A Power System Dispatch Optimization Method Based on Deep Reinforcement Learning and Its System

Abstract

This invention discloses a power system dispatch optimization method and system based on deep reinforcement learning, belonging to the field of smart grid optimization and dispatch technology. It includes: receiving and synchronizing real-time operational data, meteorological data, equipment health indicators, and renewable energy output data; constructing a power grid diagram and generating a node time-series matrix, encoding it as a multi-temporal embedding vector; generating short-term output prediction values ​​and uncertainty indicators based on meteorological and renewable energy output data, converting them into compensation factors; calculating risk scores based on equipment health indicators and meteorological data and mapping them to dynamic weights; inputting the multi-temporal embedding vectors, compensation factors, and dynamic weights into a deep reinforcement learning model to generate a dispatch strategy and performing feasibility verification; if the verification passes, the strategy is executed and the results are returned for online model updates. This invention effectively improves the security and robustness of power system dispatch by integrating multi-source data and a risk perception mechanism.

Description

Technical Field

[0001] This invention relates to the field of smart grid optimization scheduling technology, specifically to a power system scheduling optimization method and system based on deep reinforcement learning. Background Technology

[0002] Power system dispatch optimization is a core technology for ensuring the safe and stable operation of the power grid and improving energy efficiency. Balancing economic efficiency, environmental protection, and power supply reliability has become a key challenge in dispatch decision-making. In recent years, with the development of artificial intelligence technology, data-driven optimization methods have been gradually applied to the field of power system dispatch, and by integrating real-time operational data, weather forecasts, and load characteristics, dynamic adjustments to dispatch strategies can be achieved.

[0003] In existing technologies, some traditional scheduling methods have introduced reinforcement learning techniques to improve dynamic response capabilities. These methods optimize strategies through interactive learning with the power grid environment. However, they often use time-series data or simplified power grid topology as state inputs, making it difficult to dynamically adjust optimization targets based on equipment health risks (such as increased failure rates due to extreme weather) and the real-time carbon emission intensity of renewable energy. This results in scheduling strategies that, after generation, may sacrifice power supply reliability in high-risk scenarios due to excessive pursuit of economic efficiency. Furthermore, insufficient consideration of carbon constraints leads to a lack of environmental friendliness in the scheduling schemes, and there is a lack of efficient feasibility verification and correction mechanisms.

[0004] Therefore, there is an urgent need in this field for a deep reinforcement learning scheduling model to improve the dynamic optimization decision-making level of power systems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a power system scheduling optimization method and system based on deep reinforcement learning.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] In a first aspect, this invention discloses a power system scheduling optimization method based on deep reinforcement learning, comprising the following steps:

[0008] Receive multi-source data and transmit it synchronously according to timestamps; the multi-source data includes real-time operation data, meteorological data, equipment health indicators and renewable energy output data;

[0009] The multi-source data is combined with the node and line information of the power grid to construct a power grid map, and a node time series matrix is ​​generated as the node attribute of the power grid. The power grid map and the node attribute are encoded into a multi-temporal embedding vector.

[0010] Based on the meteorological data and the renewable energy output data, short-term output forecasts and uncertainty indices are generated and converted into compensation factors;

[0011] A risk score is calculated based on the equipment health indicators and the meteorological data, and then mapped to a dynamic weight.

[0012] The multi-temporal embedding vector, the compensation factor, and the dynamic weights are used to generate a scheduling strategy through a deep reinforcement learning model, and the feasibility is verified.

[0013] If the verification fails, the action sequence output by the scheduling strategy is corrected using the Euclidean projection algorithm; if the verification passes, the action is executed and the result is sent back, and the scheduling strategy is updated online.

[0014] Secondly, this invention discloses a power system dispatch optimization system based on deep reinforcement learning, which uses the aforementioned power system dispatch optimization method based on deep reinforcement learning, including:

[0015] The data acquisition module is used to receive multi-source data and transmit it synchronously according to timestamps; the multi-source data includes real-time operation data, meteorological data, equipment health indicators and renewable energy output data;

[0016] The graph modeling module is used to combine the multi-source data with the node and line information of the power grid to construct a power grid graph, generate a node time series matrix as the node attribute of the power grid, and encode the power grid graph and the node attribute into a multi-temporal embedding vector.

[0017] The prediction and compensation module is used to generate short-term output prediction values ​​and uncertainty indicators based on the meteorological data and the renewable energy output data, and convert them into compensation factors;

[0018] The risk assessment module is used to calculate a risk score based on the equipment health indicators and the meteorological data and map it to a dynamic weight.

