Hydropower station project group dynamic investment decision-making method considering uncertainty

Abstract

The invention provides a hydropower station project group dynamic investment decision-making method considering uncertainty, and relates to the technical field of hydropower project investment decision-making, and the method comprises the steps: S1, constructing a dynamic element perception and iteration engine, and obtaining a basic decision-making data source of a hydropower project group according to the dynamic element perception and iteration engine; extracting a plurality of decision-making elements from the basic decision-making data source, and dividing the plurality of decision-making elements into static basic elements and dynamic sensitive elements; according to the dynamic sensitive elements, through a real-time data and high-frequency feature analysis technology, an evaluation threshold interval is dynamically calculated and corrected, and a real-time iterative dynamic element library is formed. A multi-dimensional static evaluation system fused with expert experience can be upgraded into a decision support system capable of embedding a key uncertainty dynamic evolution model and performing adaptive simulation and optimization according to time sequence information input, so that a dynamic portfolio strategy with higher anti-risk capability is output.

Description

Technical Field

[0001] This invention relates to the field of hydropower project investment decision-making technology, and in particular to a dynamic investment decision-making method for a group of hydropower station projects that takes into account uncertainty. Background Technology

[0002] Hydropower projects, as crucial strategic investment targets for power generation companies, present a complex multi-attribute, multi-objective investment decision-making problem, involving comprehensive evaluation across multiple dimensions such as policy, technology, economics, environment, and society. For single-project decisions, the industry typically relies on detailed feasibility studies. However, when investment entities (such as large power generation groups) face multiple undeveloped integrated energy projects (hydropower, wind power, photovoltaic, etc.) within a river basin or region, scientifically determining implementation priorities and investment portfolios from numerous candidates to maximize overall benefits and minimize risks under limited funding, resources, and time constraints becomes a core challenge in engineering decision-making practice. Traditional single-project reviews and simple experience-based rankings are insufficient for systematically quantifying the advantages and disadvantages and dynamic relationships between projects. A systematic, quantifiable, and future-adaptable decision support methodology is urgently needed.

[0003] To address the aforementioned problem of prioritizing multiple projects, existing technologies have proposed several solutions based on Multi-Criterion Decision Models (MCDM). For example, the paper "Application Research of the Combined Weighting TOPSIS Method for Hydropower Station Decision Models" (Yang Yi, *China Hydropower and Electrification*, 2016) provides a decision-making method for selecting the optimal development scheme for hydropower station cascade development. This scheme constructs an evaluation index system covering dimensions such as economy, ecology, inundation, and land occupation, and employs a combination of "combined weighting" (combining subjective and objective weights) and the "Top-Approximation-to-Ideal-Solution Ranking Method (TOPSIS)" to calculate the closeness of each scheme to the ideal solution, thereby selecting the optimal development scheme.

[0004] However, such existing models have the following inherent flaws when dealing with real-world, long-term investment decisions for hydropower projects, resulting in insufficient robustness of their decision-making outcomes in dynamically changing real-world environments: The models described in the aforementioned literature are essentially static, one-off "photographic" assessments. The parameters they rely on, such as runoff, electricity prices, and investment costs, all use fixed values ​​or historical averages from the feasibility study phase, and the weights are determined in a single step. This ignores the key uncertainties that hydropower projects inevitably face throughout their entire lifecycle (up to decades), such as fluctuations in water flow, power market reforms, and policy adjustments. Using past deterministic data to plan future dynamic development is like "marking the boat to find the sword"—when the actual operating environment deviates from the preset scenario, the original "optimal" decision may quickly become invalid, resulting in poor risk resistance. This approach typically scores and ranks individual schemes or projects in isolation. It assumes that projects are independent of each other, and the ranking result is the optimal investment order. However, in reality, there are strong correlations between hydropower, electricity, and ecology among cascade hydropower stations in a river basin, and constraints such as funding competition and mutually exclusive construction resources exist between projects. Simple independent ranking cannot solve the "portfolio optimization" problem under overall budget and resource constraints. It may select a group of projects with high individual scores but conflicting benefits or concentrated risks when combined, failing to achieve the systemic synergy of "1+1>2". As mentioned above, the output of this model is a static priority list. It lacks a feedback and optimization mechanism that can continuously absorb new information such as updated climate forecasts, the latest policy documents, and real-time construction cost data, and adaptively adjust decisions. Once a decision is made, the model becomes ineffective, unable to continuously optimize subsequent investment pace and strategies and provide risk warnings based on new circumstances at different stages of project progress (such as planning, construction, and early operation).

[0005] In summary, the traditional multi-criteria ranking model, which relies on deterministic assumptions and static comparisons, is fundamentally incompatible with the complex, dynamic, uncertain, and multi-project coupled environment faced by hydropower project investment. Therefore, a method is needed to upgrade the static evaluation system into a dynamic decision-making system, thereby outputting more adaptable and resilient dynamic investment strategies. Summary of the Invention

[0006] To address the shortcomings of the existing technologies, the technical problem to be solved by this invention is to provide a dynamic investment decision-making method for hydropower station project groups that considers uncertainty. This method can upgrade a multi-dimensional static evaluation system that integrates expert experience into a decision support system that can embed a dynamic evolution model of key uncertainties and can perform adaptive simulation and optimization based on time-series information input, thereby outputting a dynamic investment portfolio strategy with stronger risk resistance.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: The present invention provides a dynamic investment decision-making method for a group of hydropower station projects considering uncertainty, which includes the following steps: S1. Construct a dynamic element perception and iteration engine to obtain the basic decision data source of the hydropower project group based on the dynamic element perception and iteration engine; extract several decision elements from the basic decision data source and classify the several decision elements into static basic elements and dynamic sensitive elements; dynamically calculate and correct the evaluation threshold range of the dynamic sensitive elements through real-time data and high-frequency feature analysis technology to form a dynamic element library that iterates in real time. S2. Construct a project cluster coupling network based on the dynamic element database; quantify the hydraulic, power, and resource coupling correlation between projects by analyzing the synchronization stability, dynamic characteristics, and interaction information of the project cluster coupling network; and generate project synergy benefit scores by combining the ecological and policy synergy dimensions. S3. Introduce risk value measurement based on synergistic benefit score, and construct a multi-objective optimization model with the goals of maximizing comprehensive benefits and minimizing risk concentration; solve the optimal hydropower station project portfolio scheme through optimization algorithm; S4. Set up monitoring nodes in stages throughout the investment process, collect data on the implementation of the plan and compare it with the expected value to generate a deviation diagnosis report; generate targeted investment strategy improvement plans based on the deviation diagnosis report and feed them back to the dynamic element perception and iteration engine of S1 to trigger threshold recalibration and form a decision-making closed loop.

[0008] In the preferred embodiment, the specific steps of step S1 are as follows: S11. Construct a dynamic element perception and iteration engine, which includes a data source docking module, a parameter calculation module, and a threshold correction module. The data source docking module is used to dock with the watershed hydrological monitoring station, the power trading platform, and the national energy policy database. The parameter calculation module is used to perform quantitative calculations of dynamic parameters. The threshold correction module is used to update the threshold range of the element scenario. S12. Obtain the basic decision-making data source for the river basin hydropower project cluster through the data source docking module of the dynamic element perception and iteration engine. The basic decision-making data source includes industry hydropower project evaluation standards, feasibility study reports of each candidate project, national energy policy documents, historical hydrological monitoring data of the river basin, real-time hydrological forecast data, historical electricity market transaction data, and real-time electricity price data. Extract several decision-making elements from the basic decision-making data source; divide the several decision-making elements into static basic elements and dynamic sensitive elements; static basic elements include topographic / geological conditions and key project structural parameters; dynamic sensitive elements include local government support policies, grid-connected electricity price, water inflow forecast value, and ecological red line range. S13. Obtain historical water inflow monitoring data and current water inflow forecast data through the parameter calculation module of the dynamic element perception and iteration engine, using the formula: ; The current water inflow forecast represents the watershed inflow forecast data released by the hydrological monitoring station for the current period; the historical average water inflow for the same period represents the average watershed inflow for the same period over the past 10 years, and the water inflow forecast deviation coefficient is calculated; simultaneously, historical electricity price data and real-time electricity price data are obtained, and the results are calculated using the formula: ; Among them, the real-time electricity price is the on-grid electricity price data released in real time by the power trading platform; the historical average electricity price for the same period is the average on-grid electricity price for the same period in the past 10 years, and the electricity price fluctuation coefficient is calculated. S14. Obtain multi-frequency data sources of dynamic sensitive elements. The multi-frequency data sources include daily on-grid electricity price data, monthly water inflow forecast data, and quarterly policy adjustment information. Input the multi-frequency data sources into the pre-trained MIDAS-QR analysis model and extract high-frequency feature coefficients through the frequency alignment algorithm and Beta multinomial weight constraint function of the model. S15. Combining the high-frequency characteristic coefficients, water inflow prediction deviation coefficients, and electricity price fluctuation coefficients obtained in the previous steps, the formula is used: ; in, This is represented as the modified dynamic sensitive element contextual threshold; This is represented as the initial baseline threshold for dynamically sensitive elements; , Represented as the watershed fit coefficient; It is represented as the Beta weight function adaptation coefficient, which corrects the scenario-based threshold range of dynamic sensitive elements. α and β are the watershed adaptation coefficients, and γ is the Beta weight function adaptation coefficient. α and β are calibrated by cross-disciplinary experts in conjunction with historical data from watershed hydropower projects over the past 10 years. S16. Connect to the National Energy Administration's policy database, river basin hydrological monitoring stations, and real-time data from the power trading platform via the engine's data source interface module, and apply the corrected dynamic sensitive element scenario-based thresholds. Real-time data is input into a pre-trained deep reinforcement learning (DQN) model, and action selection uses... The -greedy strategy verifies the validity of data using a Markov decision algorithm based on a pre-trained deep reinforcement learning (DQN) model, updates the dynamic feature database with the verified real-time data, and completes the real-time iteration of the feature database. In the preferred embodiment, the specific steps of step S2 are as follows: S21. Extract individual project time series data from the dynamic element database for each candidate hydropower project, including hydraulic time series data, power time series data, construction resource time series data, and cost time series data; input the individual project time series data into a pre-trained coupled time series network mapping model; divide the time series data into five levels of amplitude segments using an amplitude quantization algorithm; identify time series synchronicity using Pearson correlation coefficient; and construct a project group coupled network with quantized amplitude segments as nodes and time series synchronicity as links. S22. Input the coupled network into the pre-trained moment analysis model and spectral entropy rate analysis model. Through the moment analysis model, input the nonlinear dynamic parameters of the coupled network to identify the network synchronous bifurcation critical point. Then, calculate the single-item entropy rate through the hierarchical interactive analysis algorithm of the spectral entropy rate analysis model. Quantify the single-item dynamic characteristics, item pair link strength, and redundancy / cooperative coupling degree of item group hyperlinks obtained from the single-item time series data respectively. Finally, obtain three types of coupling correlation indicators: hydraulic coupling degree, power coupling degree, and resource coupling degree. S23. Obtain watershed ecological protection planning documents, watershed ecological monitoring data, and local energy policy documents through the data source docking module. Input the three types of coupled correlation indicators and the aforementioned ecological and policy data into the pre-trained multi-objective programming model of interaction effects. Based on the two-project / three-project interaction effect quantification system of the multi-objective programming model of interaction effects, supplement the dimensions of ecological synergy and policy synergy effects. Use the coupling correlation degree as the correction coefficient for the basic benefit score of the project to generate the project synergy benefit score. The objective function formula is: ;in, This is represented as the total score of the collaborative benefits of the project group; This is represented as the basic benefit score of the i-th item; Let be the synergistic gain coefficient of the i-th item; Represented as risk weight; This is represented as the risk concentration of the project group.

