A scene generation method, device and equipment based on source load uncertainty

Abstract

The application discloses a scene generation method and device based on source-load uncertainty and equipment, and relates to the technical field of energy scheduling. The method comprises the following steps: acquiring real-time source-load data of a power system; processing the real-time source-load data based on a target prediction model to obtain a multi-time-scale source-load prediction result, wherein the target prediction model is obtained by compressing the weights of a long short-term memory network prediction model; generating a collaborative scheduling strategy of source-load resources under the multi-time-scale source-load prediction result based on a multi-agent scheduling decision model; and generating a hydropower source-load scene set of the power system at different time stages under the source-load uncertainty based on the collaborative scheduling strategy. In the foregoing manner, the prediction link and the decision optimization link are closely coupled, the multi-time-scale prediction result is used to dynamically guide the reinforcement learning decision, the adaptability of the system to the source-load uncertainty is enhanced, and the organic collaboration of the multi-time-scale and the accurate generation of the uncertainty scene are realized.

Description

Technical Field

[0001] This application relates to the field of energy dispatching technology, and in particular to a method, apparatus and equipment for generating scenarios based on source-load uncertainty. Background Technology

[0002] Currently, the generation of scenarios for power system source-load uncertainty mainly adopts a phased and independent approach: first, deterministic point predictions of power output and load demand are made using time series forecasting models; then, based on the prediction results, stochastic optimization or robust optimization methods are used to generate scheduling schemes; and finally, a limited number of scenarios are generated for risk assessment using methods such as Monte Carlo simulation. In implementation, the prediction module and the decision optimization module are usually designed independently. The prediction module provides input data, while the decision optimization module performs decision analysis based on this data, lacking effective feedback and collaborative learning processes between the two. Furthermore, existing solutions generally employ a fixed-time-scale analysis framework, with short-term, medium-term, and long-term problems handled by different models. Scenarios at different stages are connected by simple boundary conditions, with insufficient consideration for coordination across multiple time scales. There is a lack of effective coupling mechanisms between scenarios at different stages, leading to decision conflicts between short-term scheduling and medium-to-long-term planning, and preventing the achievement of globally optimal resource allocation.

[0003] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main purpose of this application is to provide a method, apparatus and equipment for scene generation based on source load uncertainty, which aims to solve the technical problem in the prior art where the prediction and decision-making links are separated, making it difficult to dynamically adapt to the time-varying characteristics of source load uncertainty, resulting in the scene generation being out of sync with the actual operating conditions.

[0005] To achieve the above objectives, this application provides a scene generation method based on source load uncertainty, the method comprising: Acquire real-time source-load data of the power system; The real-time source load data is processed based on the target prediction model to obtain multi-timescale source load prediction results. The target prediction model is obtained by weight compression of the long short-term memory network prediction model. Based on the multi-agent scheduling decision model, a collaborative scheduling strategy for source load resources is generated under the source load prediction results of the multi-time scale. Based on the aforementioned collaborative scheduling strategy, a set of water and power load scenarios for the power system at different time stages under source-load uncertainty conditions is generated.

[0006] In one embodiment, the step of processing the real-time source load data based on the target prediction model to obtain multi-timescale source load prediction results includes: The preprocessed real-time source-load data is input into the target prediction model for multiple inferences to obtain multiple sets of prediction results at different time scales. The real-time source-load data includes power output data on the power supply side, demand data on the load side, associated meteorological data, and associated hydrological data. Based on the numerical distribution of multiple sets of prediction results at each time scale, the confidence intervals of multiple sets of prediction results at each time scale are determined. Based on the confidence intervals of multiple prediction results at each time scale, the target prediction results and uncertainty prediction intervals at each time scale are determined. Based on the target prediction results and uncertainty prediction intervals at various time scales, multi-time scale source load prediction results are obtained.

[0007] In one embodiment, the step of acquiring real-time source-load data of the power system further includes: Collect historical source load data, and determine training data based on the historical source load data; Based on the training data, a Long Short-Term Memory (LSTM) network prediction model is trained. The key weight matrix is ​​extracted from the Long Short-Term Memory Network prediction model, and the key weight matrix is ​​compressed to obtain the reconstructed weight matrix. Based on the reconstructed weight matrix, the long short-term memory network prediction model is converted into a compressed prediction model; The compressed prediction model is fine-tuned to obtain the target prediction model.

[0008] In one embodiment, the step of compressing the key weight matrix to obtain the reconstructed weight matrix includes: The key weight matrix is ​​reshaped into a tensor form to obtain the key weight tensor; The key weight tensor is subjected to regularized Tucker decomposition to obtain the target core tensor and the target factor matrix; Based on the target core tensor and the target factor matrix, basic trend reconstruction and detailed fluctuation reconstruction are performed to obtain the reconstructed weight tensor; The reconstructed weight tensor is reshaped into a matrix form to obtain the reconstructed weight matrix.

[0009] In one embodiment, the step of performing regularized Tucker decomposition on the key weight tensor to obtain the target core tensor and the target factor matrix includes: A joint optimization objective function is established based on the energy retention ratio, sensitivity-weighted reconstruction error, complexity penalty component, and balance hyperparameter; The joint optimization objective function is iteratively solved to determine the optimal compressed rank combination; Based on time smoothness constraints, core tensor sparsity constraints, and periodic constraints, we determine the hydropower scenario-specific regularization components. Based on the aforementioned key weight tensor and the hydropower scenario-specific regularization component, a regularization decomposition objective function is established. Based on the optimal compressed rank combination, the regularized decomposition objective function is solved iteratively to determine the target core tensor and the target factor matrix.

[0010] In one embodiment, the step of generating a collaborative scheduling strategy for source load resources based on the multi-agent scheduling decision model under the multi-timescale source load prediction results includes: The multi-timescale source-load prediction results are fused with the real-time operating status of the power system to obtain integrated status data; The integrated state data is input into the multi-agent scheduling decision model so that multiple agents in the multi-agent scheduling decision model can make collaborative decisions based on the integrated state data to obtain a collaborative scheduling strategy for source load resources under the source load prediction results of the multi-time scale.

[0011] In one embodiment, the step of generating a set of hydropower load scenarios for the power system at different time stages under source-load uncertainty conditions based on the cooperative scheduling strategy includes: Based on the aforementioned coordinated scheduling strategy, multiple Monte Carlo simulations are performed under source-load uncertainty to obtain the dynamic evolution path of the power system at different time stages under source-load uncertainty. The dynamic evolution path includes changes in power output, fluctuations in load demand, and evolution of grid status. The dynamic evolution path is taken as the corresponding hydropower load scenario, and the occurrence probability of the hydropower load scenario is calculated. Based on the water-power-load scenarios and their occurrence probabilities, a set of water-power-load scenarios for the power system at different time stages under source-load uncertainty is generated.

[0012] In one embodiment, the step of calculating the probability of occurrence of the water-powered load scenario includes: Obtain the scene time span, system state transition probability, and action probability distribution of the cooperative scheduling strategy; Obtain the correspondence between the scene time span, system state transition probability, action probability distribution and the occurrence probability of water and power load scene; Based on the scenario time span, the system state transition probability, the action probability distribution of the collaborative scheduling strategy, and the corresponding relationship, the probability of occurrence of the water and power load scenario is obtained.

[0013] Furthermore, to achieve the above objectives, this application also proposes a scene generation device based on source load uncertainty, which includes: The data prediction module is used to acquire real-time source-load data of the power system; The data prediction module is also used to process the real-time source load data based on the target prediction model to obtain multi-timescale source load prediction results. The target prediction model is obtained by weight compression of the long short-term memory network prediction model. The collaborative decision-making module, based on a multi-agent scheduling decision-making model, generates a collaborative scheduling strategy for source load resources under the source load prediction results at the multiple time scales. The scenario generation module is used to generate a set of water and power load scenarios of the power system at different time stages under the condition of source and load uncertainty, based on the cooperative scheduling strategy.

