A Multi-Time-Scale Intelligent Dispatch Method for Provincial Power Grids with New Energy Integration

Abstract

This invention relates to the field of power system dispatching technology and provides a multi-timescale intelligent dispatching method for provincial power grids with renewable energy integration. The method includes: acquiring relevant data from the provincial power grid; using a trained multi-timescale prediction model to generate the current time-ahead timescale renewable energy power point prediction curve, the intraday timescale renewable energy power fluctuation range prediction result, and the real-time timescale renewable energy power point prediction curve and renewable energy power fluctuation range prediction result; processing the data with a day-ahead deep reinforcement learning agent to output day-ahead dispatching instructions; processing the data with an intraday deep reinforcement learning agent to output intraday adjustment instructions; and processing the data with a real-time deep reinforcement learning agent to output real-time control instructions. This invention achieves global optimization from day-ahead robust planning and intraday proactive prevention to real-time multi-objective collaboration, improving the collaborative dispatching control level of dispatching decisions.

Description

Technical Field

[0001] This invention relates to the field of power system dispatching technology, and in particular to a multi-timescale intelligent dispatching method for provincial power grids with new energy access. Background Technology

[0002] The strong randomness, volatility, and low inertia exhibited by renewable energy grid integration pose significant challenges to traditional power system dispatching and operation modes, significantly increasing the technical difficulty of ensuring the safe, stable, and economical operation of the power grid. To address this challenge, multi-timescale coordinated dispatching has become a core technical framework for power system operation. This framework typically divides the dispatching process into different timescales, such as day-ahead, intraday, and real-time, and uses a progressively refined and rolling optimization approach to smooth out renewable energy fluctuations and achieve a dynamic balance between generation and load.

[0003] In existing technologies, the models at each scale in multi-timescale joint scheduling are relatively independent, or only connected by simple data transmission. The lack of effective collaborative feedback and joint optimization between the three levels of day-ahead scheduling, intraday rolling scheduling, and real-time control can lead to disconnections in the transmission of scheduling commands, preventing the formation of an organic whole to address global security risks. Furthermore, existing control strategies primarily focus on the single security objective of eliminating frequency deviations, neglecting multi-objective collaborative optimization such as smoothing tie-line power fluctuations and extending the lifespan of critical regulating resources. This can lead to accelerated wear and tear on energy storage and other equipment due to frequent charging and discharging, or cause drastic fluctuations in tie-line power, compromising the system's economy and security.

[0004] In view of this, a multi-time-scale intelligent scheduling method for provincial power grids with new energy access is proposed. Summary of the Invention

[0005] This invention provides a multi-timescale intelligent scheduling method for provincial power grids with new energy access, which is used to solve the problem that it is impossible to form an organic whole to deal with the security risks of the whole region, and that the system's economy and security may be compromised.

[0006] This invention provides a multi-timescale intelligent scheduling method for provincial power grids with renewable energy access. The method employs a multi-agent deep reinforcement learning model, composed of day-ahead, intraday, and real-time deep reinforcement learning agents, for scheduling. The method includes:

[0007] Obtain real-time meteorological data, real-time satellite cloud image data, and real-time power grid operation data of relevant new energy power stations in the provincial power grid at the current moment;

[0008] Based on the acquired real-time data, the trained multi-timescale prediction model is used to generate the prediction results for the current moment. The prediction results include the new energy power point prediction curve at the day-ahead timescale, the new energy power fluctuation range prediction results at the intraday timescale, and the new energy power point prediction curve and new energy power fluctuation range prediction results at the real-time timescale.

[0009] The predicted curve of new energy power points at the day-ahead time scale, along with load forecast data and grid topology data obtained from the grid dispatch system, are input into the trained day-ahead deep reinforcement learning agent, and day-ahead dispatch instructions are output.

[0010] Based on the day-ahead scheduling instructions, the predicted results of the new energy power fluctuation range at the intraday time scale are input into the trained intraday deep reinforcement learning agent in a rolling optimization manner, and intraday adjustment instructions are output.

[0011] Based on the day-ahead scheduling instructions and intraday adjustment instructions, combined with the real-time time-scale new energy power point prediction curve, the new energy power fluctuation range prediction results, and the real-time frequency deviation of the power grid obtained from the power grid measurement system, a trained real-time deep reinforcement learning agent outputs real-time control instructions.

[0012] Furthermore, the multi-timescale prediction model is constructed using a Long Short-Term Memory (LSTM) network model, and the construction and training process includes:

[0013] Acquire historical meteorological data, historical satellite cloud image data, and historical power grid operation data;

[0014] Based on the acquired historical data, a day-ahead timescale prediction model, an intraday timescale prediction model, and a real-time timescale prediction model were trained respectively; among which:

[0015] The day-ahead timescale prediction model is used to output a new energy power point prediction curve with a first preset time period as the period.

[0016] The intraday timescale prediction model is used to output the prediction results of the new energy power fluctuation range with the second preset time period as the cycle.

[0017] The real-time time-scale prediction model is used to output the prediction curve of new energy power points and the prediction results of new energy power fluctuation range with a third preset time period as the cycle.