[0009] In the preferred embodiment, step S3 consists of the following steps: S31. Obtain the total investment budget from the company's annual investment plan, the total amount of construction resources in the engineering resource database, and the ecological constraint ceiling in the watershed ecological protection plan; simultaneously, input historical electricity price fluctuation data and water inflow data into the pre-trained CVaR risk measurement model, and use the Monte Carlo simulation algorithm of the CVaR risk measurement model, combined with the formula... The CVaR risk index was calculated. in, Represented as conditional risk value; This is expressed as the confidence level; This is represented as the loss function of the portfolio; Represented as Value at Risk; It is represented as the probability density of the loss function; Let it be expressed as the probability density of the random variable y; S32. Input the project synergy benefit score obtained in step S23 and the total investment budget, total construction resources, ecological constraint upper limit, and CVaR risk index from step S31 into the pre-trained MOPSO model. Using the chaotic mutation search algorithm and elite individual retention algorithm of the pre-trained MOPSO model, solve the multi-objective problem of "maximizing comprehensive benefits, minimizing risk concentration, and optimizing synergy benefits," and output the optimal hydropower station project portfolio scheme. The formula for the chaotic mutation search algorithm is as follows: ; in, Represented as the particle position after a chaotic mutation; Represented as the particle position before the mutation; Represented as the variation control factor; Represented as a chaotic sequence value; ; in, Represented as the chaotic sequence value of the previous iteration step; Inertia variables in elite individual preservation algorithms As the number of iterations adaptively decreases from 0.9 to 0.2, the crowding degree is calculated using the following formula: ; in, This represents the crowding level of the i-th individual; Represented as the number of objective functions; This is represented by the value of the o-th objective function; , Let these represent the maximum and minimum values ​​of the o-th objective function, respectively.

[0010] In the preferred embodiment, step S4 consists of the following steps: S41. Divide the entire investment process into the planning, construction, and initial operation stages. Set up monitoring nodes in the planning, construction, and initial operation stages respectively. Collect execution data of the optimal hydropower station project group investment scheme through the project site monitoring system. Input the execution data into the pre-trained spectral entropy rate analysis model. Diagnose the source of decision deviation through the hierarchical interactive analysis algorithm of the spectral entropy rate analysis model and generate a deviation diagnosis report. The execution data includes actual construction costs, total water inflow, and market electricity prices; the sources of decision-making deviations include parameter fluctuation deviations, coupling imbalance deviations, and risk exceeding standards deviations. S42. Input the deviation diagnosis report into the pre-trained deep reinforcement learning DQN model, and generate a targeted investment strategy improvement plan by modeling the Markov decision process of the pre-trained deep reinforcement learning DQN model and combining the sequential decision logic of the DRL-ARMA algorithm. S43. Feedback the improved solution to the threshold correction module of the dynamic element perception and iteration engine built in S1. The threshold correction module calls the dynamic parameter threshold correction formula in S15 to recalibrate the scenario-based threshold range of dynamic sensitive elements, complete the threshold correction, and form a decision-making closed loop.

[0011] In the preferred embodiment, in step S1, the daily hydrological monitoring data of the basin extracted in S12 and the historical hydrological monitoring data of the basin for the past 10 years and the current hydrological station prediction data obtained in S13 are first acquired. Then, the multi-frequency data sources are input into the MIDAS-QR analysis model in S14. Frequency alignment is completed by the frequency multiple matching algorithm of the weekly revenue sequence and the daily electricity price / inflow characteristic variables built into the MIDAS-QR analysis model. At the same time, the Beta polynomial weight constraint function is called to perform weighted allocation of high-frequency features. The expression of the Beta polynomial weight constraint function is: ; in, This is represented as the weighting coefficient for the d-th lag order; Represented as the lag order; This is expressed as the maximum lag order of the j-th feature variable; , This is represented as the shape parameter of the Beta function, with values ​​of 1 and 1.4301 respectively; It is represented as the Beta probability density function.

[0012] The daily inflow fluctuation characteristic value weighted by the weighting function is extracted as the high-frequency characteristic coefficient.

[0013] In the preferred embodiment, in step S22, two nonlinear dynamic parameters of the project cluster coupled network in S21 are first obtained. These two nonlinear dynamic parameters are then input into the moment analysis model. The synchronization stability criterion calculation logic built into the moment analysis model identifies the bifurcation critical point of the network synchronization region and outputs the network synchronization stability criterion. Next, the time series data of single-project power generation extracted in S21 is obtained. This data is then input into the spectral entropy rate model. The dynamic characteristics of the single project are calculated using the spectral entropy rate formula built into the spectral entropy rate model. The spectral entropy rate formula is as follows: ; in, Represented as the spectral entropy rate of the i-th item; Represents the normalized angular frequency; This represents the dimension of the time series data for the i-th project; Let represent the determinant of the power spectral density matrix of the i-th project. Logarithmic operations and dimensional weighting are performed on the time-series data. Then, the power generation time-series data of the project pair are obtained, and the link strength of the project pair is calculated. Substituting this into the mutual information rate formula: ; in, This represents the mutual information rate between items i and j; This represents the determinant of the power spectral density matrix of the j-th item; The determinant of the joint power spectral density matrix of projects i and j is used to quantify the link strength of the project. Then, bootstrap data analysis is performed on the calculation results of hyperlink redundancy / cooperative coupling degree of the project group. Finally, the synchronization stability criterion is used as a correction factor to adjust the three types of coupling correlation indicators, and the corrected three types of coupling correlation indicators are obtained.

[0014] In the preferred scheme, historical electricity price fluctuation data and water inflow data are obtained and input into the CVaR risk measurement model. The CVaR risk index is calculated using the Monte Carlo simulation algorithm built into the CVaR risk measurement model, combined with the CVaR formula. Simultaneously, the total investment budget in the company's annual hydropower project investment plan, the total amount of construction resources in the engineering machinery and equipment and construction team resource pool, and the ecological constraint upper limit in the watershed ecological environment protection special plan are obtained. Then, the MOPSO model is called and the inertial variable iteration formula is configured. ;in, This represents the inertia variable in the k-th iteration; Represents the maximum inertial variable; Represents the smallest inertial variable; Indicates the current iteration number; This represents the maximum number of iterations, causing the inertia variable to adaptively decrease from 0.9 to 0.2 with the number of iterations; then, the risk-benefit tradeoff is achieved by combining the mean-CVaR utility function, as shown in the formula: ; in, This represents the investment utility value during time period t; This represents the expected return during time period t; Indicates the confidence level for time period t. Conditional Value at Risk (VaR); Indicates the number of simulation scenarios; This represents the revenue during time period t in the l-th scenario; This represents the expected loss corresponding to CVaR, and a weighted integration of expected returns and risk indicators is performed.

[0015] In the preferred scheme, real-time time-series data of hydraulic, electric, and construction resources from the dynamic element library in S1 are obtained and input into the coupled time series network mapping model. The amplitude of the time-series data is divided into five levels of amplitude segments using the amplitude quantization algorithm built into the coupled time series network mapping model. Then, the synchronicity identification algorithm is called and the Pearson correlation coefficient calculation formula is configured. ; in, This represents the Pearson correlation coefficient; , This represents the t-th sample value of two time series data sets; , This represents the sample mean of two time series data sets. The number of samples is represented by a threshold of 0.8 to identify temporal synchronicity. Then, a coupled network of items is constructed with quantization amplitude segments as nodes, temporal synchronicity as links, and the frequency of synchronicity occurrence as the link weight.

[0016] In the preferred embodiment, in the multi-objective programming model of interaction effects in step S23, the built-in objective function of the multi-objective programming model of interaction effects is called, and the formula is: ; in, This represents the total score for the synergistic benefits of the project group; This represents the basic benefit score of the i-th project; This represents the synergistic gain coefficient of the i-th item; This represents the risk weight and its value ranges from [0.1, 0.3]. This represents the risk concentration of a project group. The basic benefit score of the project is multiplied and summed with the synergy gain coefficient, and then the risk concentration is weighted and deducted. Ecological collaboration is obtained based on watershed ecological monitoring data, and policy collaboration is obtained based on local energy policy documents.

[0017] In the preferred embodiment, the dynamic sensitive element scene-based threshold modified in S15 is used. The execution data of the optimal hydropower station project portfolio scheme collected by the project site monitoring system in S41 is input into the spectral entropy rate analysis model. The source of decision deviation is diagnosed through the deviation calculation formula built into the spectral entropy rate analysis model. ; in, Indicates the parameter deviation value; This represents the actual monitored value of the parameter; This represents the threshold after parameter correction; The sources of diagnostic decision-making bias include deviations from water / electricity prices and thresholds, deviations from expected values ​​in inter-project synergy, and deviations from preset thresholds in CVaR. Combined with the synchronization stability criteria output by the moment analysis model, the root causes of coupling imbalance are determined, and a deviation diagnosis report is finally generated.

[0018] In the preferred scheme, in step S15, historical data of regional hydropower projects over the past 10 years are first collected, and 13 types of cross-disciplinary experts are organized to calibrate the basin adaptation coefficients α and β to obtain the calibrated coefficient values; then the coefficient of variation formula is called: ;in, Indicates the coefficient of variation; This represents the standard deviation of the coefficient calibration values; This represents the mean of the coefficient calibration values; the calibration results are validated and the coefficient of variation threshold is set to ≤0.1; Simultaneously, the parameter verification logic of the Beta weight function is introduced to match the calibration coefficients with the high-frequency feature coefficients. The verified coefficients are then substituted into the dynamic parameter threshold correction formula of S15 to complete the correction of the dynamic element scenario-based threshold range.

[0019] In the preferred embodiment, the targeted investment strategy improvement scheme generated by the deep reinforcement learning (DQN) model in step S42 includes a project investment timing adjustment table, a resource allocation optimization table, and a risk mitigation measure list. This targeted investment strategy improvement scheme needs to be fed back to the threshold correction module of the dynamic element perception and iteration engine in S1, triggering the recalibration of the dynamic parameter threshold in S15. The recalibrated threshold is used for the next round of dynamic element iteration. The resource adjustment problem is transformed into a Markov decision process using the DRL-ARMA algorithm built into the deep reinforcement learning (DQN) model. A scheme is generated through iterative optimization of the state space and action space, and the state transition probability formula is configured. ; in, Indicates the state transition probability; Indicates the current state; Indicates the state at the next moment; Indicates the action to be performed; Indicates the state at time t; This represents the action at time t.

[0020] In the preferred embodiment, the O-information rate calculation of the spectral entropy rate analysis model in step S22 uses an iterative formula: ; in, Represents the O-information rate of N items; Represents the O-information rate of N-1 items; Let O-information rate gradient represent the gradient of the newly added Nth item, and the formula is as follows: ; in, This represents the mutual information rate between the Nth item and the sub-item group after removing the i-th item; This represents the mutual information rate between the Nth project and the original N-1 project groups. This formula quantifies the redundancy / cooperative coupling degree of project group hyperlinks, and the calculation results need to be verified for statistical significance through Bootstrap data analysis.

[0021] In the preferred embodiment, the chaotic mutation search mechanism of the MOPSO model in step S31 adopts the following formula: and Generate chaotic sequences; in, Indicates the particle position after the mutation; Indicates the original particle position; Indicates the variation control factor; Represents chaotic sequence values; This represents the chaotic sequence value from the previous iteration step; And variation control factor ;in, Indicates the current iteration number; Indicates the maximum number of iterations; The elite individual retention mechanism is calculated based on crowding: Select the optimal individual to ensure the diversity and convergence of non-dominated solutions; in, This represents the crowding level of the i-th individual; Indicates the number of objective functions; This represents the o-th objective function value; , Let represent the maximum and minimum values ​​of the o-th objective function.