[0014] In addition, to achieve the above objectives, this application also proposes a scene generation device based on source load uncertainty. The scene generation device based on source load uncertainty includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. The computer program is configured to implement the steps of the scene generation method based on source load uncertainty as described above.

[0015] In addition, to achieve the above objectives, the present invention also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the scene generation method based on source load uncertainty as described above.

[0016] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the scene generation method based on source load uncertainty as described above.

[0017] This application provides a scenario generation method based on source-load uncertainty. The method acquires real-time source-load data of the power system; processes the real-time source-load data based on a target prediction model to obtain multi-timescale source-load prediction results. The target prediction model is obtained by weight compression of a long short-term memory network prediction model. Based on a multi-agent scheduling decision model, a collaborative scheduling strategy for source-load resources under the multi-timescale source-load prediction results is generated. Based on the collaborative scheduling strategy, a set of hydropower and power load scenarios of the power system at different time stages under source-load uncertainty is generated. Through this approach, the prediction and decision optimization stages are tightly coupled. The multi-timescale prediction results dynamically guide reinforcement learning decisions, enhancing the system's adaptability to source-load uncertainty, avoiding conflicts between decisions at different timescales, effectively mitigating operational risks caused by source-load fluctuations, achieving organic collaboration across multiple timescales and accurate generation of uncertain scenarios. Simultaneously, it achieves efficient and lightweight prediction models, significantly reducing computational resource requirements while maintaining prediction accuracy. This solves the technical problem of the separation of prediction and decision stages, making it difficult to dynamically adapt to the time-varying characteristics of source-load uncertainty, leading to a disconnect between scenario generation and actual operating conditions. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating an embodiment of the scene generation method based on source load uncertainty in this application. Figure 2 This is a flowchart illustrating Embodiment 2 of the scene generation method based on source load uncertainty in this application; Figure 3 This is a flowchart illustrating Embodiment 3 of the scenario generation method based on source load uncertainty in this application; Figure 4 This is a schematic diagram of the module structure of the scene generation device based on source load uncertainty in an embodiment of this application; Figure 5 This is a schematic diagram of the device structure of the hardware operating environment involved in the scene generation method based on source load uncertainty in the embodiments of this application.

[0021] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0023] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0024] The main solution of this application embodiment is as follows: acquiring real-time source-load data of the power system; processing the real-time source-load data based on the target prediction model to obtain multi-time-scale source-load prediction results, wherein the target prediction model is obtained by weight compression of the long short-term memory network prediction model; generating a collaborative scheduling strategy for source-load resources under the multi-time-scale source-load prediction results based on a multi-agent scheduling decision model; and generating a set of hydropower and power load scenarios of the power system at different time stages under source-load uncertainty conditions based on the collaborative scheduling strategy.

[0025] This application provides a solution for acquiring real-time source-load data of a power system; processing the real-time source-load data based on a target prediction model to obtain multi-timescale source-load prediction results, the target prediction model being obtained by weight compression of a long short-term memory network prediction model; generating a collaborative scheduling strategy for source-load resources under the multi-timescale source-load prediction results based on a multi-agent scheduling decision model; and generating a set of hydropower and power load scenarios for the power system at different time stages under source-load uncertainty conditions based on the collaborative scheduling strategy. Through this approach, the prediction and decision optimization stages are tightly coupled, and the multi-timescale prediction results are used to dynamically guide reinforcement learning decisions, enhancing the system's adaptability to source-load uncertainty, avoiding conflicts between decisions at different time scales, effectively mitigating operational risks caused by source-load fluctuations, achieving organic collaboration across multiple time scales and accurate generation of uncertain scenarios. Simultaneously, the prediction model is made highly efficient and lightweight, significantly reducing computational resource requirements while maintaining prediction accuracy. This solves the technical problem of the separation of the prediction and decision stages, making it difficult to dynamically adapt to the time-varying characteristics of source-load uncertainty, leading to a disconnect between scenario generation and actual operating conditions.

[0026] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of performing the above functions, a scene generation device based on source-load uncertainty, etc. This embodiment does not specifically limit it. The following uses a scene generation device based on source-load uncertainty as an example to describe this embodiment and the following embodiments.

[0027] This application provides a scene generation method based on source load uncertainty, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the scene generation method based on source load uncertainty in this application.

[0028] In this embodiment, the scene generation method based on source load uncertainty includes steps S10~S40: Step S10: Obtain real-time source-load data of the power system; It should be noted that a power system generally refers to a unified whole that integrates power generation, transmission, transformation, distribution, and consumption equipment, along with corresponding auxiliary systems, to achieve a continuous and reliable supply of electricity and real-time supply-demand balance. Real-time source-load data refers to data related to the power source and load sides collected in real time, reflecting the current and recent operating status of the power system. Real-time source-load data typically originates from the power system's operation monitoring platform, meteorological and hydrological monitoring networks, and user-side measurement equipment, and can be collected through real-time data interfaces.

[0029] Additionally, it should be noted that real-time source-load data includes power generation output data, load demand data, and related meteorological and hydrological data. Power generation output data includes hydropower output data, thermal power output data, and renewable energy output data. Specifically, hydropower output data includes the current real-time active power output of hydropower units, the current water level of reservoirs, and recent (e.g., within the last 3 days) inflow. Thermal power output data includes the current actual power generation of thermal power units (e.g., coal-fired, gas-fired, etc.) and their operating status. Renewable energy output data includes the current actual power generation of wind farms and photovoltaic power plants. Load demand data refers to the current total load value of the power grid, which can be further subdivided into regional load, key industry load, etc. The associated meteorological data includes meteorological data related to new energy sources and regional meteorological data for large-scale power systems. These are used to assess the comprehensive impact on load and hydrological conditions. Meteorological data related to new energy sources refers to real-time meteorological data corresponding to the location of power stations (wind farms and photovoltaic power stations), including wind speed, wind direction, solar irradiance, and ambient temperature. Regional meteorological data for large-scale power systems refers to meteorological data within a large area spanning multiple regions, interconnected power grids, and multiple power source clusters, such as precipitation, temperature, and humidity. Associated hydrological data mainly includes real-time runoff and water level information from key monitoring stations affecting hydropower output.

[0030] Understandably, real-time source and load data can be acquired through automated data acquisition protocols, continuously pulling or receiving data streams from the data source at a frequency of minutes or hours, ensuring that the real-time source and load data can reflect the latest status of the power system.

[0031] Step S20: Process the real-time source load data based on the target prediction model to obtain multi-timescale source load prediction results. The target prediction model is obtained by weight compression of the long short-term memory network prediction model. It should be noted that the Long Short-Term Memory (LSTM) network prediction model is the baseline prediction model trained on the traditional Long Short-Term Memory (LSTM) network. The target prediction model (T-LSTM model) is the efficient prediction model obtained by compressing and optimizing the weights of the LSTM prediction model.

[0032] It is understood that the multi-timescale source load prediction results refer to the prediction results at multiple time scales. In this embodiment, the time scales include short-term (1-7 days, hour level), medium-term (1-4 weeks, day level) and long-term (1-12 months, month level). At this time, the multi-timescale source load prediction results include short-term prediction results (short-term prediction sequence), medium-term prediction results (medium-term prediction trend) and long-term prediction results (long-term prediction profile).