[0018] Furthermore, the Long Short-Term Memory (LSTM) network model is also used to construct a dynamic simulation environment for the power grid. The construction process includes:

[0019] An LSTM state transition model is trained based on the historical power grid operation data. The LSTM state transition model takes the power grid operation state and dispatch instructions of the previous moment as input and outputs the predicted power grid operation state of the next moment.

[0020] The trained LSTM state transition model is used to construct a power grid dynamic simulation environment. The power grid dynamic simulation environment receives scheduling instructions, adjustment instructions or control instructions output by the deep reinforcement learning agent, simulates the execution of instructions based on the current power grid operating state, and returns new power grid operating state data.

[0021] Furthermore, the day-ahead deep reinforcement learning agent, the intraday deep reinforcement learning agent, and the real-time deep reinforcement learning agent are trained in the power grid dynamic simulation environment. The training process includes:

[0022] In the power grid dynamic simulation environment, the historical meteorological data, historical satellite cloud image data, and historical power grid operation data are used to train the day-ahead deep reinforcement learning agent.

[0023] Based on the pre-day deep reinforcement learning agent that has completed fixed training, the intra-day deep reinforcement learning agent is trained using the historical meteorological data, historical satellite cloud image data, and historical power grid operation data.

[0024] Based on the pre-day and intra-day deep reinforcement learning agents that have completed fixed training, the real-time deep reinforcement learning agents are trained using the historical meteorological data, historical satellite cloud image data, and historical power grid operation data.

[0025] Furthermore, the action space of the aforementioned deep reinforcement learning agent includes key action decisions, which include conventional unit start-up and shutdown instructions, new energy power plant output plan instructions, and inter-provincial transaction instructions.

[0026] Furthermore, the action space of the intraday deep reinforcement learning agent includes adjustment action instructions, which include unit output adjustment instructions, energy storage plan optimization instructions, and demand response activation instructions.

[0027] Furthermore, the action space of the real-time deep reinforcement learning agent includes rapid control instructions, which include AGC adjustment instructions, rapid energy storage control instructions, and new energy emergency control instructions.

[0028] Furthermore, the reward function of the aforementioned deep reinforcement learning agent includes an operating cost reward, a renewable energy curtailment penalty, and a cybersecurity risk penalty.

[0029] The cybersecurity risk penalty items The calculation is based on the predicted range of new energy power fluctuations at the intraday time scale, and the calculation formula is as follows:

[0030]

[0031] in: This represents the total number of critical transmission lines in the power grid. For the lines calculated based on the day-ahead scheduling plan The transmission power; For the line The power transmission limit; Based on the intraday timescale prediction results of renewable energy power fluctuation range, the power transfer distribution factor is used to calculate the line potential impact of renewable energy power fluctuations. The amplitude of transmission power fluctuation; This is the risk confidence level coefficient, used to adjust the level of confidence in the predicted volatility amplitude; For the line The over-limit penalty coefficient is used to characterize the line. The importance of safety in the power grid.

[0032] Furthermore, the reward function of the intraday deep reinforcement learning agent includes a plan tracking reward and a blockage prevention reward.

[0033] The blockage prevention reward item The calculation is based on the predicted range of new energy power fluctuations at the aforementioned real-time time scale, and the calculation formula is as follows:

[0034]

[0035]

[0036] in: This represents the total number of critical transmission lines in the power grid. After executing the intraday adjustment order, the line Predicted transmission power; For the line Real-time power transmission safety margin; For the line The ground-state transmission power at the current moment; Based on the intraday timescale prediction results of renewable energy power fluctuation range, the power transfer distribution factor is used to calculate the line potential impact of renewable energy power fluctuations. The amplitude of transmission power fluctuation; This is the real-time safety margin coefficient, used to adjust the degree of impact of ultra-short-term forecast fluctuations on the safety margin calculation; For the line The congestion prevention reward weight is used to incentivize agents to prioritize ensuring the safety margin of critical paths.

[0037] Furthermore, the reward function of the real-time deep reinforcement learning agent It includes a frequency stability bonus, a fluctuation smoothing bonus, and a device lifespan bonus, calculated using the following formula:

[0038]

[0039] in: The system's real-time frequency deviation is obtained from the power grid measurement system; This represents the power change of the power grid interconnection line within adjacent control cycles. The rated total power of the system is used for... Standardize; It represents the absolute value of the change in the state of charge of the energy storage within a single control cycle; This is the frequency deviation penalty coefficient, used to balance the weight of frequency stability in the total reward; This is a power fluctuation penalty factor used to suppress frequent fluctuations in tie-line power. This is the energy storage lifespan loss penalty coefficient, used to balance control effectiveness and the lifespan of energy storage equipment.