[0022] This invention provides a dynamic investment decision-making method for a group of hydropower station projects that takes into account uncertainty. Through the coordination of the above-mentioned structures, it has the following advantages compared to existing methods: First, it overcomes the shortcomings of traditional static assessments, which are often based on outdated methods, and significantly improves the resilience of decision-making. This invention constructs a dynamic element library with quantifiable uncertainty, transforming key uncertainties throughout the entire lifecycle, such as hydrological water inflow, electricity market, and policy adjustments, into quantifiable risk labels. Furthermore, it relies on the MIDAS-QR model to achieve time-series alignment and dynamic updates of multi-frequency data, no longer depending on fixed historical averages or static parameters from the feasibility study stage. This allows it to adapt to environmental changes over the decades-long lifecycle of hydropower projects, avoiding decision failures when actual operating scenarios deviate from preset parameters, and fundamentally enhancing the robustness of decision-making.

[0023] Secondly, it breaks through the limitations of isolated project ranking and maximizes the synergistic benefits of the project portfolio. This invention reconstructs the coupling relationships between hydraulic, power, and resource aspects of projects using a coupled time-series network mapping model. It corrects synchronization bifurcation points through a moment analysis model, completes hierarchical interaction quantification using a spectral entropy rate model, and achieves risk-weighted scoring of synergistic benefits using a multi-objective programming model with interaction effects. Simultaneously, it optimizes the portfolio using a MOPSO model with CVaR constraints. Instead of treating projects as independent entities for simple ranking, it avoids conflicts of benefit or concentration of risk among projects under limited funding and resource constraints, fully releasing the systemic synergistic value of watershed cascade projects.

[0024] Third, a closed-loop optimization mechanism for the entire decision-making process has been established, enabling dynamic rolling adjustments and risk warnings. The DRL-ARMA model has been introduced to dynamically correct decision-making biases, and a feedback update link between the element library and the pre-trained model has been set up. It is no longer a one-off model that terminates upon decision-making. It can continuously absorb new information such as climate predictions, policy documents, and construction costs at different stages of project planning, construction, and operation, adaptively optimize investment pace and strategies, and provide accurate warnings for decision-making biases, ensuring that decisions always align with the real-time environment. Attached Figure Description

[0025] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is the main view structure diagram of the process of this invention. Detailed Implementation

[0026] To better understand the purpose, system architecture, and functional implementation of this embodiment, the embodiments and features described herein can be combined with each other without conflict. The exemplary embodiments disclosed herein will be described below with reference to the accompanying drawings, including specific technical details disclosed to aid understanding; however, these details should be considered exemplary rather than restrictive. Therefore, those skilled in the art should understand that various improvements and adjustments can be made to the embodiments described herein without departing from the scope and core ideas of the invention. Similarly, for clarity, detailed descriptions of well-known technologies, functions, and structures are omitted in the following description.

[0027] Example 1 This embodiment addresses the dynamic investment decision-making scenario of a cascade hydropower station project group in the Lancang River Basin of Southwest my country. It completes the full-process training and deployment of the MIDAS-QR analysis model, coupled time series network mapping model, moment analysis model, interactive effect multi-objective programming model, spectral entropy rate analysis model, MOPSO optimization model, and deep reinforcement learning model, realizing the practical application of technology from dynamic element perception to closed-loop optimization of decision-making. The data used in this embodiment covers seven types of data sources, including the National Energy Administration policy database, basin hydrological monitoring stations, and the Yunnan power trading platform. It also integrates historical hydrological data of the basin from 1995 to 2024, Yunnan power market transaction data from 2018 to 2024, and feasibility study reports of 12 candidate hydropower station projects, providing sufficient sample support for model training. Furthermore, all model training processes meet the "sufficient disclosure" requirement of the Patent Law and can be directly deployed and executed on a computer system.

[0028] In the preferred scheme, the core function of the MIDAS-QR analysis model is the extraction of high-frequency features from multiple frequency data sources. Its training process needs to complete data preprocessing, model structure building, parameter calibration and validity verification in sequence to ensure that dynamic sensitive elements of different frequencies are accurately fused.

[0029] In this embodiment, the multi-frequency data sources for model input include three categories: first, daily on-grid electricity price data from the Yunnan Power Trading Platform from 2018 to 2024 (a total of 2,555 samples, with a sampling frequency of 1 day / sample); second, monthly inflow forecast data from 120 hydrological monitoring stations in the Lancang River Basin from 2018 to 2024 (a total of 84 monthly samples, with a sampling frequency of 1 month / sample); and third, quarterly hydropower development policy documents from the National Energy Administration from 2018 to 2024 (a total of 28 quarterly documents, with a sampling frequency of 1 quarter / sample).

[0030] Specifically, the data preprocessing stage requires standardization and outlier handling of heterogeneous data: for daily electricity price data, the 3σ principle is used to remove 12 abnormal samples caused by sudden market fluctuations to ensure that the data distribution meets the requirements of normality; for monthly water inflow data, linear interpolation is used to complete 5 missing monthly data from 3 remote monitoring points to ensure the continuity of time series data; for quarterly policy documents, qualitative indicators are used for quantitative coding to convert policy clauses such as "subsidy intensity" and "ecological constraint level" into quantitative scores of 0-10, thereby realizing the numericalization of unstructured policy data.

[0031] The model structure comprises four core modules: an input layer, a frequency adaptation layer, a weight constraint layer, and a quantile regression output layer. The input layer has three dimensions, corresponding to daily electricity prices, monthly water inflows, and quarterly policy quantification values. The frequency adaptation layer employs a frequency-difference matching algorithm between weekly revenue sequences and daily feature variables, upsampling daily electricity price data in 7-day weekly windows to generate 365 weekly electricity price samples. This achieves time-scale unification for daily, monthly, and quarterly data, resolving the temporal misalignment problem of multi-frequency data. The weight constraint layer embeds a Beta multinomial weight constraint function, expressed as: ; in, Represented as the first The weighting coefficients corresponding to each lag order; Represented as the lag order; Represented as the first The maximum lag order of each feature variable; , Represented as the shape parameter of the Beta function; It is represented as the Beta probability density function.

[0032] In this embodiment, the shape parameter of the Beta function The value is 1. The value was set to 1.4301. This parameter was calibrated by 13 interdisciplinary experts from the fields of hydrology, power market, and engineering economics, using historical data from nearly 10 years of hydropower projects in the basin. The calibration results were verified by the coefficient of variation (≤0.1) to ensure the parameter's suitability for high-frequency characteristics. The confidence level β of the quantile regression output layer was set to 0.95, and the optimization objective was to minimize the quantile loss function. , in, Represented as quantile loss function value; This represents the input deviation value for loss calculation; This represents the confidence level for quantile regression; Represented as an indicator function, when hour Otherwise, it is 0, thus achieving accurate extraction of high-frequency feature coefficients.

[0033] In practice, the preprocessed 365 weekly samples were divided into a training set (255 samples) and a validation set (110 samples) in a 7:3 ratio. The Nelder-Mead optimization algorithm was used for model training, with 1000 iterations and an initial learning rate of 0.01, which linearly decreased to 0.001 with each iteration. During training, the maximum lag order of each feature variable was determined using the GACV criterion: the maximum lag order of the daily electricity price. Maximum lag order of monthly water inflow The maximum lag order of quarterly policies , GACV stands for Generalized Approximate Cross-Validation Criterion, which is used to select the optimal lag order to balance model fit and generalization ability.

[0034] After training, the high-frequency feature coefficient extraction accuracy on the validation set reached 92.7%. Compared with the traditional MIDAS model, the mean square error of feature extraction was reduced by 18.3%. It can effectively output high-frequency features such as "fluctuation coefficient of water abundance and scarcity" and "time-varying characteristic coefficient of electricity price", providing a core basis for subsequent dynamic element threshold correction.

[0035] In this embodiment, the coupled time series network mapping model is used to build a coupled correlation network of hydropower station project groups. Its training process is divided into three interlocking stages: time series data amplitude quantization, synchronization identification, and network topology construction. The parameters of each stage are calibrated using actual project data.

[0036] In the preferred scheme, the model input consists of three types of time-series data from the dynamic element library: first, the hydraulic time-series data (inflow rate, 120 monitoring points, sampling frequency of 1 hour / data, total of 630,720 samples) for 12 candidate hydropower stations from 2018 to 2024; second, the power time-series data (power generation, sampling frequency of 15 minutes / data, total of 2,522,880 samples) for each power station; and third, the construction resource time-series data (equipment / team scheduling, sampling frequency of 1 day / data, total of 2,555 samples).

[0037] Specifically, the amplitude quantization stage of the time-series data uses five levels of amplitude quantization intervals to discretize the data: for inflow water flow, according to... , , , , The interval division is carried out by using the equal frequency binning method to ensure that the sample proportion of each level is controlled at 18%-22%; for power generation, it is divided into 5 quantitative levels according to the 20% gradient of the installed capacity of the power station; for construction resources, it is discretized according to the equipment idle rate intervals of 0-20%, 20%-40%, 40%-60%, 60%-80%, and 80%-100%, so as to realize the node-based transformation of continuous time series data.

[0038] In the synchronization identification stage, the Pearson correlation coefficient is used as the criterion, and its calculation formula is as follows: ;in, Expressed as the Pearson correlation coefficient; , Represented as the first of two time series data Each sample value; Represented as The sample mean of time series data; Represented as The sample mean of time series data; This represents the number of samples.

[0039] In this embodiment, the judgment threshold is set to 0.8, that is, when the correlation coefficient of the sliding window (window length 7 days) of two time series data is ≥0.8, it is determined that there is a synchronous correlation between the two. Preferably, the average sliding window correlation coefficient of the water flow time series data of upstream power station A and downstream power station B in the Lancang River cascade hydropower stations is 0.86, which is judged as a strong synchronous correlation; while the correlation coefficient of the construction resource time series data of power station A and power station C is 0.62, which is judged as no significant synchronous correlation.

[0040] In practice, during the network topology construction phase, each quantization amplitude segment is treated as a network node, and the synchronicity of the amplitudes of two time series is considered a network link, with the link weight being the frequency of synchronicity occurrence. During model training, an initial adjacency matrix is ​​first constructed based on sample data, and then the link weights are iteratively adjusted using a modular optimization algorithm. The number of iterations is set to 500, and the convergence condition is that the change in the cosine similarity of the adjacency matrix is ​​≤1e-5. The adjacency matrix represents the association between network nodes, with each element corresponding to a link weight. The cosine similarity represents the degree of similarity between two matrix vector spaces and is used to determine whether the iteration has converged.

[0041] After training, the coupled network output by the model clearly shows three types of coupling relationships: the network density for hydraulic coupling relationships is 0.78, the network density for power coupling relationships is 0.69, and the network density for resource coupling relationships is 0.42. Network density is represented by the ratio of the actual number of links in the network to the maximum possible number of links. It is used to characterize the degree of connection between network nodes, providing an accurate topological basis for subsequent coupling correlation quantification. Furthermore, the network can be directly used as input for the moment analysis model and the spectral entropy rate analysis model.

[0042] Specifically, the core function of the moment analysis model is to identify the critical point of synchronous bifurcation in the coupled network. Its training relies on the dynamic parameters of the coupled network to complete the construction of dynamic equations, the derivation and verification of synchronous stability criteria, and the criteria can be used as a correction factor for the coupling correlation degree to improve the subsequent quantization accuracy.

[0043] In this embodiment, the dynamic parameters input to the model are the upstream inflow regulation coefficient and the downstream power generation response coefficient of the coupled network. The parameter samples are from the operation and control records of the Lancang River cascade hydropower stations from 2018 to 2024 (a total of 126 parameter combinations). The upstream inflow regulation coefficient ranges from [0.1, 0.9] (characterizing the upstream reservoir's storage capacity), and the downstream power generation response coefficient ranges from [0.2, 1.0] (characterizing the downstream power station's power generation response sensitivity to upstream inflow).