[0033] In one feasible implementation, step S20 may include steps S201 to S204: Step S201: Input the preprocessed real-time source load data into the target prediction model for multiple inferences to obtain multiple sets of prediction results at different time scales. It should be noted that preprocessing includes at least data cleaning and standardization. Data cleaning includes handling missing values, outlier handling, and noise filtering, while standardization includes normalization and feature alignment. After preprocessing, a structured and standardized real-time data snapshot or short-term series can be formed, meeting the input dimension and format requirements of the target prediction model, and can be used as input to the target prediction model. Based on the preprocessed real-time source load data, a multivariate time series is formed and input into the target prediction model.

[0034] Understandably, in order to quantify the uncertainty of the prediction, Dropout is kept active during the inference phase and multiple random forward propagations are performed (e.g., 50 times, 100 times). In other words, the target prediction model is used to perform multiple inferences on the preprocessed real-time source load data. Each inference can obtain a set of prediction results at various time scales, and the prediction results are the corresponding prediction values.

[0035] It should be understood that the target forecasting model can generate forecasts at multiple time scales in parallel. Short-term forecasts are generated using a rolling forecasting strategy, using the forecast from the previous moment as input for the next moment to gradually generate a detailed forecast sequence for the next 168 hours; medium-term forecasts use a sequence-to-sequence forecasting architecture to directly output daily forecasts for the next 28 days; long-term forecasts require combining direct forecasts with trend extrapolation to generate monthly forecast trends for the next 12 months.

[0036] Step S202: Based on the numerical distribution of multiple sets of prediction results at each time scale, determine the confidence interval of multiple sets of prediction results at each time scale. Understandably, based on multiple sets of prediction results obtained at various time scales through multiple inferences, the numerical distribution of the multiple sets of prediction results at various time scales is determined, and the confidence interval of the multiple sets of prediction results at various time scales is determined according to the numerical distribution of the multiple sets of prediction results at various time scales.

[0037] Step S203: Based on the confidence intervals of multiple sets of prediction results at each time scale, determine the target prediction results and uncertainty prediction intervals at each time scale. It should be noted that the uncertainty prediction interval is a numerical range set to quantify the uncertainty of the prediction. Its lower and upper limits can be selected based on the confidence interval. For example, assuming that the confidence level of the confidence interval is 95%, the 2.5% quantile is selected as the lower limit of the uncertainty prediction interval, and the 97.5% quantile is selected as the upper limit of the uncertainty prediction interval. This embodiment does not make specific limitations on this.

[0038] It is understood that the target prediction result is the final prediction value determined based on multiple sets of prediction results. The specific value can be calculated / selected according to the confidence interval. For example, the mean of multiple sets of prediction results within the confidence interval at each time scale is taken as the target prediction result at each time scale. This embodiment does not make specific limitations on this.

[0039] Step S204: Based on the target prediction results at each time scale and the uncertainty prediction interval corresponding to the target prediction results, obtain the multi-time scale source load prediction results.

[0040] It should be noted that the final output can include multi-timescale source load prediction results that include the target prediction results and the uncertainty prediction interval.

[0041] Step S30: Based on the multi-agent scheduling decision model, generate a collaborative scheduling strategy for source load resources under the source load prediction results of the multi-time scale; It should be noted that the Multi-Agent Scheduling Decision Model (MARL model) includes multiple agents, including hydropower scheduling agents, thermal power scheduling agents, renewable energy scheduling agents, energy storage scheduling agents, and load management agents. Source and load resources refer to resources on the power supply side and the load side. Power supply side resources include at least hydropower, thermal power, wind power, photovoltaic, and energy storage resources. Different agents are responsible for scheduling / regulating / optimizing different source and load resources. The hydropower scheduling agent is responsible for optimizing the scheduling of hydropower resources, considering water uncertainty, reservoir constraints, and cross-cycle regulation capabilities. The thermal power scheduling agent is responsible for regulating thermal power resources, providing rapid backup and balancing services to compensate for fluctuations in hydropower and renewable energy. The renewable energy scheduling agent is responsible for the absorption management of wind and photovoltaic resources, balancing prediction errors and volatility. The energy storage scheduling agent is responsible for the charging and discharging scheduling of energy storage systems (energy storage resource scheduling), providing short-term balancing and frequency support. The load management agent is responsible for scheduling load-side resources, optimizing the system load curve through demand response.

[0042] Additionally, it should be noted that the environment (environment model) of the multi-agent scheduling decision-making model is constructed based on the actual operating conditions of the power system and integrates the prediction results of the T-LSTM model in a forward-looking manner, thereby simulating a decision-making environment containing future uncertainties. The core components of the environment include the static physical architecture, operational constraints, and external factors. The static physical architecture includes the power source side (operating characteristics and constraints of hydropower units, thermal power units, wind farms, and photovoltaic power plants), the load side (demand characteristics and adjustability of industrial, commercial, and residential loads), the power grid topology (models and transmission capacity constraints of power grid components such as transmission lines, transformers, and node voltages), and reservoir characteristics (reservoir capacity curves, water balance constraints, and ecological flow requirements). Operational constraints include real-time updated line transmission capacity, node voltage safety limits, and spinning reserve requirements. External factors include real-time meteorological conditions, hydrological situations, and their future forecasts.

[0043] It is understandable that the state space of a multi-agent scheduling decision model contains comprehensive operational information of the power system and can be represented as a multi-dimensional vector. It comprises four dimensions: source load prediction information, real-time operating status, time context information, and external environment status. Source load prediction information refers to the multi-timescale source load prediction results output by the T-LSTM model; real-time operating status comes from real-time measurements of the power grid monitoring system, describing the actual operating profile of the power system at the current moment and serving as the physical basis for decision-making, including generator status, load status, grid status, and reservoir status; time context information is determined by the system clock and scheduling plan, including time characteristics and prediction time scale identifiers; and external environment status comes from real-time data from meteorological and hydrological monitoring stations, including meteorological and hydrological data.

[0044] It should be understood that the action space of a multi-agent scheduling decision model defines the operations that each agent can perform, and it can usually be divided based on the scheduling dimension of the power system. The action space of a hydropower scheduling agent includes hydropower unit output setpoints, reservoir water release decisions, and reservoir water level targets; the action space of a thermal power scheduling agent includes thermal power unit start-up and shutdown decisions, thermal power unit output levels, and spinning reserve capacity; the action space of a renewable energy scheduling agent includes wind and solar curtailment ratios and predicted utilization rates; the action space of an energy storage scheduling agent includes charging and discharging power and state of charge targets; and the action space of a load management agent includes demand response commands and interruptible load control. The action space of each agent can be designed according to the characteristics of the resources it controls, including both continuous and discrete actions.

[0045] Understandably, the reward function guides the agent in learning the optimal strategy, comprehensively considering multiple objectives such as economy, safety, clean energy utilization, and supply-demand balance. In specific implementations, the reward function can be determined based on cost rewards, safety rewards, clean energy utilization rewards, supply-demand balance rewards, and prediction accuracy rewards.

[0046] Cost incentives are related to power generation costs (including operating costs of hydropower and thermal power), start-up and shutdown costs of generating units (hydropower units and thermal power units), and reserve capacity costs. The calculation formula is shown below:

[0047] In the formula, Indicates cost incentive, Indicates the cost of electricity generation. This indicates the cost of starting and stopping the generator unit. This represents the cost of spare capacity.

[0048] The safety bonus is related to the degree of voltage exceedance, frequency deviation, and line overload, and the calculation formula is shown below:

[0049] In the formula, Indicates a security reward. Indicates the degree of voltage exceeding the limit. Indicates the degree of frequency deviation. Indicates the degree of line overload. , , These are all corresponding weighting coefficients.

[0050] The clean energy consumption incentive is related to the clean energy consumption incentive and the actual amount of new energy consumed. The calculation formula is shown below:

[0051] In the formula, Indicating rewards for clean energy consumption, This indicates the actual amount of new energy consumed. This indicates the amount of electricity wasted from renewable energy sources. , These represent the reward coefficient and the penalty coefficient, respectively.