[0019] The scheduling decision module is used to generate a scheduling strategy by using the multi-temporal embedding vector, the compensation factor and the dynamic weight through a deep reinforcement learning model, and to verify its feasibility.

[0020] The execution and update module is used to correct the action sequence output by the scheduling strategy using the Euclidean projection algorithm when the verification fails; when the verification passes, it is executed and the result is sent back and the scheduling strategy is updated online.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0022] 1. This invention introduces unified modeling and spatiotemporal embedding of multi-source data, enabling scheduling optimization to reflect not only the power grid structure and real-time operating status, but also external meteorological environment and renewable energy fluctuation characteristics. This approach avoids the information gaps caused by traditional methods relying on a single data source, allowing scheduling strategies to obtain more complete support at the data level, thereby more accurately characterizing the power grid operating status in complex environments.

[0023] 2. In the process of generating the scheduling strategy, this invention utilizes the uncertainty of output prediction to construct a compensation factor, and combines the risk score calculated from equipment health indicators and meteorological data to map it as a dynamic weight. This allows the model to dynamically adjust the focus of the optimization objective according to different operational risk conditions. This design overcomes the limitation of existing methods in balancing economy and safety in scheduling, enabling the generated strategy to ensure system safety while taking into account operational efficiency, thus improving the overall scheduling effect.

[0024] 3. This invention incorporates a feasibility verification and correction mechanism based on Euclidean projection before the scheduling strategy is implemented, ensuring that the generated scheduling strategy necessarily falls within the feasible solution domain. This avoids the problem of scheduling instructions being unenforceable due to optimization model deviations or data fluctuations. Through real-time feedback of execution results and online model updates, this invention can continuously correct and improve model performance in a dynamic environment, achieving continuous evolution of scheduling optimization, thereby providing a more robust and adaptable intelligent scheduling solution for power systems. Attached Figure Description

[0025] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0026] Figure 1 This is a flowchart of the steps of the present invention;

[0027] Figure 2 This is a flowchart illustrating the working principle of the present invention;

[0028] Figure 3 This is a flowchart illustrating the power grid diagram construction and coding refinement process of the present invention.

[0029] Figure 4 This is a flowchart of the DRL decision-making and feasibility verification cycle of the present invention;

[0030] Figure 5 This is a diagram of the online update security mechanism of the present invention;

[0031] Figure 6 This is a diagram of the overall hardware architecture of the system of the present invention. Detailed Implementation

[0032] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0033] In existing technologies, power system dispatching largely relies on traditional optimization algorithms or single data source drivers, making it difficult to simultaneously address the volatility of high-proportion renewable energy grid connections, equipment operation risks, and carbon emission reduction targets. Traditional methods are susceptible to multiple constraint conflicts during extreme weather or load surges, leading to optimization results deviating from the practically feasible domain. Existing systems lack deep understanding of the spatiotemporal coupling characteristics of the power grid; especially when renewable energy output fluctuates drastically, static weighting models exhibit systematic biases, making it difficult to achieve multi-objective dynamic equilibrium.

[0034] To address the aforementioned issues, the inventors discovered a coupling relationship between power grid topology and carbon emission flow distribution, and implemented a spatial representation of carbon emissions by constructing an environmentally weighted adjacency matrix. During the research, it was found that there is a quantitative correlation between line carbon flow density and electrical parameters; the main power angle stability constraint is highly sensitive to topology changes, while the reserve capacity constraint, although less sensitive, exhibits good robustness. Therefore, a method for dynamically adjusting optimization weights based on risk thresholds was proposed. Further simulation verification incorporated the mapping relationship between risk scores and multi-constraint relaxations into the reward function's dynamic adjustment mechanism, forming an adaptive decision-making system.

[0035] Specifically, the scheduling system first synchronously collects multi-source data and constructs a spatiotemporal graph structure. By analyzing the node time-series matrix and line carbon flow density, it generates a power grid embedding vector that incorporates environmental protection characteristics. Simultaneously, it extracts output uncertainty from the prediction data as a compensation factor. When the risk score exceeds a set threshold, the system automatically switches to a reliability-first weight configuration, utilizing the relaxation property of reserve capacity constraints and combining it with a pre-established constraint-risk mapping model to correct the scheduling strategy. During continuous decision-making, the system updates the deep reinforcement learning model parameters in real time based on the temporal changes in the strategy execution results through an experience replay mechanism, forming an online optimization closed loop. For cases where the risk threshold is not exceeded, the standard optimization mode prioritizing economy continues to output the scheduling strategy.

