A multi-objective planning method and system for hydropower stations based on multimodal data fusion

Abstract

The present invention relates to the technical field of hydropower station optimization and scheduling, and specifically to a multi-objective planning method and system for a hydropower station based on multimodal data fusion. The method and system comprise the following steps: collecting multimodal data of the hydropower station, mining a first planning connection and a second planning connection between the multimodal data of the hydropower station, fusing the data according to the first planning connection and the second planning connection to obtain fused planning data; constructing multi-objective planning constraints of the hydropower station according to the fused planning data, including long-term planning constraints, medium-term planning constraints and short-term planning constraints; constructing a planning state space of a target object using the multimodal data of the hydropower station, including a long-term planning state space, a medium-term planning state space and a short-term planning state space; and sequentially planning the target objects of the long-term planning, medium-term planning and short-term planning using a reinforcement learning algorithm based on the multimodal data fusion to obtain corresponding planning strategies.

Description

Technical Field

[0001] The present invention relates to the technical field of hydropower station optimization scheduling, and in particular to a hydropower station multi-objective planning method and system based on multimodal data fusion. Background Art

[0002] As an important clean and renewable energy source, hydropower plays a key role in the global energy transition and the stable operation of power systems. It not only provides large-scale power generation capacity but also undertakes important grid ancillary services such as peak shaving, frequency regulation, and emergency backup. It also offers multiple benefits, including flood control, water supply, shipping, and ecological regulation. However, the operation and management of hydropower stations is extremely complex. Their power generation capacity is directly constrained by the randomness and volatility of water inflows within their basins, while electricity demand also fluctuates dynamically. Furthermore, hydropower station operations must strictly adhere to multiple physical and environmental constraints, including reservoir scheduling regulations, dam safety requirements, and downstream ecological and environmental protection regulations. Therefore, how to balance multiple objectives, such as maximizing power generation efficiency, ensuring water supply, flood control safety, and ecological protection, while meeting these complex constraints, and achieving optimal water resource allocation and refined, intelligent scheduling and operation of hydropower stations, remains a core technical challenge facing the hydropower industry.

[0003] Traditional hydropower station operation planning methods often rely on simplified models and limited single-source data (such as historical runoff or simple forecasts), or focus on single-objective optimization. These methods fail to fully reflect the complex dynamic characteristics of the system and the inherent trade-offs between multiple objectives. In recent years, the development of sensor technology, the Internet of Things (IoT), remote sensing technology, and big data technologies has enabled the acquisition of massive amounts of data from multiple sources, types, and scales, including hydrology, meteorology, operating conditions, power grids, and the environment (i.e., multimodal data). This data contains richer and more comprehensive information about hydropower stations and their operating environments, but it also presents challenges in data processing, information extraction, and effective utilization. How to effectively integrate this heterogeneous, multimodal data and explore its inherent correlations to achieve a more accurate understanding of the hydropower system state and more reliable predictions of future trends? Based on this, the development of advanced multi-objective planning models that dynamically optimize and coordinate various operational objectives, overcoming the limitations of traditional methods, has become a key research direction and technological frontier for improving the operational management and overall efficiency of hydropower stations.

[0004] Therefore, a multi-objective planning method and system for hydropower stations based on multimodal data fusion is proposed. Summary of the Invention

[0005] The purpose of the present invention is to provide a multi-objective planning method and system for a hydropower station based on multi-modal data fusion, which mines the first planning connection and the second planning connection between the multi-modal data of the hydropower station by collecting the multi-modal data of the hydropower station; the first planning connection represents a high-intensity association between the data with a clear physical mechanism or direct causality; through physical model analysis, high-frequency data correlation analysis and Granger causality test, a retrieval method for the first planning connection is constructed; the second planning connection represents an indirect, lagged, low-intensity association with statistical regularity between the data; time series graphs, scatter plots and distribution graphs of different modal data are drawn, and the connection between the data is analyzed through statistical description and correlation coefficient matrix between variables, and a retrieval method for the second planning connection is constructed; the multi-objective planning relationships of the long-term planning, medium-term planning and short-term planning of the hydropower station are obtained. A keyword set is prepared, and the first planning connection data vector and the second planning connection data vector of the multi-objective planning keyword are searched according to the retrieval method of the first planning connection and the second planning connection; the first planning connection data vector and the second planning connection data vector are preprocessed and data fused to obtain fused planning data corresponding to the long-term planning, the medium-term planning and the short-term planning; the multi-objective planning constraints of the hydropower station are constructed according to the fused planning data, including the long-term planning constraints, the medium-term planning constraints and the short-term planning constraints; the planning state space of the target object is constructed using the multimodal data of the hydropower station, including the long-term planning state space, the medium-term planning state space and the short-term planning state space; the target objects of the long-term planning, the medium-term planning and the short-term planning are planned in sequence using the reinforcement learning algorithm of the multimodal data fusion to obtain the corresponding planning strategies.