[0040] As can be seen from the above technical solutions, the present invention has the following advantages:

[0041] This invention employs a multi-agent model architecture comprised of day-ahead, intraday, and real-time deep reinforcement learning agents. First, based on real-time meteorological, satellite cloud imagery, and power grid operation data, a trained multi-timescale prediction model generates renewable energy power prediction results covering both point prediction and fluctuation range prediction. The day-ahead agent outputs day-ahead scheduling commands based on the point prediction curve, the intraday agent outputs adjustment commands using a rolling optimization approach based on the day-ahead commands and fluctuation range predictions, and the real-time agent outputs control commands by integrating the day-ahead commands, adjustment commands, and real-time frequency deviations. By deeply embedding prediction results from different timescales into corresponding agent levels, this invention enables current scheduling to assess potential risks, intraday rolling to proactively prevent transmission congestion, and real-time control to collaboratively optimize frequency stability and energy storage lifetime. This effectively overcomes the technical shortcomings of existing technologies, such as disconnected decision-making across timescales, delayed response to renewable energy uncertainties, and a single control objective, thus significantly improving the economy, safety, and stability of power grid operation under renewable energy integration. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the system architecture of a provincial power grid multi-timescale intelligent scheduling method for new energy access in this invention;

[0043] Figure 2 This is a schematic diagram of the system architecture of the power grid dynamic simulation environment in this invention;

[0044] Figure 3 This is a flowchart illustrating the scheduling application stage of a multi-timescale intelligent scheduling method for provincial power grids with new energy access, as described in this invention. Detailed Implementation

[0045] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “corresponding to,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0046] Example 1

[0047] Please see Figure 1 The execution of this embodiment is based on Figure 1 The system architecture shown demonstrates how the intelligent scheduling and control method is implemented through the collaborative work of each layer and module within this architecture, as explained below:

[0048] 1. Data Input Layer: This layer is the foundation of the system, responsible for collecting all necessary input information, including real-time meteorological data, real-time satellite cloud imagery data, and real-time power grid operation data. This data is cleaned and standardized by the preprocessing module to provide high-quality input for the prediction model.

[0049] 2. Data Processing Layer: This layer adopts a coupled architecture of Long Short-Term Memory (LSTM) network and multi-agent deep reinforcement learning. The LSTM prediction module of the LSTM network constructs and runs prediction models at three time scales: day-ahead, intraday, and real-time, based on the real-time data from the input layer. Specifically, the day-ahead model outputs the predicted curve of renewable energy power points, the intraday model outputs the predicted results of renewable energy power fluctuation ranges, and the real-time model outputs both the predicted curve of the points and the predicted results of the fluctuation ranges.

[0050] The multi-agent deep reinforcement learning model employs three agents with clearly defined roles and working collaboratively: day-ahead, intraday, and real-time agents. All agents are trained within a power grid dynamic simulation environment built on an LSTM state transition model. This simulation environment accurately simulates the power grid's state transitions under dispatch commands, providing a safe and efficient training platform for the agents.

[0051] The newly developed deep reinforcement learning agent's state space is composed of current-day renewable energy power point prediction curves, load forecast data, and grid topology data, focusing on global and steady-state information for the next 24 hours. Its action space comprises two key dimensions: in terms of controlled objects, it covers conventional generating units, renewable energy plants, and inter-provincial transactions; in terms of decision-making characteristics, it consists of key actions that are strategic and irreversible, laying the foundation for subsequent scheduling.

[0052] Intraday deep reinforcement learning agent: Its state space is characterized by the introduction of day-ahead scheduling instructions and intraday renewable energy power fluctuation range prediction results, achieving a combination of planning and uncertainty. Its action space also includes two dimensions: the controlled object dimension covers conventional units, energy storage systems, and flexible loads; the decision characteristics dimension consists of adjustment actions, which are tactical and adjustable, used for rolling optimization and risk prevention within the day-ahead framework.

[0053] Real-time deep reinforcement learning agent: Its state space integrates day-ahead instructions, intraday adjustment instructions, real-time point prediction and fluctuation range prediction, and real-time frequency deviation of the power grid, focusing on the instantaneous dynamic balance of the system. Its action space targets conventional units, fast energy storage, and new energy power plants in the dimension of controlled objects; in the dimension of decision characteristics, it consists of fast actions with executableness and millisecond-level response capabilities to ensure real-time safety and stability.

[0054] 3. Equipment Output Layer: This layer represents the actual controlled objects of the power grid, receiving and executing instructions from the dispatch execution layer; it mainly includes:

[0055] Conventional generator sets: receive start-stop plans (critical actions) from day-ahead agents, output adjustment commands (adjustment actions) from intraday agents, and automatic generation control (AGC, Automatic Generation Control) (rapid actions) from real-time agents to achieve full-cycle collaborative control from hourly to second-level.

[0056] New energy power plants: mainly receive planned output instructions (key actions) from the day-ahead agent, and execute emergency control instructions (rapid actions) from the real-time agent when necessary.

[0057] Energy storage system: It mainly receives charging and discharging plan optimization instructions (adjustment actions) from intraday intelligent agents and rapid power adjustment instructions (rapid actions) from real-time intelligent agents. It is a key flexible resource for smoothing power fluctuations and participating in frequency regulation.

[0058] Example 2

[0059] The scheduling method of this invention comprises two main stages: a model training stage and a scheduling application stage. The model training stage is completed offline, constructing and training the core intelligent decision-making model; the scheduling application stage, on the other hand, runs online, utilizing the trained model for real-time scheduling.