[0036] Compared to existing technologies, traditional methods rely on fixed-weight models and lack multi-constraint coordination mechanisms, making them prone to policy infeasibility issues when multiple risks overlap. This solution innovatively integrates multi-spatiotemporal graph encoding and dynamic risk perception, achieving real-time correction of scheduling strategies by establishing a constraint-risk coupling model. Unlike static optimization models, this solution can intelligently switch optimization modes based on real-time risk levels and continuously update the decision model through online learning, significantly improving scheduling reliability in high-uncertainty scenarios.

[0037] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0038] Example 1:

[0039] like Figure 1 As shown, a power system scheduling optimization method based on deep reinforcement learning includes the following steps:

[0040] It receives multi-source data and transmits it synchronously according to timestamps; the multi-source data includes real-time operation data, meteorological data, equipment health indicators, and renewable energy output data;

[0041] A power grid graph is constructed by combining multi-source data with the node and line information of the power grid, and a node time series matrix is ​​generated as the node attribute of the power grid. The power grid graph and node attributes are encoded into multi-temporal embedding vectors.

[0042] Short-term power output forecasts and uncertainty indices are generated based on meteorological data and renewable energy output data, and then converted into compensation factors.

[0043] Risk scores are calculated based on equipment health indicators and meteorological data and mapped to dynamic weights.

[0044] The scheduling strategy is generated by using multi-temporal embedding vectors, compensation factors, and dynamic weights through a deep reinforcement learning model, and its feasibility is verified.

[0045] When the verification fails, the action sequence output by the scheduling policy is corrected using the Euclidean projection algorithm; when the verification passes, the action is executed and the result is sent back, and the scheduling policy is updated online.

[0046] Multi-source data synchronization refers to aligning data from different sources (real-time operational data, meteorological data, etc.) according to a unified timestamp to ensure data consistency in time. For example, the Network Time Protocol (NTP) can be used to achieve millisecond-level time synchronization. The specific steps for receiving multi-source data and synchronizing it according to timestamps include:

[0047] Using real-time running data as the baseline timeline, all data from multiple sources are aligned to a unified timestamp sequence through interpolation.

[0048] The renewable energy volatility index VI is calculated as σ(ΔP) / P, where σ(ΔP) is the standard deviation of the renewable energy output data, and P is the rated renewable energy output data.

[0049] Dynamically select either the ARIMA model or the historical mean completion algorithm based on the renewable energy volatility index VI:

[0050] When VI > 0.3, the ARIMA model is used to predict and complete the missing values ​​of the multi-source data;

[0051] When VI≤0.3, the missing values ​​of the multi-source data are completed using the historical average completion algorithm for the same scene.

[0052] Output synchronized multi-source data.

[0053] like Figure 2 As shown, the working principle of this application is as follows: In the power system dispatch center, real-time operating data (such as bus voltage and line power flow) is acquired through a Supervisory Control and Data Acquisition (SCADA) system or a synchronous phasor measurement unit (PMU). Simultaneously, minute-by-minute updated meteorological data (such as wind speed and solar irradiance) is acquired from the meteorological service interface, equipment health indicators (such as transformer oil temperature and circuit breaker status) are acquired from the equipment monitoring system, and real-time output data of wind farms / photovoltaic power plants are acquired from the renewable energy monitoring system. These multi-source data are timestamped through high-precision clock synchronization (such as the PTP protocol) to ensure data temporal consistency.

[0054] A power grid graph structure is constructed based on the power grid topology (node ​​and line connection relationships), where each node is appended with a node time-series matrix generated in the previous step as an attribute. The power grid graph refers to a graph structure constructed with power grid nodes (buses, generators, loads, etc.) as vertices and transmission lines as edges. The node attribute is the node time-series matrix, and the edges may contain electrical parameters such as impedance. The node time-series matrix represents a matrix formed by load, voltage, and other data at multiple time granularities (e.g., 5 minutes, 1 hour) for each node, used to characterize the temporal behavior of the node. Then, a spatiotemporal graph convolutional network (ST-GCN) is used to encode the power grid graph and its node attributes into multi-spatiotemporal embedding vectors. These multi-spatiotemporal embedding vectors are then encoded using a graph neural network (GNN) to obtain a vector representation of each node, simultaneously fusing the temporal features and spatial topological relationships of the nodes. This vector captures both the spatial structure of the power grid and the temporal evolution characteristics of the nodes.