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

[0007] A multi-objective planning method for a hydropower station based on multimodal data fusion, comprising:

[0008] Collect multimodal data of the hydropower station and mine the first planning connection and second planning connection between the multimodal data of the hydropower station; the first planning connection represents a high-intensity association between data with a clear physical mechanism or direct causality; construct a retrieval method for the first planning connection through physical model analysis, high-frequency data correlation analysis and Granger causality test; the second planning connection represents an indirect, lagged, low-intensity association between data with statistical regularity; draw time series graphs, scatter plots, and distribution graphs of different modal data, analyze the connection between data through statistical descriptions and correlation coefficient matrices between variables, and construct a retrieval method for the second planning connection; obtain a set of multi-objective planning keywords for the long-term, medium-term and short-term plans of the hydropower station, and search for the first planning connection data vector and the second planning connection data vector of the multi-objective planning keywords according to the retrieval method of the first planning connection and the second planning connection; preprocess and fuse the first planning connection data vector and the second planning connection data vector to obtain the fused planning data corresponding to the long-term, medium-term and short-term plans;

[0009] Furthermore, the temporal modal data includes at least load demand, precipitation, and inflow flow; the spatial modal data includes at least upstream and downstream water level distribution, reservoir terrain, and transmission network topology; the structural modal data includes at least equipment status parameters of generator sets and flood discharge gates; and the semantic modal data includes at least operation and scheduling strategy texts, weather forecast reports, and superior scheduling commands.

[0010] Construct multi-objective planning constraints for the hydropower station based on the integrated planning data, including long-term planning constraints, medium-term planning constraints, and short-term planning constraints;

[0011] Furthermore, constructing multi-objective planning constraints includes: the planning constraint prediction model obtains and processes the fused planning data corresponding to the multi-objective planning keywords of long-term planning, medium-term planning and short-term planning respectively, and the output prediction results are the long-term planning constraints, medium-term planning constraints and short-term planning constraints.

[0012] Use the multimodal data of the hydropower station to construct the planning state space of the target object, including long-term planning state space, medium-term planning state space and short-term planning state space;

[0013] Furthermore, the planning state space includes: obtaining the planning keywords of the target object of the long-term plan, and searching for the first planning connection data vector and the second planning connection data vector of the planning keywords according to the retrieval method of the first planning connection and the second planning connection, and performing data fusion to obtain the long-term planning state space of the target object of the long-term plan; obtaining the planning keywords of the target object of the medium-term plan, and searching for the first planning connection data vector and the second planning connection data vector of the planning keywords according to the retrieval method of the first planning connection and the second planning connection, and performing data fusion to obtain the medium-term planning state space of the target object of the medium-term plan; obtaining the planning keywords of the target object of the short-term plan, and searching for the first planning connection data vector and the second planning connection data vector of the planning keywords according to the retrieval method of the first planning connection and the second planning connection, and performing data fusion to obtain the short-term planning state space of the target object of the short-term plan.

[0014] The reinforcement learning algorithm based on multimodal data fusion is used to plan the target objects of long-term planning, medium-term planning and short-term planning in sequence to obtain the corresponding planning strategies.

[0015] Furthermore, using a reinforcement learning algorithm based on multimodal data fusion to sequentially perform long-term planning, mid-term planning, and short-term planning on a target object includes: constructing three levels of intelligent agents, corresponding to long-term planning, mid-term planning, and short-term planning, respectively;

[0016] The first-level agent builds a reinforcement learning environment based on the long-term planning constraints. The state space of the first-level agent is built based on the long-term planning state space of the target object. The action space is built based on the execution strategy of the target object. The reward function is built based on the long-term planning goal. The DDPG algorithm is used to solve the long-term planning execution strategy.

[0017] The second-level agent builds a reinforcement learning environment based on the medium-term planning constraints and long-term planning execution strategy. The state space of the second-level agent is built based on the medium-term planning state space of the target object of the medium-term planning. The action space is built based on the execution strategy of the target object. The establishment function is built based on the medium-term planning goal and solved using the DDPG algorithm to obtain the execution strategy of the medium-term plan.

[0018] The third-level intelligent agent constructs a reinforcement learning environment based on the short-term planning constraints, the long-term planning execution strategy, and the medium-term planning execution strategy. The state space of the third-level intelligent agent is constructed based on the short-term planning state space of the target object of the short-term planning. The action space is constructed based on the execution strategy of the target object. The reward function is constructed based on the long-term planning goal. The solution is performed according to the DDPG algorithm to obtain the execution strategy of the short-term planning.