[0060] Part 1: Model Training Phase

[0061] This stage aims to build and train the core intelligent agent of this invention, and its process is as follows:

[0062] The LSTM model was chosen here because its gating mechanism (input gate, forget gate, output gate) can effectively capture long-term dependencies in time-series data, which is crucial for predicting renewable energy power output that is continuously affected by weather systems. The principle is that the internal state of the LSTM can transmit key information across time steps, thereby better modeling the complex nonlinear time-series relationship between meteorological conditions and renewable energy output. The purpose of building this prediction model is to provide a complete predictive information spectrum, from trend to detail and from certainty to uncertainty, for subsequent scheduling decisions at different time scales. The construction and training process includes:

[0063] 1. Acquire historical meteorological data, historical satellite cloud imagery data, and historical power grid operation data;

[0064] 2. Based on the acquired historical data, train the day-ahead timescale prediction model, the intraday timescale prediction model, and the real-time timescale prediction model respectively; wherein: the day-ahead timescale prediction model is used to output the new energy power point prediction curve with the first preset time period as the period; the intraday timescale prediction model is used to output the new energy power fluctuation range prediction result with the second preset time period as the period; the real-time timescale prediction model is used to output the new energy power point prediction curve and the new energy power fluctuation range prediction result with the third preset time period as the period.

[0065] It should be noted that although these three models are all based on the LSTM architecture, they have significant differences in the temporal resolution of the input data, output target, and training data. Therefore, they need to be trained independently, rather than simultaneously or hierarchically.

[0066] The current-day timescale forecasting model primarily takes into account historical meteorological data and corresponding historical power data from the past few days, learning the macro-cyclical patterns of weather and power output. It outputs a renewable energy power point forecast curve with a first preset time period (the next 24 hours) as the period and 15-minute intervals. This curve provides a baseline power plan for the next day, emphasizing predictive certainty and overall trend accuracy.

[0067] Intraday timescale forecasting model: The main inputs include the latest short-term weather forecasts and higher-frequency real-time satellite cloud imagery data to capture rapid changes in weather, particularly cloud cover, over the next few hours. Output is a forecast of renewable energy power fluctuations within a 15-minute interval, calculated over a second preset time period (the next 4 hours). This interval forecast quantifies the uncertainty of the forecast, providing a risk assessment basis for intraday rolling optimization.

[0068] Real-time timescale prediction model: Primarily inputs are ultra-short-term meteorological data from the last few hours and real-time power grid operation data, focusing on capturing minute-level fluctuations. Outputs include renewable energy power point prediction curves and renewable energy power fluctuation range predictions with a period of the third preset time period (the next 15 minutes) and intervals of 5 minutes or even shorter. Point predictions are used for real-time power balance, while interval predictions provide risk boundaries for real-time control.

[0069] Please see Figure 2 The Long Short-Term Memory (LSTM) network model is also used to construct dynamic power grid simulation environments. Compared to traditional numerical simulations based on precise physical equations or simple linear models, the LSTM state transition model can learn complex, nonlinear power grid dynamic responses from massive amounts of historical data. While ensuring sufficient accuracy, it effectively improves simulation speed, making large-scale, high-efficiency DRL training possible. The construction process includes:

[0070] 1. Train an LSTM state transition model based on historical power grid operation data; The LSTM state transition model takes the power grid operation state and dispatch instructions of the previous moment as input and outputs the predicted power grid operation state of the next moment;

[0071] 2. The trained LSTM state transition model is used to construct a power grid dynamic simulation environment. The power grid dynamic simulation environment receives scheduling instructions, adjustment instructions or control instructions output by the deep reinforcement learning agent, simulates the execution of instructions based on the current power grid operating state, and returns new power grid operating state data.

[0072] Specifically, historical data pairs containing grid state and corresponding dispatch commands executed at each moment are collected. The previous grid state and the current dispatch command are used as input features, and the next grid state is used as the prediction target to supervise the training of an LSTM model, enabling it to learn to predict the system's state transitions under a given command. The trained LSTM state transition model is then encapsulated into a grid dynamic simulation environment. This environment receives dispatch commands, adjustment commands, or control commands output by a deep reinforcement learning agent. Based on the current grid operating state, it simulates the execution of the command through an internal LSTM state transition model and quickly deduces new grid operating state data, returning it to the agent.

[0073] The principle is described using a specific scenario: Assume that during the training of a real-time agent, the current state of the simulation environment is a frequency of 50.01Hz and a power flow of 90MW for a certain line. The agent attempts to output a command to increase the output of unit A by 10MW. The simulation environment inputs this command along with the current state into the LSTM state transition model. Based on patterns learned from historical data, the model quickly predicts that after executing the command, the frequency may rise to 50.03Hz, and the power flow of the line will become 95MW (still within the limit), returning these new states and corresponding positive rewards (more stable frequency). If the command causes the predicted power flow value to exceed the limit, a negative reward is returned. Through several such rapid simulation interactions, the agent can efficiently learn the optimal control strategy without having to go through trial and error in slow real systems or detailed physical simulations.