[0044] In the preferred scheme, the core dynamic equations of the model are constructed based on nonlinear coupled differential equations, and the single-node dynamic equations are as follows: ; in, Represented as the first Power station nodes Rate of change of state at time t; It is represented as the self-dynamic function of the power station node, characterizing the relationship between power generation and inflow rate; Represented as global coupling strength; Represented as the Laplace matrix of the coupled network; It is represented as an inner coupling matrix (containing positive and negative parameters, which respectively represent the cooperative and competitive relationships between nodes). Represented as the first Power station nodes The state value at any given time; This represents the total number of network nodes.

[0045] The derivation of the synchronization stability criterion requires first solving the Jacobian matrix of the dynamic equations, and then calculating the real parts of the matrix eigenvalues: when all real parts of the eigenvalues ​​are less than 0, the network is considered to be in a synchronously stable state; when the real parts of the eigenvalues ​​turn from negative to positive, the corresponding parameter critical point is the synchronization bifurcation point. The Jacobian matrix is ​​represented by the first-order partial derivative matrix of the dynamic equation, which is used to analyze the stability of the system near the equilibrium point; the real part of the eigenvalues ​​is represented by the real part of the eigenvalues ​​of the Jacobian matrix, and its sign determines the stable state of the system.

[0046] During training, 126 sets of parameter combinations were substituted into the model one by one, and the eigenvalues ​​of the Jacobian matrix were solved through numerical simulation. The synchronous stability intervals under different parameter combinations were statistically analyzed. Preferably, when the upstream water inflow regulation coefficient = 0.6 and the downstream power generation response coefficient = 0.7, the real parts of all eigenvalues ​​of the Jacobian matrix are -0.23 to -0.05, and the network is in a stable state. When the upstream water inflow regulation coefficient rises to 0.8 and the downstream power generation response coefficient = 0.9, the real part of the eigenvalues ​​becomes positive, with a value of 0.02, triggering a synchronous bifurcation.

[0047] In practice, the model validation phase used 10 sets of parameter samples added in 2024 that were not used in the training. The prediction error of their synchronization bifurcation points was ≤5%, which is 22% lower than the traditional linear stability analysis method. The coupling correlation degree after the model correction improved the fit with the actual running data by 15%, providing a reliable basis for correction of the subsequent hierarchical interactive quantization of the spectral entropy rate analysis model.

[0048] In one feasible approach, the spectral entropy rate analysis model is used to quantify the individual project dynamics of a project group, the link strength of project pairs, and the hyperlink redundancy / cooperative coupling of the project group. Its training process covers three core steps: power spectral density (PSD) matrix estimation, three types of information rate calculation, and statistical significance verification.

[0049] In this embodiment, the model input is the coupled network time series data output by the coupled time series network mapping model. During the data preprocessing stage, both the weighted covariance (WC) method and the vector autoregression (VAR) model are used to estimate the PSD matrix. The nonparametric PSD estimation employs the Parzen window (bandwidth...) The formula for calculating the cross PSD matrix is: ; in, Represented as the first , Cross power spectral density of time series data for each project; This is represented as the maximum lag order (set to 50). Represented as a Parzen window function; Represented as the first , Time series data of each project Cross-autocorrelation coefficient; Represented as normalized angular frequency; Represented as the imaginary unit, .

[0050] In the formula This is the maximum lag order (set to 50). The Parzen window function is used; the PSD parameter is estimated using a VAR model, and the model order is... The value is determined to be 10 using the AIC criterion, and the VAR equation is: , in, Represented as The time-series data vector at each moment; Represented as The coefficient matrix of the VAR model of order 1; Represented as The white noise perturbation term at time step 1; the AIC criterion, denoted as the Akaike Information Criterion, is used to select the optimal order of the VAR model, and then the PSD matrix is ​​obtained through spectral decomposition. , in, Represented as the frequency response matrix of the VAR model; It is represented as the covariance matrix of the white noise perturbation term; Represented as The conjugate transpose of .

[0051] Specifically, the calculation of the three information rates relies on the following formulas: Single-item spectral entropy rate: ; in, Represented as the first Each item at angular frequency Spectral entropy rate under the following conditions; Represented as the first Dimensions of time-series data for each project; Represented as the first Determinant of the power spectral density matrix of time series data for each project; It is expressed as a natural constant, with a value of approximately 2.71828.

[0052] In this embodiment, the spectral entropy rate of upstream power station A is 0.83 nats / Hz in the low-frequency band of 0.04-0.15Hz and 0.56 nats / Hz in the high-frequency band of 0.15-0.4Hz, indicating that its power generation stability is lower in the low-frequency band.

[0053] Project mutual information rate: ; in, Represented as a project With the project At angular frequency Mutual information rate at the following levels; Represented as the first Determinant of the power spectral density matrix of time series data for each project; Represented as a project With the project The determinant of the power spectral density matrix of the joint time series data.

[0054] In this embodiment, the mutual information rate of the AB power station pair is 0.32 nats / Hz in the low-frequency band and 0.21 nats / Hz in the high-frequency band, indicating that the hydraulic complementarity between the two is stronger in the low-frequency band.

[0055] Project team O-information rate: ; in, Represented as Each item at angular frequency O-information rate; Represented as Each item at angular frequency O-information rate; This indicates the newly added number. The gradient value of O-information rate after each project.

[0056] and ; in, Represented as the first The project and the elimination of the first The sub-project teams of each project are at angular frequency Mutual information rate at the following levels; Represented as the first The project and the original Each project team at angular frequency The mutual information rate.

[0057] In this embodiment, the O-information rate of the project group consisting of 3 power plants is -0.18 nats / Hz (negative value), which is determined to be cooperative coupling; if the O-information rate is positive, it is determined to be redundant coupling.

[0058] In the preferred scheme, statistical significance verification uses bootstrap data analysis, with a number of iterations. When verifying A value < 0.05 indicates that the calculation result is significant. Among them, bootstrap data analysis refers to a method of generating multiple sample sets through resampling to verify the significance of statistics; The value is used as an indicator to determine the significance of the result in statistical hypothesis testing.

[0059] After training, the model's performance error on the validation set is ≤4%, and it can achieve frequency-specific interaction quantization. Compared with traditional time-domain information theory methods, its recognition accuracy for high-frequency interaction features is improved by 30%.

[0060] In this embodiment, the multi-objective programming model of interaction effect is used to calculate the synergistic benefit score of the project group. Its training needs to complete three key steps: objective function construction, constraint setting, and synergistic gain coefficient calibration. In addition, the model needs to integrate indicators of ecological and policy synergy dimensions.

[0061] Specifically, the objective function of the model is:

[0062] in, This is represented as the total score of the collaborative benefits of the project group; Represented as the first Basic benefit score for each project; Represented as the first Synergistic gain coefficient of each project; This is represented as a risk weight, with a value range of [0.1, 0.3]. This is expressed as the risk concentration of the project group; This represents the total number of projects within the project group.

[0063] In the formula For the first Basic benefit scores for each project (from the feasibility study report; scores for the 12 candidate projects range from 65 to 92 points). The coefficient of synergistic gain is calibrated based on a linear mapping of coupling correlation (coupling correlation 0 corresponds to...). Coupling degree 1 corresponds to (This is followed by a separate section on ecological and policy coordination, which is not directly related to the preceding sentence about water quality compliance rates.) If the vegetation coverage growth rate is ≥5%, then Provincial and national dual subsidies If the regional energy plan is fully aligned, then ; The risk weight, determined by 13 categories of experts, is 0.2. This represents the CVaR risk concentration, derived from the output of the subsequent CVaR risk measurement model.

[0064] In the preferred scheme, the model's constraints include three categories: first, total investment budget constraint (annual hydropower investment budget ≤ 5 billion yuan); second, total construction resource constraint (number of construction equipment ≤ 200 units, construction teams ≤ 30 groups); and third, ecological constraint upper limit (basin ecological flow guarantee rate ≥ 90%). During training, a linear weighted method is used to transform the multi-objective problem into a single objective, iteratively optimizing the objective function value. The number of iterations is set to 300, and the convergence condition is that the change in the objective function value ≤ 1e-4. Among them, the linear weighted method is a method of assigning weights to different objective functions and summing them up to transform them into a single objective, which is used to simplify the solution of multi-objective optimization problems.

[0065] In practice, the model training used actual operational data from 12 candidate projects from 2018 to 2024. After training, the model output synergy benefit score matched the actual data by 89%. Compared with the traditional single-objective programming model, the overall benefit of the project group was improved by 12%, the risk concentration was reduced by 8%, and the indicators of ecological and policy synergy made the score more in line with the policy orientation of regional hydropower development.

[0066] Specifically, the MOPSO optimization model is used to solve the multi-objective optimal solution of the investment portfolio of hydropower station projects. Its training process covers four core steps: particle initialization, chaotic mutation search, elite individual retention, and dynamic file update. The model also introduces a chaotic mutation mechanism to improve local search capabilities.

[0067] In this embodiment, the model input includes the synergistic benefit score output by the multi-objective programming model of interaction effect, the enterprise's annual investment budget, the total amount of construction resources, the upper limit of ecological constraints, and the CVaR risk index. The optimization objective is "maximizing comprehensive benefits, minimizing risk concentration, and optimizing synergistic benefits". The particle dimension is 12 (corresponding to the investment ratio of 12 alternative projects), the population size is set to 100, and the maximum number of iterations is 500.

[0068] In the preferred scheme, during the particle initialization phase, the position vector of each particle corresponds to the investment ratio of 12 projects (the sum of the ratios ≤ 1), the initial value of the velocity vector is a random number in the interval [-0.05, 0.05], and the inertial variable... As the number of iterations adaptively decreases from 0.9 to 0.2, the iteration formula is: ; in, Represented as the first The inertial variables of the next iteration; This is represented as the maximum inertial variable, with a value of 0.9; It is represented as the smallest inertial variable, with a value of 0.2; This represents the current iteration number; This represents the maximum number of iterations.

[0069] The chaotic mutation search mechanism uses the following formula: ; in, This represents the position of the particle after the mutation; Represented as the position of the original particle; Represented as the variation control factor; It is represented as a chaotic sequence value. ;in, This represents the chaotic sequence value from the previous iteration. The variation control factor is shown in the formula. ,in, Represented by the maximum number of iterations, chaotic sequence This mechanism can effectively prevent particles from getting trapped in local optima.

[0070] Among them, the elite individual retention mechanism selects the best individuals through crowding degree calculation, and the crowding degree formula is: ;in, Represented as the first The degree of crowding of individuals; This is expressed as the number of objective functions (3 in this example); Represented as the first The values ​​of the objective function; Represented as the first The maximum value of each objective function; Represented as the first The minimum value of an objective function.

[0071] During training, the top 30% of crowded individuals in each generation are retained as elite individuals to ensure the diversity of non-dominated solutions. A dynamic archive update mechanism stores non-dominated solutions from each generation in an external archive. When the archive size exceeds 200, individuals with low crowding are removed to maintain the archive's validity. Here, a non-dominated solution is defined as a solution for which no other solution is superior in all objectives; the external archive is defined as a dataset that stores all non-dominated solutions during the iteration process.

[0072] In practice, the model training uses 100 sets of electricity market price scenarios (with a confidence level of 95%) generated by Monte Carlo simulation. Monte Carlo simulation is a method of generating a large number of scenarios through random sampling to simulate uncertainty. The confidence level of 95% indicates that the simulation results are 95% credible. In the set of non-dominated solutions output after training, the optimal solution achieves a comprehensive benefit of 1.073 billion yuan, a risk concentration of 0.12, and a synergistic benefit score of 91 points. Compared with the traditional PSO model, the diversity of non-dominated solutions is increased by 25%, the convergence speed is accelerated by 18%, and it can meet the needs of different decision-making scenarios.