[0019] The present invention also provides a multi-objective planning system for a hydropower station based on multimodal data fusion, comprising:

[0020] The data acquisition module collects multimodal data of the hydropower station;

[0021] The data fusion module mines the first planning connection and the second planning connection between the multimodal data of the hydropower station, performs data fusion based on the first planning connection and the second planning connection, and obtains the fused planning data;

[0022] The planning constraint generation module constructs the multi-objective planning constraints of the hydropower station based on the integrated planning data, including long-term planning constraints, medium-term planning constraints and short-term planning constraints;

[0023] The planning state space generation module uses the multimodal data of the hydropower station to construct the planning state space of the target object, including long-term planning state space, medium-term planning state space and short-term planning state space;

[0024] The planning strategy generation module uses the reinforcement learning algorithm of multimodal data fusion to plan the target objects of long-term planning, medium-term planning and short-term planning in sequence to obtain the corresponding planning strategies.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. By systematically mining, storing, and establishing retrieval methods for the primary and secondary planning relationships, a powerful knowledge engine is built for intelligent planning of hydropower stations, enabling complex, multi-objective decision-making processes to be based on comprehensive, quantified system states and relationships.

[0027] 2. Driven by multi-objective planning keywords, we further search for the data vectors connecting the first and second plans, ensuring that the state space of each planning stage only contains the most important and relevant information for its decision-making, avoiding stuffing all the original multimodal data into the state space at once, significantly reducing the invalid dimensions of the state space, and improving the "signal-to-noise ratio" of information.

[0028] 3. The first-level intelligent agents focus on dealing with long-term uncertainties such as macro-climate and market; the second-level intelligent agents focus on dealing with medium-term uncertainties such as seasonal forecast errors and planned maintenance execution; the third-level intelligent agents focus on dealing with immediate uncertainties such as real-time fluctuations, ultra-short-term weather, and power grid instructions. The hierarchical structure allows the use of the most suitable method for handling the uncertainty at each level, better dealing with uncertainties at different levels, and improving system robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A flowchart of a multi-objective planning method for a hydropower station based on multimodal data fusion provided by an embodiment of the present invention;

[0030] Figure 2 A flowchart of data fusion provided by an embodiment of the present invention;

[0031] Figure 3 A schematic structural diagram of a multi-objective planning system for a hydropower station based on multimodal data fusion provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0032] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0033] Example 1:

[0034] A hydropower station has introduced a multi-objective planning method for a hydropower station based on multimodal data fusion provided by the present invention to improve the comprehensive management efficiency of the hydropower station. The method flow is as follows: Figure 1 As shown, the specific implementation is as follows:

[0035] Collecting multimodal data of the hydropower station, mining the first planning connection and the second planning connection between the multimodal data of the hydropower station, performing data fusion based on the first planning connection and the second planning connection to obtain fused planning data;

[0036] Furthermore, multimodal data includes data of several modes, including: temporal modal data including at least load demand, precipitation, and inflow flow; spatial modal data including at least upstream and downstream water level distribution, reservoir terrain, and transmission network topology; structural modal data including at least equipment status parameters of generator sets and flood discharge gates; semantic modal data including at least operation scheduling strategy text, weather forecast report, and superior scheduling command.

[0037] By collecting specific multimodal data such as load demand, precipitation, and inflow flow, we can construct a state description of the hydropower station and its environment that provides a more comprehensive perception, deeper understanding, more accurate prediction, and more scientific decision-making. This rich state description is a strong foundation for using reinforcement learning to achieve cross-time scale, multi-objective collaborative optimization planning, which can significantly improve the efficiency, safety, and reliability of hydropower station operation.

[0038] Furthermore, the first planning connection indicates a strong correlation between data with a clear physical mechanism or direct causality; through physical model analysis, high-frequency data correlation analysis and Granger causality test, a retrieval method for the first planning connection is constructed;

[0039] The second planning relationship represents the indirect, lagged, low-intensity correlation between data with statistical regularity; time series graphs, scatter plots, and distribution graphs of different modal data are drawn, and the relationship between data is analyzed through statistical descriptions and correlation coefficient matrices between variables to construct a retrieval method for the second planning relationship.

[0040] Furthermore, first formalize and quantify the storage, store physical laws, engineering formulas, equipment characteristic curves, etc. in a structured form, such as storing formula strings, parameter values, applicable equipment IDs, applicable scopes, etc., and create a "planning connection knowledge base" (which can be a relational database, graph database or customized data structure); assign a unique ID to each stored connection; establish an index based on key attributes, such as: the name of the variable involved, the ID of the equipment involved, the connection type (physical model, correlation, Granger), the level of association strength, etc.