[0052] The supply-demand balance bonus is related to the total generating capacity and total load capacity of the power system, and the calculation formula is shown below:

[0053] In the formula, Indicates a supply-demand balance reward. This represents the balance penalty coefficient. This represents the total generating capacity of the power system. This represents the total load power of the power system.

[0054] The reward for prediction accuracy is related to the mean absolute error (MAE), which measures the deviation between the predicted data on which the agent's decision is based and the actual value. The formula for calculating the reward for prediction accuracy is as follows:

[0055] In the formula, Rewards are given for accurate predictions. This represents the penalty coefficient for prediction accuracy. It represents the mean absolute error.

[0056] It should be understood that the final reward function is a weighted sum of cost incentives, safety incentives, clean energy consumption incentives, supply and demand balance incentives, and forecast accuracy incentives, as shown below:

[0057] In the formula, Indicates cost incentive, Indicates a security reward. Indicating rewards for clean energy consumption, Indicates a supply-demand balance reward. Rewards are given for accurate predictions. , , , , These represent the corresponding weight coefficients, which can be adjusted through training to balance the various objectives. The corresponding reward can be calculated based on the reward function.

[0058] It should be noted that this embodiment uses historical operating data of the power system as the data for training the multi-agent scheduling decision model. It typically needs to cover a time span of 3-5 years, including historical source and load data, historical system operating data, historical meteorological and hydrological data, and historical scheduling decision data.

[0059] It is understood that this embodiment employs a multi-agent deep deterministic policy gradient algorithm, which is applicable to continuous action spaces and can handle cooperation between agents. Each agent has an independent actor network and critic network. The critic network can access the action and state information of all agents, promoting collaborative learning.

[0060] It should be understood that the training process of a multi-agent scheduling decision model includes the following elements: experience replay, target network, and exploration strategy. Experience replay: storing the experience tuples of the agents. ,in For state, For joint operations, As a reward, The next state is defined as follows: Target Network: A target actor network and a target critic network are used, updated periodically to stabilize training. Exploration Policy: Noise is added to the actor network output to encourage exploration, gradually decreasing as training progresses. In the implementation, the network parameters and experience replay buffers of all agents are first initialized, and then the model performance is gradually improved through multiple rounds of training. In each round of training, the agent interacts with the environment based on the current policy, collects state transition experience, and stores it in the experience replay buffer. Subsequently, batch data is sampled from the experience replay buffer, and the critic network and actor network are updated sequentially. The critic network improves the value estimation accuracy by minimizing temporal difference error, while the actor network optimizes decision-making performance through policy gradient ascent. The target network mechanism is used during training to ensure learning stability, and the relationship between exploration and utilization is balanced by dynamically adjusting the exploration noise.

[0061] Understandably, during training, the model performance is periodically evaluated using a validation set to monitor the trends of various metrics. An early stopping mechanism prevents overfitting, and a grid search approach is used to optimize hyperparameter configuration, including learning rate, discount factor, and exploration parameters. The trained multi-agent scheduling decision model needs to be evaluated using multi-dimensional metrics, including economic metrics (e.g., total scheduling cost), safety metrics (e.g., number of system constraint violations), and cleanliness metrics (e.g., renewable energy absorption rate). Simultaneously, the adaptability of the multi-agent scheduling decision model under source-load uncertainty conditions is assessed to ensure that the final generated scheduling strategy combines safety, economy, and cleanliness.

[0062] In one feasible implementation, step S30 may include steps S301 to S302: Step S301: The multi-timescale source-load prediction results are fused with the real-time operating status of the power system to obtain integrated status data; It should be noted that the multi-timescale source-load prediction results output by the T-LSTM model are fused with the real-time operating status of the power system (generator status, load status, grid status, and reservoir status) to construct complete state data, i.e., integrated state data. The integrated state data includes short-term accurate forecasts for the next seven days on an hourly basis, medium-term trend forecasts for the next four weeks on a daily basis, and long-term profile forecasts for the next twelve months on a monthly basis, while also incorporating real-time operating data of generators, loads, grids, and reservoirs.

[0063] Step S302: Input the integrated state data into the multi-agent scheduling decision model so that multiple agents in the multi-agent scheduling decision model can make collaborative decisions based on the integrated state data to obtain a collaborative scheduling strategy for source load resources under the source load prediction results of the multi-time scale.

[0064] Understandably, five specialized intelligent agents make decisions in parallel based on integrated state data. Specifically, the hydropower dispatching agent optimizes hydropower output plans, the thermal power dispatching agent adjusts the operating status of thermal power units, the renewable energy dispatching agent manages wind and solar power consumption strategies, the energy storage dispatching agent controls energy storage charging and discharging behavior, and the load management agent implements demand response measures. These agents implicitly coordinate through shared integrated state data and a network of commentators, ensuring that their decisions not only optimize their own objectives but also promote overall operational efficiency. The collaborative dispatching strategy is the joint dispatching strategy generated collaboratively by the agents.

[0065] It should be understood that the environment model can receive the multi-timescale source-load prediction results output by the T-LSTM model and provide them as part of the state to all agents. The agents generate a collaborative scheduling strategy based on the integrated state data. The environment model simulates the state transition of the power system according to the agents' collaborative scheduling strategy and provides corresponding rewards (reward signals). Based on the rewards calculated by the environment model, the agents continuously adjust their strategies through an online learning mechanism to adapt to the dynamic changes in source-load uncertainty. This embodiment deeply embeds the multi-timescale source-load prediction results into the environment state, allowing the agents to see a predicted future scenario at each decision step, thereby learning to optimize scheduling strategies not only for the current moment but also for multiple future cycles.

[0066] Step S40: Based on the cooperative scheduling strategy, generate a set of water and power load scenarios of the power system at different time stages under source-load uncertainty conditions.

[0067] In this embodiment, the multi-timescale source-load prediction results are input into a pre-trained multi-agent scheduling decision model. Multiple agents in the multi-agent scheduling decision model make collaborative decisions and learn, and simulate the dynamic evolution of the power system at different time stages under source-load uncertainty, ultimately generating a multi-stage hydropower load scenario.

[0068] It should be noted that, based on the time scale corresponding to the source load prediction results, the time phase may include: short-term phase (within the next seven days), medium-term phase (within the next four weeks), and long-term phase (within the next twelve months). This embodiment does not make specific limitations on this.

[0069] Understandably, this embodiment transforms the cooperative scheduling strategy generated by the intelligent agent into a multi-stage set of hydropower and power load scenarios with probabilistic significance. Based on stochastic optimization theory, the deterministic scheduling strategy is extended into a set of probabilistic scenarios, fully characterizing the impact of source-load uncertainty. These sets of hydropower and power load scenarios can be used for power system scheduling decisions and risk assessment.

[0070] This embodiment provides a scenario generation method based on source-load uncertainty. It acquires real-time source-load data of the power system; processes the real-time source-load data based on a target prediction model to obtain multi-timescale source-load prediction results (the target prediction model is obtained by weight compression of an LSTM prediction model); generates a collaborative scheduling strategy for source-load resources under the multi-timescale source-load prediction results based on a multi-agent scheduling decision model; and generates a set of hydropower and power load scenarios for the power system at different time stages under source-load uncertainty conditions based on the collaborative scheduling strategy. This embodiment tightly couples the prediction and decision optimization stages, dynamically guiding reinforcement learning decisions using multi-timescale prediction results. This enhances the system's adaptability to source-load uncertainty, avoids conflicts between decisions at different timescales, effectively mitigates operational risks caused by source-load fluctuations, achieves organic collaboration across multiple timescales and accurate generation of uncertain scenarios, and simultaneously realizes efficient and lightweight prediction models, significantly reducing computational resource requirements while ensuring prediction accuracy.