[0073] In this embodiment, the deep reinforcement learning model includes the DQN model (data validity verification) and the DRL-ARMA model (investment strategy improvement). Both models are based on Markov decision process (MDP) modeling to achieve closed-loop support from data verification to decision optimization.

[0074] In this embodiment, the DQN model is used to verify the validity of real-time data sources. Its training requires first defining the three elements of MDP, then building the network structure and performing iterative optimization.

[0075] Specifically, the MDP state space It includes three dimensions: data integrity (0-1 variables), data timeliness (difference between collection time and current time, in hours), and format standardization (0-1 variables); action space. It includes three discrete actions: "Validation passed", "Validation failed and needs to be completed", and "Validation failed and needs to be removed". The reward function is defined as follows: +10 reward for validation, +5 reward for passing after completion, and -5 reward for removing invalid data, ensuring the quality of the data source for the dynamic feature database. Here, MDP stands for Markov Decision Process, which is a mathematical framework for describing sequential decision problems; state space Represented as the set of environmental states that the agent can perceive; action space It represents the set of actions that the agent can perform; the reward function represents the feedback obtained by the agent after performing an action, which is used to guide policy learning.

[0076] In the preferred scheme, the DQN network is a 3-layer fully connected structure with an input layer dimension of 3, a hidden layer dimension of 64 (ReLU activation), and an output layer dimension of 3 (outputting the Q value of each action). Here, the Q-value represents the long-term cumulative reward expectation of an agent performing an action in a certain state; ReLU activation represents the modified linear unit activation function, used to enhance the nonlinear fitting ability of the network.

[0077] The model training experience replay pool capacity is set to 10,000, the target network synchronization period is 100 steps, and action selection adopts... -greedy strategy, The value gradually decreases from 0.9 to 0.1, using the following formula: ; in, Represented as the first The exploration probability of the next iteration; This represents the initial exploration probability, with a value of 0.9; This represents the final exploration probability, with a value of 0.1; This represents the current iteration number; This represents the maximum number of iterations.

[0078] The training data consisted of 15,600 real-time data validation records from 2018 to 2024, divided into training and validation sets in a 7:3 ratio. The dataset underwent 2000 iterations with a learning rate of 0.001. Among them, the experience replay pool represents a dataset that stores the interaction experience between the agent and the environment, which is used to break the correlation between samples; the target network represents a network with the same structure as the main network but with lagging parameter updates, which is used to improve training stability; and the learning rate represents the step size of the network parameter updates, which controls the training convergence speed.

[0079] After training, the data validation accuracy on the validation set reached 98.2%, which is 7% lower than the traditional rule-based validation method, enabling real-time and accurate iteration of the dynamic element library.

[0080] In one feasible approach, the DRL-ARMA model is used to generate investment strategy improvement solutions. Its training requires the completion of MDP modeling, sequential decision logic construction and strategy verification, and the solutions need to be fed back to the dynamic element perception engine to form a closed loop.

[0081] Specifically, the MDP state space includes three dimensions: resource load (construction equipment utilization rate, 0-1), project progress (actual progress / planned progress, 0-1.2), and deviation degree (the ratio of decision deviation value to warning threshold, 0-1.5); the action space includes 15 discrete actions, covering investment timing adjustments (advanced / delayed by 1-3 months), resource reallocation (equipment / team transfer ratio 0-0.3), and risk mitigation measures (ecological compensation / market hedging); the reward function is: ; in, Represented as the immediate reward value for reinforcement learning; , , These are the weighting coefficients for each indicator, with values ​​of 0.5, 0.3, and 0.2 respectively. This is expressed as the rate of improvement in the effectiveness of the plan; This represents the time taken to generate the solution; This is expressed as the resource utilization rate.

[0082] In this context, the sequential decision logic transforms the resource adjustment problem into a Markov decision process, with the state transition probability formula as follows: ; in, Represented as state transition probability; This represents the current state; This represents the state at the next moment; This indicates the action to be performed; Represented as The state at any given moment; Represented as Actions at any given moment.

[0083] The model combines the time-series forecasting capabilities of the ARMA algorithm to make short-term predictions of project progress and resource load, assisting in action selection. Among them, ARMA stands for Autoregressive Moving Average, which is used for predictive analysis of time series data.

[0084] Training was performed using the SARSA algorithm, with an experience replay pool size of 20,000, a learning rate of 0.0005, and a discount factor. 5000 iterations Among them, the SARSA algorithm is represented as an online temporal difference learning algorithm; discount factor This is represented as a discount factor for future rewards, used to balance immediate rewards versus long-term rewards.

[0085] In practice, the model training used 36 decision-making bias cases from 2018 to 2024. During the validation phase, six new cases from 2024 were added. The improved solutions generated by the model reduced decision-making bias by 22%-35%, with investment timing adjustment solutions improving efficiency by 1.2%-2.5% and resource reallocation solutions reducing resource conflict rates by 18%. These solutions need to be fed back to the threshold correction module of the dynamic element perception engine, triggering the recalibration of dynamic parameter thresholds, forming a complete decision-making closed loop of "perception-quantification-optimization-feedback".

[0086] Example 2 like Figure 1 As shown below, the present invention provides a dynamic investment decision-making method for a group of hydropower station projects that considers uncertainty. The dynamic investment decision-making method for a group of hydropower station projects that considers uncertainty described below can be referred to in correspondence with the MIDAS-QR analysis model, coupled time series network mapping model, moment analysis model, interaction effect multi-objective programming model, spectral entropy rate analysis model, MOPSO optimization model and deep reinforcement learning model pre-trained in Example 1 above. like Figure 1 As shown, Figure 1 This application provides an overall process for a dynamic investment decision-making method for hydropower station project groups that considers uncertainty. This method addresses uncertainties in hydropower station project group investment decisions, such as market price fluctuations, random hydrological inflows, and dynamic adjustments to policy constraints. Based on the model system pre-trained in Example 1, the method employs four steps: S1, constructing and updating a dynamic element library with quantified uncertainty; S2, quantifying the collaborative benefits of the project group under uncertainty; S3, generating a risk-controllable investment portfolio scheme; and S4, implementing decision monitoring and closed-loop optimization. This achieves dynamic adaptation and risk control of the decision-making scheme. The overall architecture can be directly deployed in the investment decision-making computer system of a hydropower station enterprise. Those skilled in the art can reproduce the entire process based on the detailed description of this embodiment.

[0087] In this embodiment, for each hydropower station project to be decided, 18 decision-making factors are collected around the four core objectives of "strategic adaptation, construction feasibility, economic rationality, and risk controllability". The factors are as follows: (1) Local government support, (2) Synergy between the project and other businesses of the investment entity, (3) Project inclusion in national planning, (4) Project functional positioning and location conditions, (5) Project needs, (6) Power system access and transmission conditions, (7) Project topographic conditions, (8) Project geological conditions, (9) Project water source conditions, (10) Electromechanical and metal structure installation conditions, (11) Environmental protection and soil and water conservation requirements, (12) Number and difficulty of land acquisition and resettlement, (13) Construction conditions, (14) Project unit cost, (15) Involvement of water source protection areas, (16) Involvement of permanent basic farmland, (17) Involvement of ecological protection red lines and (18) Ancient trees, famous trees, cultural relics and historical sites.

[0088] Specifically, the data sources collected in this step are supplemented with uncertain correlation data based on Example 1, covering three types of data, and the collection devices, frequencies, and data dimensions are as follows: Hydrological uncertainty data: Relying on the remote telemetry units (RTUs) of 120 hydrological monitoring stations in the Lancang River Basin, hourly inflow data (630,720 data points) were collected from 2018 to 2024. At the same time, daily precipitation probability data (2,555 data points) from meteorological stations in the basin were also collected. The equipment selected was the SL651-2014 hydrological remote telemetry terminal certified by the Ministry of Water Resources. The sampling frequency was consistent with that in Example 1 to ensure data time sequence alignment. Market uncertainty data: Daily electricity price data (2555 records) and monthly bilateral transaction winning probability data (84 records) from 2018 to 2024 were obtained from the open interface of the Yunnan Power Exchange Platform. At the same time, 100 sets of electricity price fluctuation scenario data (confidence level 95%) were generated through Monte Carlo simulation. Data transmission adopted an encrypted VPN channel to avoid market data leakage. Policy uncertainty data: Quarterly hydropower policy documents (28 documents) from 2018 to 2024 were collected from the National Energy Administration's policy disclosure platform, and uncertainty indicators such as the delay time of policy implementation and the fluctuation range of subsidy intensity were added. OCR recognition technology was used to convert unstructured policy texts into quantitative scores.

[0089] In the preferred embodiment, the data preprocessing process relies on the frequency adaptation layer of the MIDAS-QR model pre-trained in Example 1 to correct the temporal misalignment problem of data at different frequencies. Simultaneously, a new uncertainty noise filtering module is added, the specific operation of which is as follows: Outlier handling: For hydrological inflow data, an improved 3σ principle is adopted, introducing inflow probability weights (the weight coefficients are determined by the synchronization bifurcation coefficients output by the moment analysis model pre-trained in Example 1), and extreme inflow samples with a probability of less than 5% are removed (a total of 18 samples are removed); for electricity price data, the iAAFT algorithm is used to generate substitute data, and outlier samples that deviate from the 95% confidence interval of the substitute data distribution are screened out (a total of 12 samples are removed). This method is based on the uncertainty noise separation logic of nonlinear time series data. Frequency adaptation: Using the weekly time window of Example 1, the scale of daily electricity price, monthly water inflow, and quarterly policy data is unified, generating 365 weekly samples. At the same time, uncertainty labels (such as electricity price fluctuation coefficient and water inflow probability) are added to each weekly sample. The label values ​​are obtained by mapping the high-frequency feature coefficients output by the spectral entropy rate model pre-trained in Example 1. Format standardization: Quantifying policy scores The Min-Max standardization of an interval is given by the following formula: ;in, Represented as a standardized policy quantitative score; This represents the raw value of the policy quantitative score; This is represented as the minimum value of the policy quantitative score; This represents the maximum value of the policy quantitative score; Z-score standardization is applied to hydrological and market data, and the formula is as follows: ; in, Represented as standardized hydrological or market data values; Represented as raw values ​​of hydrological or market data; Represented as the sample mean of hydrological or market data; The standard deviation of a sample of hydrological or market data; To ensure that the data dimensions of the input model are consistent, its standardized formula is the basic formula of Example 1, with only the addition of an independent standardized channel for uncertainty labels to avoid interference of labels on feature values.

[0090] In practice, this step requires quantifying the three core uncertainty factors. The calculation method for the fluctuation coefficient of water inflow during periods of abundance and scarcity is as follows: ;in, This is expressed as the coefficient of variation in water inflow during periods of high and low water levels. The sample standard deviation is expressed as the inflow water flow rate. This is expressed as the sample mean of the inflow water flow rate; Represented as hydraulic coupling correlation degree; The calculation method for the electricity price fluctuation coefficient is as follows: ;in, This is expressed as the electricity price fluctuation coefficient; The coefficient of variation of electricity price , For the standard deviation of electricity prices, This represents the average electricity price. Represented as electricity price and water supply frequency at angular frequency Mutual information rate The calculation method for the policy implementation delay coefficient is as follows: ;in, This is represented as the policy implementation delay coefficient; This indicates the actual time when the policy was implemented; This indicates the timeframe for the implementation of the policy plan; Indicated as policy constraint level The quantitative indicators and calculation methods are shown in Table 1. The indicator design in Table 1 combines the random fluctuation characteristics of time series data with the industry characteristics of hydropower station investment. The calculation basis of the "fluctuation coefficient of water inflow" is the hydraulic coupling correlation degree output by the coupled time series network mapping model pre-trained in Example 1, and the "policy implementation delay coefficient" is obtained by statistical fitting based on the time difference of historical policy implementation.