[0041] Furthermore, the following example illustrates how to perform the first planning contact search:

[0042] Furthermore, we first search by variables: retrieve all first planning connections involving the variable "generator unit 3 output", including possible returns of the physical model (output = f (head, flow)), high correlation with current / voltage, etc.; Table 1 shows some generator unit data.

[0043] Table 1. Generator set data

[0044]

[0045] The mined connections are converted from implicit knowledge into explicit, structured data, allowing the planning system to directly query and utilize these data, improving data utilization efficiency, enhancing the interpretability of planning, and promoting the practical application of multimodal data fusion; through systematic mining, storage and establishment of retrieval methods for the first and second planning connections, a powerful knowledge engine is built for the intelligent planning of hydropower stations, enabling complex, multi-objective decision-making processes to be based on comprehensive, quantified system states and relationships.

[0046] Furthermore, the data fusion process is as follows Figure 2 As shown, it specifically includes: obtaining a multi-objective planning keyword set of the long-term planning, medium-term planning and short-term planning of the hydropower station, and searching for a first planning connection data vector and a second planning connection data vector of the multi-objective planning keyword according to the retrieval method of the first planning connection and the second planning connection;

[0047] The first planning connection data vector and the second planning connection data vector are preprocessed and data fused to obtain fused planning data corresponding to the long-term plan, the medium-term plan and the short-term plan.

[0048] Furthermore, taking long-term planning as an example, its multi-objective planning keyword set includes maximizing total power generation benefits, sustainable utilization of water resources, preservation and appreciation of equipment assets, and long-term flood control safety; according to the planning keywords, the relevant first and second planning connections are searched from the "planning connection knowledge base", and the corresponding specific values ​​or features are extracted from the real-time, historical, and predicted original multimodal data to construct the first planning connection data vector and the second planning connection data vector of the long-term planning.

[0049] Furthermore, the extracted first and second planning connection data vectors undergo necessary cleaning, format conversion, and integration to form a unified input for the final planning model. Feature fusion is then performed. This can be layered, feature-level, or model-level fusion. Feature-level fusion involves concatenating the preprocessed first and second planning connection data vectors to form a longer feature vector. Feature interaction can be performed before or after the concatenation to create new features. Model-level fusion involves processing the two vectors using different models rather than directly concatenating the raw data. For example, the first type of connection data can be used to calculate real-time available output, while the second type of connection data can be used through a predictive model to derive future risk indicators. The outputs of these models are then fused, ultimately yielding fused planning data corresponding to the long-term plan.

[0050] Taking the multi-objective planning keywords as the starting point, it ensures that only the data and connections most relevant to the current planning goals are extracted and integrated, avoiding the interference of irrelevant information, improving the efficiency and pertinence of data processing, providing high-quality input for subsequent planning algorithms, and supporting the effectiveness and practicality of subsequent hierarchical planning.

[0051] Construct multi-objective planning constraints for the hydropower station based on the integrated planning data, including long-term planning constraints, medium-term planning constraints, and short-term planning constraints;

[0052] Furthermore, constructing multi-objective planning constraints includes: the planning constraint prediction model obtains and processes the fused planning data corresponding to the multi-objective planning keywords of long-term planning, medium-term planning and short-term planning respectively, and the output prediction results are the long-term planning constraints, medium-term planning constraints and short-term planning constraints.

[0053] Furthermore, the planning constraint prediction model can be a machine learning model, a deep learning model, or a time series neural network model. By separately processing the integrated planning data of long-term planning, medium-term planning, and short-term planning, long-term planning constraints are obtained, including the average storage capacity control range of the reservoir in the next few years, the annual average runoff utilization rate target, the equipment replacement plan, the total installed capacity adjustment direction, etc.; medium-term planning constraints include the target water level range of the reservoir in the next few months, the monthly total power generation target range, the equipment shutdown plan, the minimum / maximum downstream flow requirements in a specific time period, etc.; short-term planning constraints include the real-time maximum / minimum output, ramp rate, available status, gate opening range, and downstream flow values ​​that must be met of each unit in the next few hours.

[0054] Traditional planning often uses fixed constraint values. However, by dynamically constructing constraints based on fused planning data, constraints can be adjusted in real time or periodically based on the system's real-time status, predicted changes, and potential risks. For example, when abundant water inflow is predicted in the future, the maximum downstream flow constraint may be relaxed; when equipment health is predicted to decline, its maximum output constraint may be reduced. This makes planning more realistic and flexible.