[0071] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 Steps S01 to S05 may be included before step S10: Step S01: Collect historical source load data, and determine training data based on the historical source load data; It should be noted that historical source-load data refers to historical data related to both the power supply and load sides over a long period, typically covering 3 to 5 years or longer, to ensure that the historical source-load data can include complete seasonal cycles and various extreme events. Historical source-load data includes historical power output data on the power supply side (time-series output records of hydropower, thermal power, wind power, and photovoltaic power), historical demand data on the load side (load records of the entire grid and by region), historical meteorological data (refined meteorological data contemporaneous with both the power supply and load sides), and historical hydrological data (refined hydrological data contemporaneous with both the power supply and load sides).

[0072] Understandably, the acquired multi-source historical load data are aligned along the timeline to construct a unified multivariate time series dataset. This multivariate time series dataset is then divided into training, validation, and test sets in chronological order. The training, validation, and test sets are then independently standardized using the same parameters to obtain the corresponding training, validation, and test data.

[0073] Step S02: Based on the training data, a long short-term memory network prediction model is trained. It should be noted that one or more deep LSTM networks are constructed for different prediction targets. A deep LSTM network can contain an input layer, hidden layers (1 to 3 layers), and an output layer (fully connected layer). The hidden layers are typically composed of LSTM units, which include input gates, forget gates, output gates, and candidate cell states, used to learn and remember long-term dependencies in the time series. Furthermore, hyperparameters such as the input time window length and hidden layer dimensions must be set in advance.

[0074] Understandably, the deep LSTM network is trained using training data with the goal of minimizing prediction error, employing a time backpropagation algorithm and a gradient descent optimizer to obtain the LSTM prediction model. During training, validation data is used to monitor performance, and overfitting is prevented through strategies such as early stopping. After training, the model's performance is evaluated on test data, which serves as the accuracy benchmark to be recovered or even surpassed after compression.

[0075] Step S03: Extract the key weight matrix from the Long Short-Term Memory Network prediction model, compress the key weight matrix, and obtain the reconstructed weight matrix. Understandably, this embodiment uses Tucker decomposition to compress the pre-trained LSTM prediction model, transforming it into a T-LSTM model with fewer parameters and higher computational efficiency. Considering the multi-timescale characteristics of hydropower load prediction, this ensures that key patterns and features for predicting source load uncertainties are preserved to the maximum extent possible while compressing the model.

[0076] In one feasible implementation, the step of compressing the key weight matrix to obtain the reconstructed weight matrix may include steps A11-A14: Step A11: Reshape the key weight matrix into tensor form to obtain the key weight tensor; It should be noted that the key weight matrix refers to the weight matrix from the input layer to the hidden layer. and the weight matrix from hidden layer to hidden layer ,in, For the hidden layer dimension, For the dimensions of the input data, This represents the four gates of the corresponding LSTM unit: input gate, forget gate, output gate, and candidate cell state.

[0077] It is understandable that the weight matrix from the input layer to the hidden layer... Remodeled into a three-dimensional tensor The first dimension is the gating dimension, with a size of 4, and the second dimension is the hidden state dimension, with a size of [missing value]. The third dimension is the input feature dimension, with a size of [value missing]. Correspondingly, the weight matrix from hidden layer to hidden layer... Remodeled into a three-dimensional tensor At this point, the key weight tensor is... and .

[0078] Step A12: Perform regularized Tucker decomposition on the key weight tensor to obtain the target core tensor and the target factor matrix; In one feasible implementation, step A12 may include: establishing a joint optimization objective function based on the energy retention ratio, sensitivity-weighted reconstruction error, complexity penalty component, and balance hyperparameter; iteratively solving the established joint optimization objective function to determine the optimal compressed rank combination; determining the hydropower scenario-specific regularization component based on time smoothness constraints, core tensor sparsity constraints, and periodic constraints; establishing a regularization decomposition objective function based on the key weight tensor and the hydropower scenario-specific regularization component; and iteratively solving the regularization decomposition objective function based on the optimal compressed rank combination to determine the target core tensor and target factor matrix.

[0079] It should be noted that in this embodiment, Tucker decomposition is performed on each key weight tensor, and the optimal compression rank is determined by the joint optimization criterion of energy and sensitivity, thereby obtaining the optimal compression rank combination, which maximizes the compression effect while ensuring accuracy.

[0080] It is understandable that the process of finding the optimal compression rank can be considered as the process of solving the joint optimization objective function. The joint optimization objective function is related to the energy retention ratio, sensitivity-weighted reconstruction error, and complexity penalty components, as shown below:

[0081] In the formula, , , This represents the balancing hyperparameters, used to control the relative importance of each item. Indicates the energy retention ratio. This represents the sensitivity-weighted reconstruction error. Represents the complexity penalty component. , , The three compressed ranks obtained from the Tucker decomposition are the target variables that need to be optimized.

[0082] The energy retention ratio measures the proportion of the original tensor information retained in the reconstructed tensor. The calculation formula is shown below:

[0083] In the formula, Indicates the energy retention ratio. Let Frobenius be the norm, which represents the square root of the sum of squares of all elements of a tensor. This is the reconstructed tensor.

[0084] The formula for calculating the sensitivity-weighted reconstruction error is shown below:

[0085] In the formula, This represents the sensitivity-weighted reconstruction error. This is the weight sensitivity matrix. , , , The corresponding weighting coefficients can be determined through cross-validation. Mean squared error, used to measure prediction accuracy. To predict the probability of interval coverage, used to measure the quality of interval prediction. The probability scores are continuously graded and used to comprehensively evaluate the performance of probability prediction. This represents the original tensor, i.e., the key weight tensor. This represents the reconstructed tensor. , , The three dimensions represent the key weight tensor.

[0086] The complexity penalty component is used to measure model complexity and can promote the selection of more compact rank combinations. The calculation relationship is shown below:

[0087] In the formula, Represents the complexity penalty component. , , The three compressed ranks obtained from Tucker decomposition. , , The three dimensions represent the key weight tensor.

[0088] It should be understood that a non-dominated sorting genetic algorithm is used to solve the joint optimization objective function to obtain the Pareto optimal solution, from which the final rank combination is selected based on actual deployment requirements. As the optimal compressed rank combination.

[0089] It should be noted that this embodiment performs regularized Tucker decomposition according to the optimal compressed rank combination, and incorporates prior knowledge of the hydropower scenario into the decomposition process to improve the domain adaptability of the decomposed model.

[0090] Understandably, the regularization decomposition objective function is the same as the regularized Tucker decomposition objective function. Solving the regularization decomposition objective function yields the final required core tensor and factor matrix, i.e., the target core tensor and target factor matrix. The regularization decomposition objective function is shown below:

[0091] In the formula, It is the Frobenius norm. For the reconstructed tensor, For hydropower scenario-specific regularization components, For the core tensor, , , For factor matrix, This is the original tensor, i.e., the key weight tensor.

[0092] It should be understood that the hydropower scenario-specific regularization component is determined based on time smoothness constraints, core tensor sparsity constraints, and periodic constraints, and the calculation relationship is shown below:

[0093] In the formula, This represents a regularization component specific to hydropower scenarios. This represents the time smoothness constraint. This represents the sparsity constraint of the core tensor. Indicates periodic constraints. , , This represents the regularization coefficient.

[0094] Time smoothness constraints can promote smooth changes in the factor matrix over time, and the calculation relationship is shown below:

[0095] In the formula, This represents the time smoothness constraint. Factor matrix The List, This is a time difference operator used to calculate the changes between adjacent time steps.