[0091] Table 1. Quantitative Indicators of Uncertainty Factors in Hydropower Station Project Groups

[0092] In one feasible approach, the update of the feature database relies on the DQN data validity verification model pre-trained in Example 1, and adds an uncertainty threshold triggering mechanism, the specific execution flow of which is as follows: Real-time data access: Data streams from hydrological RTUs and power trading platforms are accessed in real time through the edge computing nodes of the computer system, with an access frequency of once per hour. Data caching uses a Redis database to ensure low-latency access. Validity verification: Real-time data is input into the DQN model, and the model outputs action instructions such as "verification passed", "needs to be completed", and "needs to be removed". The verification logic is the same as in Example 1, except that a new verification branch for uncertainty indicators is added. When the uncertainty indicators of the data exceed the high-risk threshold in Table 1, a secondary verification is automatically triggered, and historical data from the same period are called for comparison. Library update trigger: When feature data passes verification for three consecutive time windows and the change in uncertainty index exceeds 5%, a batch update of the feature library is triggered. During the update process, the Beta multinomial weight constraint function of Example 1 is used. , The weights of the new data are assigned, and the weight constraint function is as follows: ;in, Represented as the first The weighting coefficients corresponding to each lag order; Represented as the lag order; Represented as the first The maximum lag order of each feature variable; This is represented as the first shape parameter of the Beta function; This is represented as the second shape parameter of the Beta function; Represented as the Beta probability density function, its expression is: , Represented as the Gamma function; ensuring the temporal continuity of data within the database.

[0093] Table 2 shows the update threshold table for the dynamic feature library. This threshold is derived from the validation set error of the pre-trained model in Example 1 and can be directly used as a configuration parameter for the computer system.

[0094] Table 2 Dynamic Feature Library Update Threshold Configuration Table

[0095] In this embodiment, the quantification of the collaborative benefits of the project group is based on the spectral entropy rate analysis model, moment analysis model and interaction effect multi-objective programming model in Embodiment 1. An uncertainty correction coefficient is introduced to realize the dynamic calibration of the coupling correlation degree and the risk weighting of the collaborative benefits. Its core operation includes four sub-steps: coupling network reconstruction, synchronous bifurcation point correction, hierarchical interaction quantification, and collaborative benefit scoring.

[0096] Specifically, the coupled network reconstruction relies on the coupled time series network mapping model pre-trained in Example 1. To address the temporal synchronization deviation caused by uncertainty, a correlation correction module is added, and its operation method is as follows: Synchronization determination threshold correction: In Example 1, the Pearson correlation coefficient threshold for synchronization determination was 0.8. In this step, the threshold is adjusted according to the uncertainty risk level in Table 1—the threshold is increased to 0.85 in high-risk scenarios (reducing false synchronization associations), and decreased to 0.75 in low-risk scenarios (preserving weak coupling associations). The threshold adjustment formula is as follows: ;in, This is represented as the modified threshold for the Pearson correlation coefficient used to determine synchronicity. It is expressed as a weighted sum of three types of uncertainty, and its calculation formula is: The weights were determined by experts to be 0.4 for hydrology, 0.4 for the market, and 0.2 for policy. Network topology reconstruction: Based on the corrected synchronization threshold, the hydraulic, power, and resource coupling relationships of 12 candidate hydropower stations are recalculated, generating a coupling adjacency matrix with uncertainty labels. The matrix elements are tuples of "correlation strength and uncertainty level". For example, "0.86-medium risk" means that the hydraulic coupling strength is 0.86 and the corresponding hydrological uncertainty is medium level. Network density calibration: The modular optimization algorithm of Example 1 is used to iteratively adjust the link weights. The number of iterations is set to 500, and the convergence condition is that the change in cosine similarity of the adjacency matrix is ​​≤1e-5. The cosine similarity calculation formula is: ;in, It is expressed as the cosine similarity of the adjacency matrix; , These are respectively represented as the adjacency matrix vectors of adjacent iteration steps; Represented as the dot product of a matrix and a vector; , These are respectively represented as the magnitudes of the matrix and the vector; At the same time, an uncertainty attenuation coefficient is added to each network node, and the link weight of high-risk nodes is attenuated by 10%, ensuring that the network topology reflects the coupling state under actual risks.

[0097] In the preferred scheme, the identification of the synchronous bifurcation critical point is the core of quantifying coupling stability. This step introduces an uncertainty perturbation term based on the moment analysis model in Example 1, and its algorithm implementation is as follows: Correction of dynamic equations: The single-node dynamic equations of Example 1 are as follows: This step adds an uncertainty disturbance term. The corrected equation is: ;in, Represented as the first Power station nodes Rate of change of state at time t; It is represented as the self-dynamic function of the power station node, characterizing the relationship between power generation and inflow rate; Represented as global coupling strength; Represented as the Laplace matrix element of the coupled network; Represented as an internal coupling matrix; Represented as the first Power station nodes The state value at any given time; Represented as an uncertainty disturbance term, its expression is: , for The distribution of the random disturbance term in the interval is obtained by fitting historical uncertainty data; Represented as the total number of network nodes; Jacobian matrix solution: Solve for the Jacobian matrix of the modified dynamic equations. The expression for the Jacobian matrix is ​​as follows: ;in, Represented as the Jacobian matrix of the first... Line number Column elements; Calculate the real part of the eigenvalues ​​of the matrix. When the real part of the eigenvalue changes from negative to positive, the corresponding critical point of the parameter is the corrected synchronization bifurcation point. Bifurcation point verification: The model was verified using 10 sets of parameter samples added in 2024 that were not used in the training. The results showed that the corrected bifurcation point prediction error was ≤4%, which was further reduced compared to the 5% error in Example 1. The core reason for the reduction in error is that the introduction of the uncertainty perturbation term makes the model closer to the actual dynamics.

[0098] Table 3 shows the correction coefficients for coupling correlation under different uncertainty scenarios. These coefficients are obtained by mapping the real parts of the eigenvalues ​​of the moment analysis model and can be directly used for input calibration in subsequent spectral entropy rate analysis.

[0099] Table 3. Coupling correlation correction coefficients under different uncertainty scenarios

[0100] In practice, the correction coefficients in Table 3 are introduced into the calculation of the three types of information rates for spectral entropy rate analysis. The correction of the core formula is as follows: Single-item spectral entropy rate: The formula in Example 1 is as follows The corrected version is: ;in, Represented as the corrected number Each item at angular frequency Spectral entropy rate under the following conditions; This is represented as the original spectral entropy rate calculated in Example 1; For nodes Uncertainty attenuation coefficient (taken from Table 3); Represented as the first Dimensions of time-series data for each project; Represented as the first Determinant of the power spectral density matrix of time series data for each project; It is expressed as a natural constant, with a value of approximately 2.71828; Project mutual information rate: The corrected formula is: ; in, The item is shown as the revised version. With the project At angular frequency Mutual information rate at the following levels; This is expressed as the original mutual information rate calculated in Example 1; This is a correction coefficient for the coupling correlation degree of the project pair; Project Group O-Information Rate: Iterative Formula Maintains the Information Rate of Example 1 Its gradient term The formula is: ;in, Represented as the first The project and the elimination of the first The sub-project teams of each project are at angular frequency Mutual information rate at the following levels; Represented as the first The project and the original Each project team at angular frequency Mutual information rate at the following levels; only the mutual information rate at the following levels; The mutual information rate in the original code is replaced with the corrected value, while the number of bootstrap iterations remains set to [value missing]. To ensure statistical significance.

[0101] The quantitative results of this step show that, under high-risk scenarios, the low-frequency mutual information rate of the Lancang River cascade hydropower stations A and B decreased from 0.32 nats / Hz to 0.27 nats / Hz, while the O-information rate increased from -0.18 nats / Hz to -0.12 nats / Hz. This indicates that uncertainty weakens the synergistic coupling strength between projects, a result consistent with the risk response logic of actual hydropower station operation.

[0102] In this embodiment, the synergistic benefit scoring relies on the multi-objective programming model of interaction effects in Embodiment 1, and introduces a risk weighting coefficient. Its objective function is modified as follows: ;in, This is represented as the revised total score for the synergistic benefits of the project cluster. Represented as the first Basic benefit score for each project; Represented as the first Synergistic gain coefficient of each project; Represented as a weighted sum of uncertainties; This is represented as a risk weight (calibrated to 0.2 in Example 1). This is expressed as CVaR risk concentration. This represents the total number of projects within the project group; Specifically, the cooperative gain coefficient The calibration is based on Example 1, with the addition of a risk correction item: when the water quality compliance rate is ≥90% and the uncertainty is low risk, If it is high risk, then only When provincial and national subsidies are provided and policy uncertainty is considered low-risk, High risk After training, the model's output synergistic benefit score matched the actual data by 91%, a further improvement from 89% in Example 1, indicating that the risk-weighted correction improved the model's actual adaptability.

[0103] In this embodiment, the optimization of the portfolio scheme relies on the MOPSO optimization model pre-trained in Embodiment 1, introduces a set of uncertain scenarios and CVaR risk constraints, and achieves a multi-objective balance of "maximizing comprehensive benefits, minimizing risk concentration, and optimizing synergistic benefits". Its core operations include four sub-steps: uncertainty scenario generation, MOPSO model adaptation, CVaR risk constraint embedding, and non-dominated solution screening.

[0104] In the preferred scheme, Monte Carlo simulation is used to generate 100 uncertainty scenarios. The scenario variables are the three types of uncertainty indicators in Table 1. The generation logic is as follows: Variable distribution fitting: The normal distribution was used for the hydrological inflow fluctuation coefficient. , Its probability density function is: ; in, Represented as the probability density value of a normal distribution; This is represented by the value of the inflow fluctuation coefficient; It is represented as the mean of a normal distribution; Expressed as the standard deviation of a normal distribution; The electricity price fluctuation coefficient is adopted according to a log-normal distribution ( , Its probability density function is: ;in, This is represented by the value of the electricity price fluctuation coefficient; It is expressed as the natural logarithm of the electricity price fluctuation coefficient; The policy implementation delay coefficient is calculated using a Beta distribution. , Its probability density function is: ;in, This represents the value of the policy implementation delay coefficient; , Represented as the shape parameter of the Beta distribution; The distribution parameters were obtained by fitting historical data from 2018 to 2024; Scenario Combination: The three variables are combined using Latin hypercube sampling (LHS) to avoid scenario duplication, generating a total of 100 independent scenarios. Each scenario corresponds to a synergistic benefit score and risk concentration. Scene weight allocation: Each scene group is assigned a weight based on the occurrence probability of historical scenes. The weight of high-probability scenes (occurrence probability > 5%) is increased to 0.02, and the weight of low-probability scenes (< 1%) is reduced to 0.005, to ensure that the optimization results are biased towards actual high-frequency scenes.

[0105] Specifically, the MOPSO model, based on the chaotic mutation search mechanism and elite individual preservation mechanism of Example 1, undergoes two core adaptations: Adaptive adjustment of inertial variables: The iterative formula for inertial variables in Example 1 is as follows: This step adds an uncertainty feedback item and is revised as follows: ; in, Represented as the corrected number The inertial variables of the next iteration; This represents the original inertial variable calculated in Example 1; This is represented as the maximum inertial variable, with a value of 0.9; It is represented as the smallest inertial variable, with a value of 0.2; This represents the current iteration number; This represents the maximum number of iterations. Represented as a weighted sum of uncertainties; In high-risk scenarios, the inertial variable is increased, enhancing the global search capability; in low-risk scenarios, the inertial variable is decreased, accelerating local convergence. Chaotic mutation probability adjustment: mutation control factor Unchanged, chaotic sequence The initial value is determined by the uncertainty level—in high-risk scenarios. (Enhancing mutation amplitude), in low-risk scenarios (Reduce mutation amplitude), among which This represents the maximum number of iterations. It is represented as the chaotic sequence value of the previous iteration step.