[0055] Use the multimodal data of the hydropower station to construct the planning state space of the target object, including long-term planning state space, medium-term planning state space and short-term planning state space;

[0056] Furthermore, the planning state space includes: obtaining the planning keywords of the target object of the long-term plan, and searching for the first planning connection data vector and the second planning connection data vector of the planning keywords according to the retrieval method of the first planning connection and the second planning connection, and performing data fusion to obtain the long-term planning state space of the target object of the long-term plan; obtaining the planning keywords of the target object of the medium-term plan, and searching for the first planning connection data vector and the second planning connection data vector of the planning keywords according to the retrieval method of the first planning connection and the second planning connection, and performing data fusion to obtain the medium-term planning state space of the target object of the medium-term plan; obtaining the planning keywords of the target object of the short-term plan, and searching for the first planning connection data vector and the second planning connection data vector of the planning keywords according to the retrieval method of the first planning connection and the second planning connection, and performing data fusion to obtain the short-term planning state space of the target object of the short-term plan.

[0057] Furthermore, taking the short-term planning state space as an example, the multi-objective planning keyword set of short-term planning includes maximizing the current hourly power generation profit, responding to power grid dispatching instructions, maintaining river ecological flow, ensuring equipment operation safety, etc., finding the relevant first planning connection and second planning connection, constructing the first planning connection data vector and the second planning connection data vector, and directly splicing or encoding and fusing the two vectors to obtain the short-term planning state space of the target object of short-term planning.

[0058] Furthermore, the short-term planning state space has the finest granularity, including real-time, high-frequency data; the medium-term planning state space has a medium granularity, including forecast sequences and planned events; and the long-term planning state space has the coarsest granularity, including aggregated statistics and long-term trends.

[0059] Driven by multi-objective planning keywords, it ensures that the state space of each planning stage only contains the most important and relevant information for its decision-making, avoids stuffing all the original multimodal data into the state space, significantly reduces the invalid dimensions of the state space, and improves the "signal-to-noise ratio" of information; it ensures that the state space of different planning levels not only fully reflects key information but also has appropriate granularity, which is the core component of realizing an intelligent hierarchical planning system based on multimodal data fusion.

[0060] The reinforcement learning algorithm based on multimodal data fusion is used to plan the target objects of long-term planning, medium-term planning and short-term planning in sequence to obtain the corresponding planning strategies.

[0061] Furthermore, using a reinforcement learning algorithm based on multimodal data fusion to sequentially perform long-term planning, mid-term planning, and short-term planning on a target object includes: constructing three levels of intelligent agents, corresponding to long-term planning, mid-term planning, and short-term planning, respectively;

[0062] The first-level agent constructs a reinforcement learning environment based on the long-term planning constraints. The state space of the first-level agent is constructed based on the long-term planning state space of the long-term planning target object. The action space is constructed based on the execution strategy of the target object. The reward function is constructed based on the long-term planning goal. The DDPG algorithm is used to solve the long-term planning execution strategy shown in Table 2.

[0063] The second-level agent builds a reinforcement learning environment based on the medium-term planning constraints and long-term planning execution strategy. The state space of the second-level agent is built based on the medium-term planning state space of the target object of the medium-term planning. The action space is built based on the execution strategy of the target object. The establishment function is built based on the medium-term planning goal and solved using the DDPG algorithm to obtain the execution strategy of the medium-term plan.

[0064] The third-level intelligent agent constructs a reinforcement learning environment based on the short-term planning constraints, the long-term planning execution strategy, and the medium-term planning execution strategy. The state space of the third-level intelligent agent is constructed based on the short-term planning state space of the target object of the short-term planning. The action space is constructed based on the execution strategy of the target object. The reward function is constructed based on the long-term planning goal. The solution is performed according to the DDPG algorithm to obtain the execution strategy of the short-term planning.

[0065] Furthermore, the core idea of ​​this framework is to decompose the complex cross-time-scale hydropower station planning problem into several levels, each of which is managed by a reinforcement learning agent. The decisions of the upper-level agents set goals or constraints for the lower-level agents. The planning state space and planning constraints constructed by multimodal data fusion provide the basis for perception and decision-making for these agents.