[0074] In this embodiment, within the aforementioned constructed power grid dynamic simulation environment, historical data is used to train the day-ahead deep reinforcement learning agent, the intraday deep reinforcement learning agent, and the real-time deep reinforcement learning agent in stages. The training process includes:

[0075] 1. In a power grid dynamic simulation environment, historical meteorological data, historical satellite cloud image data, and historical power grid operation data are used to train a day-ahead deep reinforcement learning agent;

[0076] In a dynamic power grid simulation environment, a day-ahead deep reinforcement learning agent is trained using historical meteorological data, historical satellite cloud imagery data, and historical power grid operation data. During this phase, the agent learns to make decisions on key actions such as generator start-up and shutdown, renewable energy planned output, and inter-provincial power transactions for the next 24 hours based on predicted curves. Its reward function not only incentivizes reductions in total operating costs but also includes a cybersecurity risk penalty term. This penalty term is calculated using the predicted renewable energy power fluctuation range over an intraday timescale. For example, when the day-ahead plan causes the power flow of a critical line to approach its limit, the system assesses the risk of the plan exceeding the power flow limit when faced with actual power fluctuations based on the intraday predicted power fluctuation range and applies a corresponding penalty. This drives the day-ahead agent to no longer focus solely on a single predicted scenario when formulating plans, but rather to proactively reserve necessary safety margins for potential power fluctuations in subsequent intraday periods, thereby outputting a more robust day-ahead scheduling plan.

[0077] 2. Based on the daily deep reinforcement learning agent that has been trained in a fixed manner, the intraday deep reinforcement learning agent is trained using historical meteorological data, historical satellite cloud image data, and historical power grid operation data;

[0078] A pre-trained day-ahead deep reinforcement learning agent is used, and then trained intraday deep reinforcement learning agents based on historical data. At this stage, the simulation starts with a day-ahead plan output by the trained agent, conforming to global economic efficiency and security. The intraday agent's task is to learn, within this given framework, to continuously output adjustment commands such as unit output adjustment and energy storage plan optimization based on the predicted range of renewable energy power fluctuations on the intraday timescale. One of the core components of its reward function is the plan tracking reward, which incentivizes smooth following of the day-ahead plan and avoids large deviations. Its congestion prevention reward drives the intraday agent to proactively and preventatively adjust resource allocation when it identifies certain lines with congestion risks in the near future, eliminating congestion before it actually occurs. By learning within this framework of a fixed upper layer and a trained lower layer, the intraday agent can gain a deep understanding of how to make optimal fine-tuning within the boundaries of the day-ahead plan, achieving a seamless connection between the tactical and strategic layers.

[0079] 3. Based on the pre-day and intra-day deep reinforcement learning agents that have been trained, the real-time deep reinforcement learning agent is trained using historical meteorological data, historical satellite cloud image data, and historical power grid operation data.

[0080] A fixed day-ahead and intraday deep reinforcement learning agent are established, and a real-time deep reinforcement learning agent is trained using historical data. The simulated operating state is simultaneously influenced by both the planned instructions from the fixed day-ahead agent and the adjusted instructions from the fixed intraday agent. The real-time agent needs to learn information such as real-time renewable energy power point prediction and fluctuation range prediction, as well as real-time grid frequency deviation, to output rapid control instructions such as AGC adjustment and rapid energy storage control. The reward function directly addresses the instantaneous safety and stability of the grid, including frequency deviation penalties and power fluctuation smoothness rewards. Since the upper-level agents are fixed, the real-time agent can deeply understand how its rapid actions function within the environment created by the upper-level plan during training. For example, it learns how to most effectively utilize rapid resources such as energy storage to smooth out random fluctuations without violating the existing scheduling framework, while also considering equipment lifespan. This ensures that real-time control actions are not only fast but also highly coordinated with the global scheduling objectives.

[0081] This phased, fixed training ensures that the decision-making strategies of the three agents are top-down and coordinated with each other.

[0082] Part Two: Scheduling Application Phase

[0083] Please see Figure 3 This stage utilizes all the models trained in the first part to perform daily intelligent scheduling online. The specific process is as follows:

[0084] S1. Obtain real-time meteorological data, real-time satellite cloud image data, and real-time power grid operation data of relevant new energy power stations in the provincial power grid at the current moment;

[0085] This step utilizes a data acquisition and communication system deployed at the provincial power grid dispatch center to acquire the underlying data required for dispatch decisions in real time from multiple heterogeneous data sources. Specifically, real-time meteorological data acquisition involves establishing a data interface between the power dispatch data network and the meteorological department's database to obtain high-precision numerical weather forecast data covering all renewable energy power plants in the province. Key parameters collected include, but are not limited to, wind speed, wind direction, irradiance, ambient temperature, and relative humidity for the next 0-72 hours; these data are the core inputs driving the renewable energy power prediction model. Real-time satellite cloud image data acquisition involves obtaining frequently updated real-time satellite cloud image data from the National Satellite Meteorological Center or regional meteorological satellite ground stations via dedicated lines. This data is primarily used to track the movement trajectory, thickness, and coverage of cloud clusters, providing crucial visual evidence for predicting power fluctuations in ultra-short-term photovoltaic power plants. Real-time grid operation data acquisition involves collecting real-time operating status information of the entire grid at second or minute intervals through the energy management system and data acquisition and monitoring control system of the grid dispatch master station. On the power supply side, this includes the active / reactive power output and operating status of each conventional generator unit, the real-time active power output and inverter operating data of each new energy power station, the power flow of key transmission lines, transformer load rate, node voltage amplitude and phase angle, the actual power value of loads in each region, and the system level, including the real-time grid frequency and tie-line power.