[0106] Table 4 shows the parameter configuration and convergence index of MOPSO in uncertain scenarios. These parameters were obtained by tuning the model validation set of Example 1 and can be directly used as the algorithm configuration parameters of the computer system. The convergence speed is improved by 12% compared with Example 1, and the diversity of non-dominated solutions is improved by 15%.

[0107] Table 4. Parameter Configuration and Convergence Indicators of MOPSO under Uncertainty Scenarios

[0108] In practical implementation, the risk constraint relies on the CVaR risk measurement model of Example 1, which is embedded into the optimization objective of MOPSO. The core formula is derived as follows: CVaR calculation: for the th The first scenario Cyclical market returns Calculate the loss function: ;in, Represented as the first The first scenario Market loss value over a period of time; Represented as Average electricity price over the period; Represented as Market allocation ratio during the cycle; Represented as a scene Down The actual returns over the period; CVaR approximation: using Example 1 ,in: This is expressed as an approximation of CVaR; Represented as VaR value; Represented as the number of scenes; This is expressed as the confidence level; It is represented as the positive part of the difference between the loss value and the VaR value, that is, when Time to take Otherwise, take 0; Constraint embedding: As a hard constraint of MOPSO ( To ensure that the risk loss of the portfolio does not exceed 10% of the expected return, the penalty factor for the constraint is set to 10 to prevent particles from violating the constraint.

[0109] In one feasible approach, the non-dominated solution screening relies on the crowding calculation mechanism of Example 1, with the addition of risk preference weights, and the operation process is as follows: Congestion calculation: The congestion formula from Example 1 is used: ; in, Represented as the first The degree of crowding of individuals; Represented as the number of objective functions; Represented as the first The values ​​of the objective function; Represented as the first The maximum value of each objective function; Represented as the first The minimum value of an objective function; , Represented as the first , The first individual One objective function value; Risk preference weighting: Based on the decision-maker's risk preference, weights are assigned to the three objective functions—the weights for risk-averse preferences are "overall benefit 0.3, risk concentration 0.5, synergistic benefit 0.2", for risk-neutral preferences are "0.4, 0.3, 0.3", and for risk-aggressive preferences are "0.5, 0.2, 0.3"; Solution output: The non-dominated solutions are sorted according to risk preference weights, and the top 10 solutions are output for decision-makers to choose from. Each solution is accompanied by a three-dimensional index of "comprehensive benefit - risk concentration - synergistic benefit" and the uncertainty adaptation range. For example, the optimal solution (risk neutral type) achieves a comprehensive benefit of 1.051 billion yuan, a risk concentration of 0.10, and a synergistic benefit score of 89 points, which can be adapted to uncertainty scenarios with medium risk and below.

[0110] In this embodiment, decision monitoring and closed-loop optimization rely on the DRL-ARMA model pre-trained in Embodiment 1, combined with real-time monitoring data from the coupled network, to achieve dynamic correction of decision bias and feedback update of the feature database. Its core operations include four sub-steps: deployment of real-time monitoring indicators, DRL-ARMA bias correction, feature database feedback update, and verification of decision effect.

[0111] Specifically, real-time monitoring sensors were deployed in the central control systems of the 12 candidate hydropower stations. The monitoring indicators were divided into three categories, and their monitoring frequencies and thresholds are as follows: Resource load indicators: Construction equipment utilization rate (monitoring frequency 1 time per day, early warning threshold >80%), relying on sensor collection of the Industrial Internet of Things (IIoT), and data is connected to the SCADA platform of the central control system; Project progress indicators: actual progress / planned progress (monitoring frequency once a week, warning threshold <0.8 or >1.2), compared based on the progress module of the BIM model; Decision deviation index: The deviation between actual investment return and planned return (monitoring frequency: once a month, warning threshold > ±5%). The deviation value is obtained by comparing the output value of the interaction effect multi-objective programming model in Example 1 with the actual value. The formula for calculating the deviation rate is as follows: ; in, This is expressed as the decision bias rate; This is expressed as actual investment return; This is expressed as planned investment returns; All monitoring data is synchronized to the real-time database of the decision-making computer system and stored using the time-series database InfluxDB to ensure efficient querying of massive amounts of time-series data.

[0112] In the preferred embodiment, the state space, action space, and reward function of the DRL-ARMA model are modified based on Example 1 by adding an uncertainty bias dimension, as follows: State space expansion: Based on the original state space (resource load, project schedule, deviation degree), an uncertainty deviation dimension (the difference between actual uncertainty and predicted value) is added, expanding the state space dimensions from 3 to 4. Additional action space: In addition to the original 15 discrete actions, "risk mitigation measures adjustment" actions (such as adding ecological compensation and market hedging) are added, expanding the action space to 18; Reward function correction: The reward function in Example 1 is as follows The corrected version is: ; in, Represented as the modified immediate reward value for reinforcement learning; This is represented as the original reward value calculated in Example 1; , , These are the weighting coefficients for each indicator, with values ​​of 0.5, 0.3, and 0.2 respectively. This is expressed as the rate of improvement in the effectiveness of the plan; This represents the time taken to generate the solution; This is expressed as the resource utilization rate. This is expressed as the uncertainty deviation value, which is the difference between the actual uncertainty and the predicted value; it ensures that the reward function penalizes the uncertainty deviation.

[0113] Table 5 is a comparison table of the decision bias correction effect of DRL-ARMA. The table was obtained through the verification of 6 newly added bias cases in 2024. The results show that the decision bias was reduced by 28%-38% after correction, the time for generating the solution was shortened by 15%, and the computational resource utilization rate was reduced by 12%. Its correction effect is better than that of Example 1 (22%-35%), indicating that the introduction of the uncertainty dimension improved the correction accuracy of the model.

[0114] Table 5 Comparison of the Decision Bias Correction Effects of DRL-ARMA

[0115] In this embodiment, after the DRL-ARMA model generates a correction scheme, the dynamic feature library is automatically updated in response. The process is as follows: Feature extraction of the revised scheme: Extract key parameters (such as the magnitude of investment timing adjustment and the proportion of resource reallocation) from the revised scheme and convert them into quantitative feature values; Feature database threshold adjustment: Based on the effect of the correction scheme, adjust the update threshold in Table 2. For example, if the deviation correction range is >35%, the validity verification pass rate threshold of the corresponding data type will be reduced by 1% to enhance data inclusiveness. Model parameter iteration: Input the feature values ​​of the corrected scheme into the pre-trained model of Example 1, perform a mini-batch iteration (set the number of iterations to 100, and the learning rate to 0.0001), and update the model's weight parameters. The parameter update formula is as follows: ;in, Represented as the updated model parameters; This represents the model parameters before the update; This is expressed as the learning rate for mini-batch iterations; Represented as loss function For parameters Gradient; ensure the model dynamically adapts to the decision-making scenario.

[0116] In practice, the effectiveness of the decision-making process is verified using a two-dimensional approach: Short-term verification: Monthly verification of the deviation between actual investment returns and planned returns. A deviation rate of ≤5% is considered acceptable. The deviation rate formula is the same as in 4.1. Calculation; Long-term verification: The synergistic benefit score of the project group and the year-end energy storage are verified annually. A synergistic benefit score of ≥85 points and year-end energy storage of ≥1 billion kWh are considered qualified.

[0117] The actual application results of the Lancang River cascade hydropower stations in 2024 show that the decision-making scheme of this embodiment achieved a comprehensive benefit of RMB 1.051 billion (an increase of 12.3% year-on-year), a risk concentration of 0.10 (a decrease of 15.2% year-on-year), and an energy storage of RMB 1.12 billion kWh at the end of the year (an increase of 7.6% year-on-year). The accuracy rate of decision deviation early warning reached 95%, which is a further improvement compared to 93% in Embodiment 1, fully verifying the effectiveness and reliability of this embodiment in uncertain scenarios.

[0118] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0119] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for dynamic investment decision of hydropower station project group considering uncertainty, characterized in that, Comprise the following steps: S1, construct a dynamic element perception and iteration engine, and acquire the basic decision data source of the hydropower project group according to the dynamic element perception and iteration engine; Extract several decision elements from the basic decision data source, and divide the several decision elements into static basic elements and dynamic sensitive elements; According to the dynamic sensitive elements, the evaluation threshold interval is dynamically calculated and corrected through real-time data and high-frequency feature analysis technology, and a real-time iterative dynamic element library is formed; S2, construct a project group coupling network according to the dynamic element library; by analyzing the synchronous stability, dynamic characteristics and interaction information of the project group coupling network, the hydraulic, electric and resource coupling correlation degree between the projects is quantified; combined with the ecological and policy coordination dimensions, the project coordination benefit score is generated; S3, according to the coordination benefit score, introduce risk value measurement, construct a multi-objective optimization model with the goal of maximizing comprehensive benefit and minimizing risk concentration; through optimization algorithm, the optimal hydropower station project group investment portfolio scheme is output; S4, set monitoring nodes in the investment whole process in stages, collect scheme implementation data and compare with expected values, generate deviation diagnosis report; according to the deviation diagnosis report, generate targeted investment strategy improvement scheme, and feed back to the dynamic element perception and iteration engine of S1, trigger threshold re-calibration, form a decision closed loop.

2. The method for dynamic investment decision of hydropower station project group considering uncertainty according to claim 1, characterized in that, The specific steps of step S1 are as follows: S11, construct a dynamic element perception and iteration engine, which comprises a data source docking module, a parameter calculation module and a threshold correction module; The data source docking module is used to dock the basin hydrological monitoring station, the electric power transaction platform and the national energy policy database, the parameter calculation module is used to perform quantitative calculation of dynamic parameters, and the threshold correction module is used to update the element scenario threshold interval; S12, acquire the basic decision data source of the basin hydropower project group through the data source docking module of the dynamic element perception and iteration engine, the basic decision data source includes industry hydropower project evaluation standard, feasibility study report of each selected project, national energy policy document, basin hydrological historical monitoring data, real-time hydrological prediction data, historical electric power market transaction data and real-time electricity price data, and several decision elements are extracted from the basic decision data source; the several decision elements are divided into static basic elements and dynamic sensitive elements; the static basic elements include topographic / geological conditions and hub engineering structure parameters; the dynamic sensitive elements include local government support policy, on-grid electricity price, water inflow prediction value and ecological red line range; S13, acquire historical water inflow monitoring data and current water inflow prediction data through the parameter calculation module of the dynamic element perception and iteration engine, and calculate the water inflow prediction deviation coefficient through the formula: ； Wherein, the current water inflow prediction value represents the basin water inflow flow prediction data published by the hydrological monitoring station at the current period; the historical same period water inflow mean value represents the average value of the basin water inflow flow at the same period in the past 10 years; at the same time, historical same period electricity price data and real-time electricity price data are acquired, and the formula is as follows: ； Wherein, the real-time electricity price is represented as the online electricity price data published by the electricity trading platform in real time; the historical same period electricity price average is represented as the average of the online electricity price in the same period in the past 10 years, and the electricity price fluctuation coefficient is calculated; S14, acquire the multi-frequency data source of the dynamic sensitive elements, the multi-frequency data source includes daily online electricity price data, monthly water inflow prediction data, and quarterly policy adjustment information, input the multi-frequency data source into the pre-trained MIDAS-QR analysis model, and extract high-frequency characteristic coefficients through the frequency alignment algorithm and the Beta polynomial weight constraint function of the model; S15, combine the high-frequency characteristic coefficients, the water inflow prediction deviation coefficient, and the electricity price fluctuation coefficient obtained in the previous steps, and solve the multi-objective problem of "maximizing comprehensive benefit, minimizing risk concentration, and optimizing collaborative benefit" through the chaos mutation search algorithm and the elite individual reservation algorithm of the pre-trained MOPSO model, and output the optimal hydropower station project portfolio scheme; ； wherein, is expressed as a modified dynamic sensitive element contextualization threshold value; is expressed as an initial baseline threshold value for the dynamic sensitive element; , is expressed as a basin adaptation coefficient; is expressed as a Beta weight function adaptation coefficient, which modifies the dynamic sensitive element contextualization threshold interval, wherein a, b are basin adaptation coefficients, and g is a Beta weight function adaptation coefficient, and a, b are calibrated by cross-professional experts in combination with historical data of the basin hydropower projects in the past 10 years. S16, through the data source of the engine docking module, docking the real-time data of the state energy bureau policy database, the basin hydrological monitoring station and the power transaction platform, the corrected dynamic sensitive element is sceneized threshold and real-time data is input into the pre-trained deep reinforcement learning DQN model, and the action selection adopts -greedy strategy, through the Markov decision algorithm of the pre-trained deep reinforcement learning DQN model, the validity of the data is checked, the real-time data that passes the check is updated to the dynamic element library, and the real-time iteration of the element library is completed.