[0096] Core tensor sparsity constraints are used to promote the sparsity of the core tensor to automatically select important feature combinations. The computational relation is shown below:

[0097] In the formula, This represents the sparsity constraint of the core tensor. For the core tensor, , , This represents the optimal compression rank obtained from the Tucker decomposition.

[0098] Periodic constraints can reflect the periodic pattern of hydropower output, and the calculation relationship is shown below:

[0099] In the formula, Indicates periodic constraints. Representing vectors Circular shift One location, The cycle length is set according to the hydrological characteristics. Factor matrix The List.

[0100] Understandably, by introducing domain-specific regularization constraints, the compressed prediction model can better adapt to the temporal characteristics of hydropower load data, thereby improving prediction accuracy and robustness.

[0101] Step A13: Based on the target core tensor and the target factor matrix, perform basic trend reconstruction and detailed fluctuation reconstruction to obtain the reconstructed weight tensor; It should be noted that the weight tensor is reconstructed in stages, prioritizing the accurate restoration of the basic trend before supplementing with detailed information. The final reconstructed tensor is the reconstructed weight tensor.

[0102] Understandably, the basic trend reconstruction uses the following calculation formula:

[0103] In the formula, the subscript This indicates the core component with the highest retained energy, determined by eigenvalue truncation, and the eigencomponent that contributes more than 90% of the cumulative energy. , , , The core tensor and factor matrix are truncated. This represents the tensor after reconstruction based on the basic trend.

[0104] It should be understood that the detailed fluctuation reconstruction uses the following calculation formula:

[0105] In the formula, the subscript This represents the detail component, which contains truncated high-frequency information. , , , For the remaining feature components, This represents the tensor after reconstruction following the detail fluctuations. This is the final reconstructed weight tensor. .

[0106] Step A14: Reconstruct the reconstructed weight tensor into a matrix form to obtain the reconstructed weight matrix.

[0107] It is understandable that the reconstructed weight tensor corresponds to the weight matrix from the input layer to the hidden layer. Remodeling into a reconstructed weight matrix Reconstruct the weight tensor corresponding to the weight matrix from hidden layer to hidden layer. Remodeling into a reconstructed weight matrix Create a new LSTM network structure, maintaining the original network's number of layers and connection method, and reconstruct the weight matrix. and Initialize the corresponding weight matrix, keeping the bias term and other parameters unchanged or scaling them by the same proportion (during this process, it is necessary to verify the dimensionality matching of the reconstructed weights). Through tensor reshaping, convert the decomposition results back to the standard LSTM weight format, ensuring that the T-LSTM model can directly replace the original LSTM model.

[0108] Step S04: Based on the reconstructed weight matrix, the long short-term memory network prediction model is converted into a compressed prediction model; Step S05: Fine-tune the compressed prediction model to obtain the target prediction model.

[0109] Understandably, since Tucker decomposition is a lossy compression method, the accuracy of the directly obtained T-LSTM model will decrease, requiring fine-tuning to restore performance. The fine-tuning process uses the same training and validation data as the LSTM prediction model, employing a small learning rate, and retraining on the compressed baseline LSTM prediction model for a few epochs. An early stopping strategy is used during fine-tuning to prevent overfitting, and the performance of the fine-tuned model is validated on test data to ensure it reaches or approaches the prediction accuracy of the baseline LSTM prediction model. After fine-tuning, the final target prediction model is obtained, which combines high prediction accuracy with low computational complexity, making it suitable for deployment in real-world power system environments.

[0110] In this embodiment, the flat, large weight matrix is ​​reshaped into a high-order weight tensor. The original weight tensor is approximated by the product of a small core tensor and a set of factor matrices. Model compression is achieved by significantly reducing the dimensionality of the core tensor, resulting in an efficient, accurate, and lightweight prediction model.

[0111] This embodiment provides a scene generation method based on source load uncertainty. It collects historical source load data and determines training data based on this data. An LSTM prediction model is trained using this training data. Key weight matrices are extracted from the LSTM prediction model and compressed to obtain a reconstructed weight matrix. Based on the reconstructed weight matrix, the LSTM prediction model is converted into a compressed prediction model. The compressed prediction model is then fine-tuned to obtain the target prediction model. This embodiment constructs a T-LSTM model with fewer parameters, higher computational efficiency, and maintained accuracy by performing tensor reshaping and Tucker decomposition on the key weight matrix. This achieves efficient and lightweight prediction model design, significantly reducing computational resource requirements while ensuring prediction accuracy.

[0112] Based on the above embodiments of this application, in the third embodiment of this application, the same or similar content as the above embodiments can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 3 Step S40 may include steps S401 to S403: Step S401: Based on the cooperative scheduling strategy, perform multiple Monte Carlo simulations under source-load uncertainty to obtain the dynamic evolution path of the power system at different time stages under source-load uncertainty. The dynamic evolution path includes changes in power output, fluctuations in load demand, and evolution of grid status. It should be noted that, by utilizing a coordinated scheduling strategy and considering the uncertainty of power source and load, multiple Monte Carlo simulations are conducted. Each simulation is based on different initial conditions and random seeds, resulting in unique scenario trajectories that all conform to the operating laws of the power system. These scenario trajectories fully record the dynamic evolution path of the power system from the current moment to various time stages within the next year, including key information such as changes in power output, fluctuations in load demand, and the evolution of grid status.

[0113] Understandably, Monte Carlo simulations can be implemented using a multi-agent scheduling decision model. This model simulates the evolution of the power system at different time scales. In the short term, it simulates the system's operation over the next seven days, with agents making decisions at the hourly level, focusing on real-time power balance, fluctuation suppression, and reserve configuration. In the medium term, it simulates the system's evolution over the next four weeks, with agents developing scheduling plans at the daily level, emphasizing optimization of reservoir scheduling, unit combination, and maintenance arrangements. In the long term, it simulates the system's development over the next twelve months, with agents planning at the monthly level, primarily focusing on strategic issues such as seasonal resource allocation, power generation planning, and inter-cycle coordination.

[0114] Step S402: Take the dynamic evolution path as the corresponding hydropower load scenario and calculate the probability of occurrence of the hydropower load scenario; In one feasible implementation, the scenario time span, system state transition probability, and action probability distribution of the collaborative scheduling strategy are obtained; the correspondence between the scenario time span, system state transition probability, action probability distribution, and the occurrence probability of the water and power load scenario is obtained; based on the scenario time span, the system state transition probability, the action probability distribution of the collaborative scheduling strategy, and the correspondence, the occurrence probability of the water and power load scenario is obtained.

[0115] It should be noted that each hydropower load scenario represents a possible dynamic evolution path. The probability of occurrence for each hydropower load scenario needs to be calculated, which is related to the scenario's time span, system state transition probabilities, and action probability distribution. The calculation formula is shown below:

[0116] In the formula, This indicates a scenario involving water and electricity supply. For the scene The probability of occurrence For the time span of the scene, The system state transition probability is determined by the physical characteristics of the power system. Indicates the cooperative scheduling strategy The probability distribution of the actions to be taken, i.e., the action probability distribution of the collaborative scheduling strategy. Indicates the joint actions at the current moment. Indicates the current state. It indicates the state at the next moment.

[0117] Step S403: Based on the water-power-load scenario and the probability of its occurrence, generate a set of water-power-load scenarios for the power system at different time stages under source-load uncertainty.

[0118] It should be understood that the generated hydropower load scenario set is divided into three levels according to the time dimension: the short-term scenario provides detailed data on hourly hydropower output, renewable energy generation, and load demand for the next seven days, and marks the power flow distribution and reserve capacity of key operating sections; the medium-term scenario shows the daily system operation status for the next four weeks, focusing on the reservoir water storage and release process, unit combination arrangement, and inter-regional power exchange plan, while also marking the maintenance arrangements and fuel supply of major equipment; the long-term scenario is the monthly system development outline for the next twelve months, including seasonal power generation plans, annual maintenance plans, and inter-basin hydropower compensation strategies, providing a reference for strategic decision-making. Each hydropower load scenario contains a complete system status description and an estimated probability of the scenario occurring, forming a complete probabilistic scenario set.