[0086] All collected real-time data will be sent to the system's data preprocessing module for processing, including data cleaning, outlier detection and correction, format standardization, and time series alignment, to form a consistent and reliable dataset, providing high-quality input for subsequent prediction and decision-making modules.

[0087] S2. Based on the acquired real-time data, the trained multi-timescale prediction model is used to generate the prediction results for the current moment. The prediction results include the new energy power point prediction curve for the day-ahead timescale, the new energy power fluctuation range prediction results for the intraday timescale, and the new energy power point prediction curve and new energy power fluctuation range prediction results for the real-time timescale.

[0088] This step differs from the training phase, which uses historical data. In this application phase, the system inputs real-time meteorological data, real-time satellite cloud imagery data, and real-time power grid operation data acquired in S1 into a pre-trained multi-timescale prediction model, generating online prediction results for different decision-making levels at the current moment. This process fully leverages the LSTM model's rapid forward inference capability for time-series data, achieving multi-dimensional and multi-precision perception of future renewable energy output, and providing dynamically updated environmental state information for subsequent real-time decision-making by the intelligent agent.

[0089] S3. Input the day-ahead timescale new energy power point prediction curve and the load prediction data and grid topology data obtained from the grid dispatch system into the trained day-ahead deep reinforcement learning agent, and output the day-ahead dispatch instructions.

[0090] 1. The state space of a deep reinforcement learning agent currently includes the current-timescale renewable energy power point prediction curve, the 24-hour load forecast data obtained from the power grid dispatch system, the power grid topology, and the DC power receiving plan. These state variables together describe the expected scenario and physical constraints of the power grid operation in the coming day.

[0091] 2. The action space of a deep reinforcement learning agent currently includes two dimensions: the controlled object and the decision characteristics. The controlled object includes conventional generating units, renewable energy power plants, and inter-provincial transactions. The decision characteristics dimension consists of key actions, which have far-reaching and irreversible effects. Specifically, these include start-up and shutdown instructions for conventional generating units, used to determine which units will operate in the next 24 hours; output plan instructions for renewable energy power plants, used to formulate benchmark power generation plans for renewable energy power plants; and inter-provincial transaction instructions, used to determine the amount of electricity traded with the external grid.

[0092] 3. Currently, the reward function for deep reinforcement learning agents includes operating cost rewards, renewable energy curtailment penalties, and cybersecurity risk penalties; the current reward function for agents... In the application phase, it is used to guide the trained policy in decision-making, and its expression is:

[0093]

[0094] in: This is an operating cost incentive item, negatively correlated with the total system operating cost, and is calculated using the following formula: ,in For the unit's fuel cost, For unit start-up and shutdown costs, For system backup costs; This is a penalty item for curtailment of renewable energy, negatively correlated with the amount of wind and solar power curtailed, and the calculation formula is as follows: ,in The total amount of electricity wasted, This is the penalty coefficient for power curtailment; This is a penalty item for cybersecurity risks.

[0095] Specifically, cybersecurity risk penalties The calculation is based on the predicted range of renewable energy power fluctuations over an intraday time scale. The calculation formula is as follows:

[0096]

[0097] in: This represents the total number of critical transmission lines in the power grid. For the lines calculated based on the day-ahead scheduling plan The transmission power; For the line The power transmission limit; Based on the intraday timescale prediction results of renewable energy power fluctuation range, the power transfer distribution factor is used to calculate the line potential impact of renewable energy power fluctuations. The amplitude of transmission power fluctuation; This is the risk confidence level coefficient, used to adjust the level of confidence in the predicted volatility amplitude; For the line The over-limit penalty coefficient is used to characterize the line. The importance of safety in the power grid.

[0098] S4. Based on the day-ahead scheduling instructions, the prediction results of the new energy power fluctuation range at the intraday time scale are input into the trained intraday deep reinforcement learning agent in a rolling optimization manner, and the intraday adjustment instructions are output.

[0099] 1. The state space of the intraday deep reinforcement learning agent includes the day-ahead scheduling instructions (as the adjustment benchmark), the intraday timescale prediction results of the new energy power fluctuation range (representing uncertainty), the current unit status, the energy storage charge status, and the latest short-term load forecast.

[0100] 2. The action space of an intraday deep reinforcement learning agent includes two dimensions: the controlled object and the decision characteristics. The controlled object dimension covers conventional generating units, energy storage systems, and flexible loads. The decision characteristics dimension consists of adjustment actions, which are continuous and reversible fine-tuning within the day-ahead framework. Specifically, these include generating unit output adjustment commands, which fine-tune the output within the generating unit ramp rate range; energy storage plan optimization commands, which adjust the charging and discharging power plan of energy storage; and demand response activation commands, which call upon flexible load resources.

[0101] 3. The reward function of an intraday deep reinforcement learning agent includes a plan tracking reward and a blocking prevention reward; the reward function of an intraday agent. The expression is:

[0102]

[0103] in: To track reward items, the incentive tracking date plan is calculated using the following formula: ,in Contribute to actual / predicted efforts, This is the planned value for the previous day. To track weights, This is a reward item for preventing blockages.