3. The method for dynamic investment decision-making of hydropower station project groups considering uncertainty according to claim 2, characterized in that, The specific steps of step S2 are as follows: S21, extract the single-project time series data of each alternative hydropower project from the dynamic element library, including hydraulic time series data, power time series data, construction resource time series data, and cost time series data; input the single-project time series data into the pre-trained coupled time series network mapping model, divide the time series data into 5 amplitude segments through the amplitude quantization algorithm, identify the time series synchronicity through the Pearson correlation coefficient, and construct a project group coupling network with the quantized amplitude segments as nodes and the time series synchronicity as links; S22, input the coupling network into the pre-trained moment analysis model and spectral entropy rate analysis model, input the nonlinear dynamic parameters of the coupling network through the moment analysis model, identify the network synchronization bifurcation critical point, and calculate the single-project entropy rate through the hierarchical interaction analysis algorithm of the spectral entropy rate analysis model, to quantize the single-project dynamic characteristics, project pair link strength, and project group hyperlink redundancy / collaborative coupling degree obtained from the single-project time series data, and finally obtain three types of coupling correlation indexes, i.e., hydraulic coupling degree, power coupling degree, and resource coupling degree; S23, acquire the basin ecological protection planning file, basin ecological monitoring data, and local energy policy file through the data source docking module, input the three types of coupling correlation indexes and the above ecological and policy data into the pre-trained interactive effect multi-objective planning model, supplement the ecological synergy and policy synergy effect dimensions according to the two-project / three-project interactive effect quantification system of the interactive effect multi-objective planning model, take the coupling correlation degree as the correction coefficient of the project basic benefit score, generate the project collaborative benefit score, and the target function formula is: ； wherein, represents the total score of the synergy benefits of the project portfolio; represents the base benefit score of the i-th project; represents the synergy gain coefficient of the i-th project; represents the risk weight; represents the risk concentration of the project portfolio.

4. The method for dynamic investment decision-making of hydropower station project groups considering uncertainty according to claim 3, characterized in that, The specific steps of step S3 are as follows: S31, acquire the total investment budget in the enterprise annual investment plan, the total amount of construction resources in the engineering resource library, and the ecological constraint upper limit in the basin ecological protection planning; At the same time, the historical electricity price fluctuation data and the water inflow rich and poor data are input into the pre-trained CVaR risk measurement model, and the CVaR risk index is calculated through the Monte Carlo simulation algorithm of the CVaR risk measurement model combined with the formula ​ wherein, is expressed as a conditional value at risk; is expressed as a confidence level; is expressed as a loss function of the portfolio; is expressed as a value at risk; is expressed as a probability density of the loss function; is expressed as a probability density of the random variable y; S32, input the project collaborative benefit score obtained in step S23, the total investment budget, the total amount of construction resources, the ecological constraint upper limit, and the CVaR risk index in step S31 into the pre-trained MOPSO model, solve the multi-objective problem of "maximizing comprehensive benefit, minimizing risk concentration, and optimizing collaborative benefit" through the chaos mutation search algorithm and the elite individual reservation algorithm of the pre-trained MOPSO model, and output the optimal hydropower station project portfolio scheme; The formula of the chaotic mutation search algorithm is as follows: ; wherein, represents the position of the particle after the chaotic mutation; represents the position of the particle before the mutation; represents the mutation control factor; represents the chaotic sequence value, ； wherein represents the chaotic sequence value of the previous iteration step; Inertia variable of elitist individual preserving algorithm The congestion level is calculated as follows as the iteration number is adapted from 0.9 to 0.2: ； wherein, denotes the crowding distance of the i-th individual; denotes the number of objective functions; denotes the value of the o-th objective function; , denote the maximum and minimum value of the o-th objective function, respectively.

5. The method for dynamic investment decision-making of hydropower station project groups considering uncertainty according to claim 4, characterized in that, The specific steps of step S4 are as follows: S41, divide the investment whole process into planning, construction and operation initial stage, set monitoring nodes in investment planning, construction and operation initial stage, collect the execution data of the optimal hydropower station project group investment portfolio scheme through the project site monitoring system, input the execution data into the pre-trained spectral entropy rate analysis model, diagnose the decision bias source through the hierarchical interactive analysis algorithm of the spectral entropy rate analysis model, and generate a bias diagnosis report; The execution data includes actual construction cost, total water inflow and market electricity price; the decision bias source includes parameter fluctuation bias, coupling imbalance bias and risk exceeding bias; S42, input the bias diagnosis report into the pre-trained deep reinforcement learning DQN model, model the Markov decision process through the pre-trained deep reinforcement learning DQN model, and generate a targeted investment strategy improvement scheme combining the sequence decision logic of the DRL-ARMA algorithm; S43, feed back the improvement scheme to the threshold correction module of the dynamic element perception and iteration engine constructed in S1, call the dynamic parameter threshold correction formula of S15 by the threshold correction module, recalibrate the scenario threshold interval of the dynamic sensitive element, complete the threshold correction, and form a decision closed loop.

6. The method for dynamic investment decision of hydropower station project group considering uncertainty according to claim 2, characterized in that, In step S1, the basin hydrological monitoring daily data extracted in S12 and the basin hydrological historical monitoring data and current hydrological station prediction data obtained in S13 in the past 10 years are obtained, and then the multi-frequency data source is input into the MIDAS-QR analysis model in S14. The frequency alignment is completed through the frequency difference fitting algorithm of the weekly yield sequence and daily electricity price / water inflow characteristic variables built in the MIDAS-QR analysis model, and the Beta polynomial weight constraint function is called to weight and distribute the high-frequency characteristics. The expression of the Beta polynomial weight constraint function is: ； wherein, represents a weight coefficient for the dth lag order; represents a lag order; represents a maximum lag order for the jth feature variable; , represents a Beta function shape parameter, and takes values of 1 and 1.4301, respectively; represents a Beta probability density function; The daily water inflow fluctuation characteristic value weighted by the weight function is extracted as the high-frequency characteristic coefficient.

7. The method for dynamic investment decision-making of hydropower station project groups considering uncertainty according to claim 3, characterized in that, In step S22, two nonlinear dynamic parameters of the project group coupling network in S21 are obtained, the two nonlinear dynamic parameters are input into the moment analysis model, the network synchronization stability criterion calculation logic built in the moment analysis model is used to identify the network synchronization region bifurcation critical point, and the network synchronization stability criterion is output; the single project power time series data extracted in S21 are obtained, the single project power time series data are input into the spectral entropy rate model, and the single project dynamic characteristics are calculated through the spectral entropy rate formula built in the spectral entropy rate model; the historical electricity price fluctuation data and water inflow data are obtained, the historical electricity price fluctuation data and water inflow data are input into the CVaR risk measurement model, the CVaR risk index is calculated through the Monte Carlo simulation algorithm built in the CVaR risk measurement model combined with the CVaR formula; at the same time, the total investment budget in the enterprise annual hydropower project investment plan, the total amount of construction resources in the engineering machinery equipment and construction team resource library, and the ecological constraint upper limit in the basin ecological environment protection special plan are obtained, and then the MOPSO model is called and the inertia variable iteration formula is configured: ； wherein, denotes the spectral entropy rate of the i-th project; denotes the normalized angular frequency; denotes the time series data dimension of the i-th project; denotes the i-th project power spectral density matrix determinant, the logarithmic operation and dimension weighting processing are performed on the time series data, then the power generation power time series data of the project pair are obtained, the project pair link strength is calculated, and substituted into the mutual information rate formula: ； wherein, represents the mutual information rate of item i and j; represents the determinant of the power spectral density matrix of the jth item; represents the determinant of the joint power spectral density matrix of item i and j, which quantifies the link strength between items; bootstrap data analysis is performed on the calculation results of the redundancy / coupling degree of the hyperlink of the item group, and finally the synchronous stability criterion is used as a correction factor to adjust the three types of coupling correlation indicators, to obtain the corrected three types of coupling correlation indicators.

8. The method for dynamic investment decision-making of hydropower station project groups considering uncertainty according to claim 4, characterized in that, ​ ； wherein, denotes the inertia variable at the kth iteration; denotes the maximum inertia variable; denotes the minimum inertia variable; denotes the current iteration number; denotes the maximum iteration number, which makes the inertia variable adaptively decrease from 0.9 to 0.2 with the iteration number; and in combination with the mean-CVaR utility function to complete the risk-benefit trade-off, the formula is: ； wherein, represents the investment utility value for period t; represents the expected return for period t; represents the conditional value at risk at confidence level for period t; represents the number of simulated scenarios; represents the return for period t under the lth scenario; represents the expected loss corresponding to the CVaR, performing a weighted integration of the expected return and the risk indicator.

9. The method for dynamic investment decision-making of hydropower station project groups considering uncertainty according to claim 3, characterized in that, In the dynamic element library in S1, hydraulic, electric power, construction resource real-time time series data is obtained, which is input into the coupled time series network mapping model, the amplitude of the time series data is divided into 5 amplitude sections by the amplitude quantization algorithm built in the coupled time series network mapping model, and then the synchronism identification algorithm is called and the Pearson correlation coefficient calculation formula is configured: ； wherein, denotes the Pearson correlation coefficient; , denotes the tth sample value of two time series data; , denotes the sample mean of two time series data; denotes the sample number, the determination threshold is set to 0.8 to identify the time series synchronicity; and the project group coupling network is constructed by taking the quantized amplitude segment as a node, the time series synchronicity as a link, and the synchronicity frequency as a link weight.

10. The method for dynamic investment decision of hydropower station project group considering uncertainty according to claim 4, characterized in that, In the interactive effect multi-objective planning model of step S23, the built-in objective function of the interactive effect multi-objective planning model is called, and the formula is: ； wherein, represents the total score of the synergy benefit of the project group; represents the basic benefit score of the i-th project; represents the synergy gain coefficient of the i-th project; represents the risk weight and has a value range of [0.1, 0.3]; represents the risk concentration degree of the project group, performs a product summation operation on the basic benefit score and the synergy gain coefficient of the project, and deducts the risk concentration degree after weighting; obtains the ecological synergy according to the basin ecological monitoring data, and obtains the policy synergy according to the local energy policy document.

Images