[0119] Understandably, the final output set of water, electricity, and load scenarios is presented in the form of a structured data file, which includes scenario number, time label, probability weight (probability of occurrence), and detailed state variable values.

[0120] This embodiment provides a scenario generation method based on source-load uncertainty. Based on a collaborative scheduling strategy, multiple Monte Carlo simulations are performed under source-load uncertainty conditions to obtain the dynamic evolution path of the power system at different time stages under these conditions. The dynamic evolution path includes changes in power output, load demand fluctuations, and grid state evolution. The dynamic evolution path is used as the corresponding hydropower-load scenario, and the probability of its occurrence is calculated. Based on the hydropower-load scenario and its probability of occurrence, a set of hydropower-load scenarios for the power system at different time stages under source-load uncertainty is generated. This embodiment tightly couples the prediction and decision optimization stages, using multi-timescale prediction results to dynamically guide reinforcement learning decisions. This enhances the system's adaptability to source-load uncertainty, avoids conflicts between decisions at different time scales, effectively mitigates operational risks caused by source-load fluctuations, and achieves organic coordination across multiple time scales and accurate generation of uncertain scenarios.

[0121] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the scene generation method based on source load uncertainty in this application. Any simple transformations based on this technical concept are within the protection scope of this application.

[0122] This application also provides a scene generation device based on source load uncertainty, please refer to... Figure 4 The scene generation device based on source load uncertainty includes: Data prediction module 10 is used to acquire real-time source-load data of the power system; The data prediction module 10 is also used to process the real-time source load data based on the target prediction model to obtain multi-timescale source load prediction results. The target prediction model is obtained by weight compression of the long short-term memory network prediction model. The collaborative decision-making module 20 is used to generate a collaborative scheduling strategy for source load resources under the source load prediction results of the multi-time scale based on the multi-agent scheduling decision model. The scenario generation module 30 is used to generate a set of water and power load scenarios of the power system at different time stages under the condition of source and load uncertainty, based on the cooperative scheduling strategy.

[0123] In one feasible implementation, the data prediction module 10 is further used to input the preprocessed real-time source load data into the target prediction model for multiple inferences to obtain multiple sets of prediction results at different time scales. The real-time source load data includes power output data on the power supply side, demand data on the load side, associated meteorological data, and associated hydrological data. Based on the numerical distribution of multiple sets of prediction results at each time scale, the confidence intervals of multiple sets of prediction results at each time scale are determined. Based on the confidence intervals of multiple prediction results at each time scale, the target prediction results and uncertainty prediction intervals at each time scale are determined. Based on the target prediction results and uncertainty prediction intervals at various time scales, multi-time scale source load prediction results are obtained.

[0124] In one feasible implementation, the data prediction module 10 is further configured to collect historical source load data and determine training data based on the historical source load data; Based on the training data, a Long Short-Term Memory (LSTM) network prediction model is trained. The key weight matrix is ​​extracted from the Long Short-Term Memory Network prediction model, and the key weight matrix is ​​compressed to obtain the reconstructed weight matrix. Based on the reconstructed weight matrix, the long short-term memory network prediction model is converted into a compressed prediction model; The compressed prediction model is fine-tuned to obtain the target prediction model.

[0125] In one feasible implementation, the data prediction module 10 is further configured to reshape the key weight matrix into a tensor form to obtain a key weight tensor. The key weight tensor is subjected to regularized Tucker decomposition to obtain the target core tensor and the target factor matrix; Based on the target core tensor and the target factor matrix, basic trend reconstruction and detailed fluctuation reconstruction are performed to obtain the reconstructed weight tensor; The reconstructed weight tensor is reshaped into a matrix form to obtain the reconstructed weight matrix.

[0126] In one feasible implementation, the data prediction module 10 is further configured to establish a joint optimization objective function based on the energy retention ratio, sensitivity-weighted reconstruction error, complexity penalty component, and balance hyperparameter. The joint optimization objective function is iteratively solved to determine the optimal compressed rank combination; Based on time smoothness constraints, core tensor sparsity constraints, and periodic constraints, we determine the hydropower scenario-specific regularization components. Based on the aforementioned key weight tensor and the hydropower scenario-specific regularization component, a regularization decomposition objective function is established. Based on the optimal compressed rank combination, the regularized decomposition objective function is solved iteratively to determine the target core tensor and the target factor matrix.

[0127] In one feasible implementation, the collaborative decision-making module 20 is further configured to fuse the multi-timescale source-load prediction results with the real-time operating status of the power system to obtain integrated status data. The integrated state data is input into the multi-agent scheduling decision model so that multiple agents in the multi-agent scheduling decision model can make collaborative decisions based on the integrated state data to obtain a collaborative scheduling strategy for source load resources under the source load prediction results of the multi-time scale.

[0128] In one feasible implementation, the scenario generation module 30 is further configured to perform multiple Monte Carlo simulations under source-load uncertainty conditions based on the cooperative scheduling strategy, to obtain the dynamic evolution path of the power system at different time stages under source-load uncertainty conditions, the dynamic evolution path including power output changes, load demand fluctuations and grid state evolution. The dynamic evolution path is taken as the corresponding hydropower load scenario, and the occurrence probability of the hydropower load scenario is calculated. Based on the water-power-load scenarios and their occurrence probabilities, a set of water-power-load scenarios for the power system at different time stages under source-load uncertainty is generated.

[0129] In one feasible implementation, the scene generation module 30 is further used to obtain the scene time span, system state transition probability, and action probability distribution of the cooperative scheduling strategy. Obtain the correspondence between the scene time span, system state transition probability, action probability distribution and the occurrence probability of water and power load scene; Based on the scenario time span, the system state transition probability, the action probability distribution of the collaborative scheduling strategy, and the corresponding relationship, the probability of occurrence of the water and power load scenario is obtained.

[0130] The scene generation device based on source load uncertainty provided in this application, employing the scene generation method based on source load uncertainty in the above embodiments, can solve the technical problem of the separation of the prediction and decision-making stages, making it difficult to dynamically adapt to the time-varying characteristics of source load uncertainty, resulting in a disconnect between scene generation and actual operating conditions. Compared with the prior art, the beneficial effects of the scene generation device based on source load uncertainty provided in this application are the same as those of the scene generation method based on source load uncertainty provided in the above embodiments, and other technical features in the scene generation device based on source load uncertainty are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0131] This application provides a scene generation device based on source-load uncertainty. The scene generation device based on source-load uncertainty includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the scene generation method based on source-load uncertainty in the above embodiment 1.

[0132] The following is for reference. Figure 5 This document illustrates a structural schematic diagram of a scene generation device based on source-load uncertainty suitable for implementing embodiments of this application. The scene generation device based on source-load uncertainty in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 5 The scene generation device based on source load uncertainty shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0133] like Figure 5 As shown, the scene generation device based on source-load uncertainty may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in ROM (Read Only Memory) 1002 or a program loaded from storage device 1003 into RAM (Random Access Memory) 1004. RAM 1004 also stores various programs and data required for the operation of the scene generation device based on source-load uncertainty. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via bus 1005. Input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the scene generation device based on source-load uncertainty to communicate wirelessly or wiredly with other devices to exchange data. Although a scene generation device based on source-load uncertainty with various systems is shown in the figure, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0134] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0135] The scene generation device based on source load uncertainty provided in this application, employing the scene generation method based on source load uncertainty in the above embodiments, can solve the technical problem of the separation of the prediction and decision-making stages, making it difficult to dynamically adapt to the time-varying characteristics of source load uncertainty, resulting in a disconnect between scene generation and actual operating conditions. Compared with the prior art, the beneficial effects of the scene generation device based on source load uncertainty provided in this application are the same as those of the scene generation method based on source load uncertainty provided in the above embodiments, and other technical features in this scene generation device based on source load uncertainty are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0136] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0137] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0138] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the scene generation method based on source load uncertainty in the above embodiments.