[0104] Blockage Prevention Rewards The calculation is based on the prediction results of the fluctuation range of new energy power at a real-time time scale. The calculation formula is as follows:

[0105]

[0106]

[0107] in: This represents the total number of critical transmission lines in the power grid. After executing the intraday adjustment order, the line Predicted transmission power; For the line Real-time power transmission safety margin; For the line The ground-state transmission power at the current moment; Based on the intraday timescale prediction results of renewable energy power fluctuation range, the power transfer distribution factor is used to calculate the line potential impact of renewable energy power fluctuations. The amplitude of transmission power fluctuation; This is the real-time safety margin coefficient, used to adjust the degree of impact of ultra-short-term forecast fluctuations on the safety margin calculation; For the line The congestion prevention reward weight is used to incentivize agents to prioritize ensuring the safety margin of critical paths.

[0108] S5. Based on the day-ahead scheduling instructions and intraday adjustment instructions, combined with the real-time time-scale prediction curve of renewable energy power points, the prediction results of renewable energy power fluctuation range, and the real-time frequency deviation of the power grid obtained from the power grid measurement system, the trained real-time deep reinforcement learning agent outputs real-time control instructions.

[0109] 1. The state space of the real-time deep reinforcement learning agent includes day-ahead scheduling instructions and intraday adjustment instructions, real-time time-scale new energy power point prediction curves and fluctuation range prediction results, real-time grid frequency deviation, actual unit output, and real-time SOC of energy storage obtained from the grid measurement system.

[0110] 2. The action space of a real-time deep reinforcement learning agent comprises two dimensions: the controlled object and the decision characteristics. The controlled object dimension includes conventional generating units, rapid energy storage systems, and renewable energy power plants. The decision characteristics dimension consists of rapid actions, which have fast response times and are used for balance and stability control at the second / minute level. Specifically, these include AGC (Automatic Generation Control) commands, sent to generating units participating in automatic generation control; rapid energy storage control commands, which directly set the charging and discharging power of energy storage; and renewable energy emergency control commands, which rapidly limit the power output of renewable energy power plants.

[0111] 3. Reward function for real-time deep reinforcement learning agents It includes a frequency stability bonus, a fluctuation smoothing bonus, and a device lifespan bonus, calculated using the following formula:

[0112]

[0113] in: The system's real-time frequency deviation is obtained from the power grid measurement system; This represents the power change of the power grid interconnection line within adjacent control cycles. The rated total power of the system is used for... Standardize; It represents the absolute value of the change in the state of charge of the energy storage within a single control cycle; This is the frequency deviation penalty coefficient, used to balance the weight of frequency stability in the total reward; This is a power fluctuation penalty factor used to suppress frequent fluctuations in tie-line power. This is the energy storage lifespan loss penalty coefficient, used to balance control effectiveness and the lifespan of energy storage equipment.

[0114] This invention employs a power grid dynamic simulation environment built upon a long short-term memory network. Compared to methods relying on traditional physical equation simulations or simplified linear models, this significantly improves training speed while maintaining accuracy, providing an efficient and secure training platform for deep reinforcement learning agents. By defining the state space, action space, and reward function for three agents—day-ahead, intraday, and real-time—it achieves refined division of labor and collaboration in scheduling decisions across three dimensions: time scale, controlled object, and decision characteristics. The day-ahead agent makes strategic critical decisions based on point predictions, and its reward function incorporates risk penalties based on intraday fluctuation ranges, reserving safety margins. The intraday agent makes tactical adjustment decisions based on fluctuation ranges, and its reward function proactively prevents congestion risks through real-time interval predictions. The real-time agent performs rapid execution control based on multi-point and multi-interval information, and its reward function achieves multi-objective collaborative optimization of frequency stability, fluctuation smoothing, and equipment lifespan. This architecture enables the method of this invention to achieve intelligent collaborative control throughout the entire process from global planning to instantaneous execution, effectively improving the economy, safety, and stability of power grid operation under high-proportion renewable energy integration.

[0115] It is understood that those skilled in the art can combine various implementation methods in the above embodiments under the guidance of the above examples to obtain technical solutions with multiple implementation methods.