[0066] Table 2. Implementation strategies for long-term planning

[0067]

[0068] Furthermore, the first-tier agent focuses on the hydropower station's long-term operational strategy and macro-resource allocation. The long-term planning execution strategy is a macro, highly abstract decision. Built upon long-term planning objectives, it aims to guide the first-tier agent in learning to maximize long-term cumulative benefits. Examples include: cumulative power generation revenue over multiple years (positive rewards), penalties for violating long-term ecological flow or sustainable water resource utilization targets (negative rewards), penalties for major floods or droughts (negative rewards), and rewards for achieving asset preservation and appreciation goals. The long-term planning execution strategy does not directly control medium- or short-term behavior. Instead, it serves as part of the second-tier agent's environment, influencing the goals or constraints of its medium-term plan. For example, an annual average reservoir capacity target translates into monthly / quarterly reservoir capacity ranges to be achieved in the medium-term plan.

[0069] An extreme cross-timescale planning problem involving decades to minutes is decomposed into three relatively independent sub-problems. Each intelligent agent only needs to focus on the dynamics and goals of its specific level, which significantly reduces the complexity of learning. The first-level intelligent agent focuses on dealing with long-term uncertainties such as macroclimate and market; the second-level intelligent agent focuses on dealing with medium-term uncertainties such as seasonal forecast errors and planned maintenance execution; the third-level intelligent agent focuses on dealing with immediate uncertainties such as real-time fluctuations, ultra-short-term weather, and power grid instructions. The hierarchical structure allows the use of the most suitable method for handling uncertainty at different levels at different levels, better dealing with uncertainties at different levels and improving system robustness.

[0070] By comprehensively collecting and mining the complex associations between primary and secondary planning connections among massive multimodal data from hydropower stations, this method can build a profound and multi-dimensional understanding of the system's current state, future trends, and inherent constraints, forming fused planning data with high information density. Based on this fused data, multi-objective planning constraints reflecting physical, environmental, regulatory, and operational constraints at different time scales (long, medium, and short) are dynamically and accurately generated, and a layered, highly correlated hierarchical planning state space is constructed, providing a solid foundation for perception and decision-making for the planning agent. Finally, the hierarchical reinforcement learning algorithm is used instead of traditional mathematical programming to provide a powerful, data-driven, end-to-end solution for solving the cross-timescale, multi-objective, and highly complex planning problems of hydropower stations, which can significantly improve the intelligence level of planning, operational efficiency, and system robustness.

[0071] Example 2:

[0072] A company has introduced a multi-objective planning system for hydropower stations based on multimodal data fusion provided by the present invention to improve the efficiency of resource management of hydropower stations. The system structure is as follows: Figure 3 As shown, the specific implementation is as follows:

[0073] The data acquisition module collects multimodal data of the hydropower station;

[0074] Furthermore, the temporal modal data includes at least load demand, precipitation, and inflow flow; the spatial modal data includes at least upstream and downstream water level distribution, reservoir terrain, and transmission network topology; the structural modal data includes at least equipment status parameters of generator sets and flood discharge gates; and the semantic modal data includes at least operation and scheduling strategy texts, weather forecast reports, and superior scheduling commands.

[0075] The data fusion module mines the first planning connection and the second planning connection between the multimodal data of the hydropower station, performs data fusion based on the first planning connection and the second planning connection, and obtains the fused planning data;

[0076] Furthermore, the first planning connection indicates a strong correlation between data with a clear physical mechanism or direct causality; through physical model analysis, high-frequency data correlation analysis and Granger causality test, a retrieval method for the first planning connection is constructed;

[0077] Furthermore, the second planning connection represents the indirect, lagged, low-intensity correlation between data with statistical laws; time series graphs, scatter plots, and distribution graphs of different modal data are drawn, and the connection between data is analyzed through statistical descriptions and correlation coefficient matrices between variables to construct a retrieval method for the second planning connection.

[0078] Furthermore, through the aforementioned mining methods, we modeled and regularized the statistical, predictive, and indirect relationships identified, creating an index. This secondary planning relationship is stored in the same or another planning relationship knowledge base, using an index based on key attributes such as the name of the variable involved, the target variable for prediction, the relationship type (statistical correlation, time series model, association rule, etc.), the time span (short-term, medium-term, long-term), and the strength of the correlation / level of prediction accuracy. For example, analyzing data from past years, we calculated the correlation between average summer temperature and total inflow runoff from the autumn of the previous year to the spring of the following year, and plotted a scatter plot. We found a certain negative correlation (hotter and drier summers are likely to result in less water inflow later in the year). Table 3 shows the data for average temperature and total inflow runoff, establishing a secondary planning relationship between average temperature and total inflow runoff.

[0079] Table 3. Average temperature and total inflow runoff data

[0080]

[0081] Furthermore, the data fusion includes: obtaining a set of multi-objective planning keywords for the long-term planning, the mid-term planning, and the short-term planning of the hydropower station, and searching for a first planning connection data vector and a second planning connection data vector of the multi-objective planning keywords according to a retrieval method of the first planning connection and the second planning connection;

[0082] The first planning connection data vector and the second planning connection data vector are preprocessed and data fused to obtain fused planning data corresponding to the long-term plan, the medium-term plan and the short-term plan.