[0139] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0140] The aforementioned computer-readable storage medium may be included in a scene generation device based on source-load uncertainty; or it may exist independently and not assembled into a scene generation device based on source-load uncertainty.

[0141] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by a scenario generation device based on source-load uncertainty, the scenario generation device based on source-load uncertainty performs the following actions: acquires real-time source-load data of the power system; processes the real-time source-load data based on a target prediction model to obtain multi-time-scale source-load prediction results, wherein the target prediction model is obtained by weighting and compressing a long short-term memory network prediction model; generates a collaborative scheduling strategy for source-load resources under the multi-time-scale source-load prediction results based on a multi-agent scheduling decision model; and generates a set of hydropower and power load scenarios of the power system at different time stages under source-load uncertainty conditions based on the collaborative scheduling strategy.

[0142] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0143] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0144] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0145] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described scene generation method based on source load uncertainty. This solves the technical problem of the separation of the prediction and decision-making stages, making it difficult to dynamically adapt to the time-varying characteristics of source load uncertainty, leading to a disconnect between scene generation and actual operating conditions. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the scene generation method based on source load uncertainty provided in the above embodiments, and will not be repeated here.

[0146] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the scene generation method based on source load uncertainty as described above.

[0147] The computer program product provided in this application can solve the technical problem that the separation of the prediction and decision-making stages makes it difficult to dynamically adapt to the time-varying characteristics of source load uncertainty, resulting in a disconnect between scene generation and actual operating conditions. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the scene generation method based on source load uncertainty provided in the above embodiments, and will not be repeated here.

[0148] The above are only some embodiments of this application and do not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A scene generation method based on source load uncertainty, characterized in that, The method includes: Acquire real-time source-load data of the power system; The real-time source load data is processed based on the target prediction model to obtain multi-timescale source load prediction results. The target prediction model is obtained by weight compression of the long short-term memory network prediction model. Based on the multi-agent scheduling decision model, a collaborative scheduling strategy for source load resources is generated under the source load prediction results of the multi-time scale. Based on the aforementioned collaborative scheduling strategy, a set of water and power load scenarios for the power system at different time stages under source-load uncertainty conditions is generated.

2. The method as described in claim 1, characterized in that, The steps of processing the real-time source load data based on the target prediction model to obtain multi-timescale source load prediction results include: The preprocessed real-time source-load data is input into the target prediction model for multiple inferences to obtain multiple sets of prediction results at different time scales. The real-time source-load data includes power output data on the power supply side, demand data on the load side, associated meteorological data, and associated hydrological data. Based on the numerical distribution of multiple sets of prediction results at each time scale, the confidence intervals of multiple sets of prediction results at each time scale are determined. Based on the confidence intervals of multiple prediction results at each time scale, the target prediction results and uncertainty prediction intervals at each time scale are determined. Based on the target prediction results and uncertainty prediction intervals at various time scales, multi-time scale source load prediction results are obtained.

3. The method as described in claim 1, characterized in that, The step of acquiring real-time source-load data of the power system includes, prior to: Collect historical source load data, and determine training data based on the historical source load data; Based on the training data, a Long Short-Term Memory (LSTM) network prediction model is trained. The key weight matrix is ​​extracted from the Long Short-Term Memory Network prediction model, and the key weight matrix is ​​compressed to obtain the reconstructed weight matrix. Based on the reconstructed weight matrix, the long short-term memory network prediction model is converted into a compressed prediction model; The compressed prediction model is fine-tuned to obtain the target prediction model.

4. The method as described in claim 3, characterized in that, The step of compressing the key weight matrix to obtain the reconstructed weight matrix includes: The key weight matrix is ​​reshaped into a tensor form to obtain the key weight tensor; The key weight tensor is subjected to regularized Tucker decomposition to obtain the target core tensor and the target factor matrix; Based on the target core tensor and the target factor matrix, basic trend reconstruction and detailed fluctuation reconstruction are performed to obtain the reconstructed weight tensor; The reconstructed weight tensor is reshaped into a matrix form to obtain the reconstructed weight matrix.

5. The method as described in claim 4, characterized in that, The step of performing regularized Tucker decomposition on the key weight tensor to obtain the target core tensor and the target factor matrix includes: A joint optimization objective function is established based on the energy retention ratio, sensitivity-weighted reconstruction error, complexity penalty component, and balance hyperparameter; The joint optimization objective function is iteratively solved to determine the optimal compressed rank combination; Based on time smoothness constraints, core tensor sparsity constraints, and periodic constraints, we determine the hydropower scenario-specific regularization components. Based on the aforementioned key weight tensor and the hydropower scenario-specific regularization component, a regularization decomposition objective function is established. Based on the optimal compressed rank combination, the regularized decomposition objective function is solved iteratively to determine the target core tensor and the target factor matrix.

6. The method as described in claim 1, characterized in that, The step of generating a collaborative scheduling strategy for source load resources under the multi-timescale source load prediction results based on a multi-agent scheduling decision model includes: The multi-timescale source-load prediction results are fused with the real-time operating status of the power system to obtain integrated status data; The integrated state data is input into the multi-agent scheduling decision model so that multiple agents in the multi-agent scheduling decision model can make collaborative decisions based on the integrated state data to obtain a collaborative scheduling strategy for source load resources under the source load prediction results of the multi-time scale.

7. The method as described in claim 1, characterized in that, The step of generating a set of hydropower load scenarios for the power system at different time stages under source-load uncertainty based on the cooperative scheduling strategy includes: Based on the aforementioned coordinated scheduling strategy, multiple Monte Carlo simulations are performed under source-load uncertainty to obtain the dynamic evolution path of the power system at different time stages under source-load uncertainty. The dynamic evolution path includes changes in power output, fluctuations in load demand, and evolution of grid status. The dynamic evolution path is taken as the corresponding hydropower load scenario, and the occurrence probability of the hydropower load scenario is calculated. Based on the water-power-load scenarios and their occurrence probabilities, a set of water-power-load scenarios for the power system at different time stages under source-load uncertainty is generated.

8. The method as described in claim 7, characterized in that, The steps for calculating the probability of occurrence of the water-powered load scenario include: Obtain the scene time span, system state transition probability, and action probability distribution of the cooperative scheduling strategy; Obtain the correspondence between the scene time span, system state transition probability, action probability distribution and the occurrence probability of water and power load scene; Based on the scenario time span, the system state transition probability, the action probability distribution of the collaborative scheduling strategy, and the corresponding relationship, the probability of occurrence of the water and power load scenario is obtained.

9. A scene generation device based on source load uncertainty, characterized in that, The device includes: The data prediction module is used to acquire real-time source-load data of the power system; The data prediction module is also used to process the real-time source load data based on the target prediction model to obtain multi-timescale source load prediction results. The target prediction model is obtained by weight compression of the long short-term memory network prediction model. The collaborative decision-making module is used to generate a collaborative scheduling strategy for source load resources under the source load prediction results at the multiple time scales, based on a multi-agent scheduling decision-making model. The scenario generation module is used to generate a set of water and power load scenarios of the power system at different time stages under the condition of source and load uncertainty, based on the cooperative scheduling strategy.

10. A scene generation device based on source load uncertainty, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the scene generation method based on source load uncertainty as described in any one of claims 1 to 8.

Images