[0116] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-time-scale intelligent scheduling method for provincial power grids with new energy access, characterized in that, Scheduling is performed using a multi-agent deep reinforcement learning model, which consists of day-ahead deep reinforcement learning agents, intraday deep reinforcement learning agents, and real-time deep reinforcement learning agents. This includes: Obtain real-time meteorological data, real-time satellite cloud image data, and real-time power grid operation data of relevant new energy power stations in the provincial power grid at the current moment; Based on the acquired real-time data, the trained multi-timescale prediction model is used to generate the prediction results for the current moment. The prediction results include the new energy power point prediction curve at the day-ahead timescale, the new energy power fluctuation range prediction results at the intraday timescale, and the new energy power point prediction curve and new energy power fluctuation range prediction results at the real-time timescale. The predicted power point curves of renewable energy sources at the day-ahead time scale, along with load forecast data and grid topology data obtained from the grid dispatch system, are input into a trained day-ahead deep reinforcement learning agent, which outputs day-ahead dispatch instructions. The reward function of the day-ahead deep reinforcement learning agent includes an operating cost reward, a renewable energy curtailment penalty, and a cybersecurity risk penalty. The cybersecurity risk penalty... The calculation formula is: in: This represents the total number of critical transmission lines in the power grid. For the lines calculated based on the day-ahead scheduling plan The transmission power; For the line The power transmission limit; Based on the intraday timescale prediction results of renewable energy power fluctuation range, the power transfer distribution factor is used to calculate the line potential impact of renewable energy power fluctuations. The amplitude of transmission power fluctuation; This refers to the risk confidence level coefficient. For the line The penalty coefficient for exceeding the limit; Based on the day-ahead scheduling instructions, the predicted results of the new energy power fluctuation range at the intraday time scale are input into the trained intraday deep reinforcement learning agent using a rolling optimization method, and intraday adjustment instructions are output. The reward function of the intraday deep reinforcement learning agent includes a plan tracking reward and a congestion prevention reward. The calculation formula is: in: After executing the intraday adjustment order, the line Predicted transmission power; For the line Real-time power transmission safety margin; For the line The ground-state transmission power at the current moment; This represents the real-time safety margin coefficient. For the line The weight of the blocking prevention reward; Based on the day-ahead scheduling and intraday adjustment instructions, combined with the real-time time-scale renewable energy power point prediction curve, renewable energy power fluctuation range prediction results, and real-time grid frequency deviation obtained from the grid measurement system, a trained real-time deep reinforcement learning agent outputs real-time control instructions. The reward function of the real-time deep reinforcement learning agent is calculated based on the state space, which includes the real-time time-scale renewable energy power point prediction curve and fluctuation range prediction results. It includes a frequency stability bonus, a fluctuation smoothing bonus, and a device lifespan bonus, calculated using the following formula: in: The system's real-time frequency deviation is obtained from the power grid measurement system; This represents the power change of the power grid interconnection line within adjacent control cycles. This refers to the system's rated total power. It represents the absolute value of the change in the state of charge of the energy storage within a single control cycle; This is the frequency deviation penalty coefficient; This is the power fluctuation penalty coefficient; This is the penalty coefficient for energy storage lifespan loss.

2. The provincial power grid multi-timescale intelligent scheduling method for new energy access according to claim 1, characterized in that, The multi-timescale prediction model is constructed using a Long Short-Term Memory (LSTM) network model. The construction and training process includes: Acquire historical meteorological data, historical satellite cloud image data, and historical power grid operation data; Based on the acquired historical data, a day-ahead timescale prediction model, an intraday timescale prediction model, and a real-time timescale prediction model were trained respectively; among which: The day-ahead timescale prediction model is used to output a new energy power point prediction curve with a first preset time period as the period. The intraday timescale prediction model is used to output the prediction results of the new energy power fluctuation range with the second preset time period as the cycle. The real-time time-scale prediction model is used to output the prediction curve of new energy power points and the prediction results of new energy power fluctuation range with a third preset time period as the cycle.

3. The provincial power grid multi-timescale intelligent scheduling method for new energy access according to claim 2, characterized in that, The Long Short-Term Memory (LSTM) network model is also used to construct a dynamic power grid simulation environment. The construction process includes: An LSTM state transition model is trained based on the historical power grid operation data. The LSTM state transition model takes the power grid operation state and dispatch instructions of the previous moment as input and outputs the predicted power grid operation state of the next moment. The trained LSTM state transition model is used to construct a power grid dynamic simulation environment. The power grid dynamic simulation environment receives scheduling instructions, adjustment instructions or control instructions output by the deep reinforcement learning agent, simulates the execution of instructions based on the current power grid operating state, and returns new power grid operating state data.

4. The provincial power grid multi-timescale intelligent scheduling method for new energy access according to claim 3, characterized in that, The day-ahead deep reinforcement learning agent, intraday deep reinforcement learning agent, and real-time deep reinforcement learning agent are trained in the power grid dynamic simulation environment. The training process includes: In the power grid dynamic simulation environment, the historical meteorological data, historical satellite cloud image data, and historical power grid operation data are used to train the day-ahead deep reinforcement learning agent. Based on the pre-day deep reinforcement learning agent that has completed fixed training, the intra-day deep reinforcement learning agent is trained using the historical meteorological data, historical satellite cloud image data, and historical power grid operation data. Based on the pre-day and intra-day deep reinforcement learning agents that have completed fixed training, the real-time deep reinforcement learning agents are trained using the historical meteorological data, historical satellite cloud image data, and historical power grid operation data.

5. The provincial power grid multi-timescale intelligent scheduling method for new energy access according to claim 1, characterized in that, The action space of the aforementioned deep reinforcement learning agent includes key action decisions, which include conventional unit start-up and shutdown instructions, new energy power plant output plan instructions, and inter-provincial transaction instructions.

6. The provincial power grid multi-timescale intelligent scheduling method for new energy access according to claim 1, characterized in that, The action space of the intraday deep reinforcement learning agent includes adjustment action instructions, which include unit output adjustment instructions, energy storage plan optimization instructions, and demand response activation instructions.

7. The provincial power grid multi-timescale intelligent scheduling method for new energy access according to claim 1, characterized in that, The action space of the real-time deep reinforcement learning agent includes rapid control commands, which include AGC adjustment commands, rapid energy storage control commands, and new energy emergency control commands.

Images