[0083] The planning constraint generation module constructs the multi-objective planning constraints of the hydropower station based on the integrated planning data, including long-term planning constraints, medium-term planning constraints and short-term planning constraints;

[0084] Furthermore, constructing multi-objective planning constraints includes: the planning constraint prediction model obtains and processes the fused planning data corresponding to the multi-objective planning keywords of long-term planning, medium-term planning and short-term planning respectively, and the output prediction results are the long-term planning constraints, medium-term planning constraints and short-term planning constraints.

[0085] The planning state space generation module uses the multimodal data of the hydropower station to construct the planning state space of the target object, including long-term planning state space, medium-term planning state space and short-term planning state space;

[0086] Furthermore, the planning state space includes: obtaining the planning keywords of the target object of the long-term plan, and searching for the first planning connection data vector and the second planning connection data vector of the planning keywords according to the retrieval method of the first planning connection and the second planning connection, and performing data fusion to obtain the long-term planning state space of the target object of the long-term plan; obtaining the planning keywords of the target object of the medium-term plan, and searching for the first planning connection data vector and the second planning connection data vector of the planning keywords according to the retrieval method of the first planning connection and the second planning connection, and performing data fusion to obtain the medium-term planning state space of the target object of the medium-term plan; obtaining the planning keywords of the target object of the short-term plan, and searching for the first planning connection data vector and the second planning connection data vector of the planning keywords according to the retrieval method of the first planning connection and the second planning connection, and performing data fusion to obtain the short-term planning state space of the target object of the short-term plan.

[0087] The planning strategy generation module uses the reinforcement learning algorithm of multimodal data fusion to plan the target objects of long-term planning, medium-term planning and short-term planning in sequence to obtain the corresponding planning strategies.

[0088] Furthermore, using a reinforcement learning algorithm based on multimodal data fusion to sequentially perform long-term planning, mid-term planning, and short-term planning on a target object includes: constructing three levels of intelligent agents, corresponding to long-term planning, mid-term planning, and short-term planning, respectively;

[0089] The first-level agent builds a reinforcement learning environment based on the long-term planning constraints. The state space of the first-level agent is built based on the long-term planning state space of the target object. The action space is built based on the execution strategy of the target object. The reward function is built based on the long-term planning goal. The DDPG algorithm is used to solve the long-term planning execution strategy.

[0090] The second-level agent builds a reinforcement learning environment based on the medium-term planning constraints and long-term planning execution strategy. The state space of the second-level agent is built based on the medium-term planning state space of the target object of the medium-term planning. The action space is built based on the execution strategy of the target object. The establishment function is built based on the medium-term planning goal and solved using the DDPG algorithm to obtain the execution strategy of the medium-term plan.

[0091] The third-level intelligent agent constructs a reinforcement learning environment based on the short-term planning constraints, the long-term planning execution strategy, and the medium-term planning execution strategy. The state space of the third-level intelligent agent is constructed based on the short-term planning state space of the target object of the short-term planning. The action space is constructed based on the execution strategy of the target object. The reward function is constructed based on the long-term planning goal. The solution is performed according to the DDPG algorithm to obtain the execution strategy of the short-term planning.

[0092] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A multi-objective planning method for a hydropower station based on multimodal data fusion, characterized in that: include: Collecting multimodal data of a hydropower station, and mining the first planning relationship and the second planning relationship between the multimodal data of the hydropower station; The first planning relationship indicates a high-intensity correlation between data with a clear physical mechanism or direct causality. Through physical model analysis, high-frequency data correlation analysis and Granger causality test, a retrieval method for the first planning relationship is constructed. The second planning relationship indicates an indirect, lagged, and low-intensity correlation between data with statistical regularity. Draw time series graphs, scatter plots, and distribution graphs of different modal data, analyze the relationship between data through statistical descriptions and correlation coefficient matrices between variables, and construct a retrieval method for the second planning relationship; Obtain a set of multi-objective planning keywords for the long-term plan, the medium-term plan, and the short-term plan of the hydropower station, and search for a first planning connection data vector and a second planning connection data vector of the multi-objective planning keywords according to a retrieval method of the first planning connection and the second planning connection; preprocess and fuse the first planning connection data vector and the second planning connection data vector to obtain fused planning data corresponding to the long-term plan, the medium-term plan, and the short-term plan; Construct multi-objective planning constraints for the hydropower station based on the integrated planning data, including long-term planning constraints, medium-term planning constraints, and short-term planning constraints; Use the multimodal data of the hydropower station to construct the planning state space of the target object, including long-term planning state space, medium-term planning state space and short-term planning state space; The reinforcement learning algorithm based on multimodal data fusion is used to plan the target objects of long-term planning, medium-term planning and short-term planning in turn, obtain the corresponding planning strategies, and construct three levels of intelligent agents, corresponding to long-term planning, medium-term planning and short-term planning respectively.

