledge base is traversed; if the mandatory intervention instruction is instantiated as inactive, the path from the supply and demand state to the theoretical price direction is deduced; if the instruction is instantiated as active, the path from the intervention instruction to price control is deduced, interrupting the supply and demand path. Based on the activated final causal path, a corresponding basic price range is mapped; then, based on network congestion status data, the basic price range is widened and corrected, and finally, the causal expectation range is output.

5. The method for calculating the revenue of a photovoltaic energy storage system based on dynamic energy management according to claim 1, characterized in that, The step of assigning confidence levels to the price prediction sequence based on the comparison results specifically includes: Determine whether each prediction point in the price prediction sequence falls within the causal expectation interval, count the proportion of points falling outside the interval to the total number of points, record the interval deviation, and record the maximum length of consecutive deviation points. If the interval deviation is lower than the first threshold and there is no continuous long sequence deviation, it is determined to be conflict-free; If the deviation of the interval is higher than the first threshold but lower than the second threshold, it is judged as a mild causal conflict; if the deviation of the interval is higher than the second threshold, the conflict is judged to originate from a supply and demand logic conflict or an intervention rule conflict, based on the dominant path inferred from the knowledge base. Based on the conflict pattern recognition results, the entire price prediction sequence is assigned a corresponding confidence level label: no conflict corresponds to high confidence, mild causal conflict corresponds to medium confidence, and clear supply and demand logic conflict or intervention rule conflict corresponds to low confidence.

6. The method for calculating the revenue of a photovoltaic energy storage system based on dynamic energy management according to claim 1, characterized in that, The process of constructing the dynamic optimization model is as follows: The input to the dynamic optimization model is a market clearing price prediction sequence with confidence labels, and the output is the optimal charge and discharge power command sequence for each time unit in multiple future operating cycles of the photovoltaic energy storage system. An objective function is constructed with the total economic benefit over the entire simulation period as the core objective and the health loss of the battery energy storage system as the soft constraint. The economic benefit objective is calculated in terms of the energy exchange cost and revenue with the power grid, while the health loss constraint is reflected by limiting the charge-discharge cycle depth and the average state of charge. In the objective function, when marked as high confidence, priority is given to ensuring economic benefits; when marked as low confidence, priority is given to ensuring the safety of the energy storage system and the reserve of electricity.

7. The method for calculating the revenue of a photovoltaic energy storage system based on dynamic energy management according to claim 1, characterized in that, The simulated revenue stream corresponding to the output specifically includes: The optimal charge and discharge power command sequence generated by the dynamic optimization model is substituted into a system simulation environment that includes photovoltaic power generation simulation, battery state of charge calculation and electricity trading accounting, and the system's operating status throughout its entire lifespan is simulated step by step in chronological order. Within each time step of the simulation, the electricity cost or electricity sales revenue for the corresponding step is calculated based on the current charging and discharging command, the measured photovoltaic power generation, and the corresponding market settlement price, and the annual net cash flow is accumulated. The net cash flow for each simulated year is arranged in chronological order to form a simulated revenue stream that spans the entire lifecycle of the photovoltaic energy storage system. The simulated revenue stream also records the revenue details generated under different strategy modes of the dynamic optimization model.

8. The method for calculating the revenue of a photovoltaic energy storage system based on dynamic energy management according to claim 1, characterized in that, The process of integrating all simulated revenue streams to generate the full lifecycle revenue calculation results for the photovoltaic energy storage system specifically includes: Each revenue stream is assigned a corresponding weight based on the confidence level of the price prediction sequence on which it was generated. The cash flow of each simulated revenue stream at each time point is multiplied by its corresponding weight and then superimposed to form a weighted composite revenue curve; the range of change of all possible cash flows at different time points is statistically analyzed to generate a band distribution chart that reflects the uncertainty of revenue. Based on the strip distribution map, calculate the weighted average net present value and internal rate of return of the project; From the band distribution chart, the lower limit of returns and the range of return fluctuations at a specific confidence level are extracted, and the value indicators and risk indicators are packaged together into the final full life cycle return calculation result report.

Images