2. The multi-objective planning method for a hydropower station based on multimodal data fusion according to claim 1 is characterized in that: Temporal modal data shall at least include load demand, precipitation, and inflow flow; spatial modal data shall at least include upstream and downstream water level distribution, reservoir terrain, and transmission network topology; structural modal data shall at least include equipment status parameters of generator sets and flood discharge gates; semantic modal data shall at least include operation and scheduling strategy texts, weather forecast reports, and superior scheduling commands.

3. The multi-objective planning method for a hydropower station based on multimodal data fusion according to claim 1 is characterized in that: Constructing multi-objective planning constraints includes: the planning constraint prediction model obtains and processes the fused planning data corresponding to the multi-objective planning keywords of long-term planning, medium-term planning and short-term planning respectively, and the output prediction results are the long-term planning constraints, medium-term planning constraints and short-term planning constraints.

4. The multi-objective planning method for a hydropower station based on multimodal data fusion according to claim 1 is characterized in that: The planning state space includes: obtaining the planning keywords of the target object of the long-term planning, and searching for the first planning connection data vector and the second planning connection data vector of the planning keywords according to the retrieval method of the first planning connection and the second planning connection, and performing data fusion to obtain the long-term planning state space of the target object of the long-term planning; obtaining the planning keywords of the target object of the medium-term planning, and searching for the first planning connection data vector and the second planning connection data vector of the planning keywords according to the retrieval method of the first planning connection and the second planning connection, and performing data fusion to obtain the medium-term planning state space of the target object of the medium-term planning; obtaining the planning keywords of the target object of the short-term planning, and searching for the first planning connection data vector and the second planning connection data vector of the planning keywords according to the retrieval method of the first planning connection and the second planning connection, and performing data fusion to obtain the short-term planning state space of the target object of the short-term planning.

5. The multi-objective planning method for a hydropower station based on multimodal data fusion according to claim 1 is characterized in that: The method utilizes a reinforcement learning algorithm based on multimodal data fusion to sequentially perform long-term planning, mid-term planning, and short-term planning on a target object, including: constructing three levels of intelligent agents, corresponding to long-term planning, mid-term planning, and short-term planning respectively; The first-level agent builds a reinforcement learning environment based on the long-term planning constraints. The state space of the first-level agent is built based on the long-term planning state space of the target object. The action space is built based on the execution strategy of the target object. The reward function is built based on the long-term planning goal. The DDPG algorithm is used to solve the long-term planning execution strategy. The second-level agent builds a reinforcement learning environment based on the medium-term planning constraints and long-term planning execution strategy. The state space of the second-level agent is built based on the medium-term planning state space of the target object of the medium-term planning. The action space is built based on the execution strategy of the target object. The establishment function is built based on the medium-term planning goal and solved using the DDPG algorithm to obtain the execution strategy of the medium-term plan. The third-level intelligent agent constructs a reinforcement learning environment based on the short-term planning constraints, the long-term planning execution strategy, and the medium-term planning execution strategy. The state space of the third-level intelligent agent is constructed based on the short-term planning state space of the target object of the short-term planning. The action space is constructed based on the execution strategy of the target object. The reward function is constructed based on the long-term planning goal. The solution is performed according to the DDPG algorithm to obtain the execution strategy of the short-term planning.

6. A multi-objective planning system for a hydropower station based on multi-modal data fusion, which executes the multi-objective planning method for a hydropower station based on multi-modal data fusion according to claim 1, characterized in that: include: The data acquisition module collects multimodal data of the hydropower station; The data fusion module mines the first planning connection and the second planning connection between the multimodal data of the hydropower station, performs data fusion based on the first planning connection and the second planning connection, and obtains the fused planning data; The planning constraint generation module constructs the multi-objective planning constraints of the hydropower station based on the integrated planning data, including long-term planning constraints, medium-term planning constraints and short-term planning constraints; The planning state space generation module uses the multimodal data of the hydropower station to construct the planning state space of the target object, including long-term planning state space, medium-term planning state space and short-term planning state space; The planning strategy generation module uses the reinforcement learning algorithm of multimodal data fusion to plan the target objects of long-term planning, medium-term planning and short-term planning in sequence to obtain the corresponding planning strategies.

Images