[0055] Based on meteorological data and historical renewable energy output data, a time series forecasting model (such as LSTM or Transformer) is used to predict renewable energy output for the next 5-60 minutes, and the uncertainty of the prediction (such as standard deviation) is calculated. The uncertainty index is converted into a compensation factor (derived from the uncertainty index of renewable energy output forecast, used to adjust the sensitivity to renewable energy fluctuations in the reinforcement learning state; for example, when the uncertainty is high, the compensation factor will reduce the impact of renewable energy nodes; for example, compensation factor = 1 / (1+uncertainty)). This is used to adjust the focus on renewable energy nodes in the reinforcement learning state; when the uncertainty is high, the compensation factor will decrease, reducing the dependence on highly volatile renewable energy nodes.

[0056] A risk score is calculated based on equipment health indicators and meteorological data (such as thunderstorm and high-temperature warnings). (For example, the equipment risk score is a weighted sum of health indicators, meteorological risk indicators, and other indicators, normalized to 0-1). This risk score is mapped to dynamic weights, which are used to adjust the weight ratios of economy, reliability, and environmental friendliness in the reinforcement learning reward function. The reliability weight is increased for higher risks. For example, when the risk score is <0.3, the economy weight is 0.6, the reliability weight is 0.3, and the environmental weight is 0.1; when the risk score is >0.7, the reliability weight increases to 0.7, the economy weight decreases to 0.2, and the environmental weight remains at 0.1.

[0057] Multi-temporal embedding vectors, compensation factors, and dynamic weights are input into a deep reinforcement learning model. The model output includes action sequences such as output adjustment of each generator, energy storage charging and discharging commands, and controllable load adjustments. This scheduling strategy needs to meet the safety requirements of power system operation, therefore, a feasibility verification is performed. The feasibility verification involves verifying the power system safety constraints of the reinforcement learning-output scheduling strategy, including power flow, voltage, and equipment capacity.

[0058] Feasibility verification is conducted based on real-time power flow constraints, equipment capacity limitations, reserve capacity requirements, and carbon emission constraints (i.e., the four-dimensional constraint space D). If the verification fails, the Euclidean projection algorithm is used to correct the action sequence to the boundary of the feasible region (i.e., when the scheduling strategy does not meet the constraints, the strategy actions are projected to the nearest feasible point that meets the constraints (in the sense of Euclidean distance)).

[0059] If the verification passes, an execution scheduling command is issued. After the command is executed, the execution results (such as actual output, line power flow, and equipment status) are collected and compared with the expected results. The results are then fed back to the deep reinforcement learning model for online updates, such as using new empirical data to update the policy network parameters, thereby achieving continuous optimization of the policy.

[0060] This scheme utilizes a deep reinforcement learning model combined with multi-source data and multi-dimensional constraints to generate dispatch strategies while considering renewable energy fluctuations and operational risks. Compared to traditional dispatch methods, this scheme can dynamically balance economic efficiency, environmental friendliness, and reliability, and continuously adapt to grid changes through online learning. Simultaneously, feasibility verification and Euclidean projection ensure the safe execution of the dispatch strategy. This application enables intelligent power system dispatch supported by multi-source information, not only improving the renewable energy absorption rate but also enhancing the safety and robustness of power system operation.

[0061] like Figure 3 As shown, this application further proposes the following specific steps for constructing a power grid map by combining multi-source data with the node and line information of the power grid, and generating a node time series matrix as the node attribute of the power grid:

[0062] Constructing the adjacency matrix of the power grid topology This identifies the connection relationship between node i and node j in the power grid;

[0063] Multi-source data also includes the unit's carbon emission coefficient. Obtain the unit's carbon emission coefficient And calculate the line carbon flux density of the power grid. ;

[0064] in For the unit On the line The power components are obtained through a power flow tracing algorithm; For line power flow; G represents the set of all generators in the system;

[0065] Based on the grid topology adjacency matrix A and the line carbon flux density Generate an environmentally friendly weighted adjacency matrix Using the node temporal matrix as a vertex attribute, carbon flux density As an edge attribute embedded in the power grid graph, where α is a preset environmental protection weight coefficient. To balance economic efficiency and environmental protection.

[0066] The power grid graph constructed in this application is a dynamic topology model that incorporates environmental protection attributes. Through spatial carbon flow density quantification and weighting mechanisms, carbon emission constraints are transformed into computable graph structure features. The system obtains node connection relationships from the power grid GIS platform to generate a basic power grid topology adjacency matrix. At the same time, the carbon emission coefficient of the unit is collected through the carbon monitoring system. The carbon flow density of the line is calculated by combining the power flow tracing algorithm. Line carbon flux density Carbon emission intensity, which reflects the unit of transmitted power, is a spatial carrier of environmental constraints.

[0067] To strengthen the influence of carbon constraints in decision-making, the carbon flux density of the transmission line will be increased. As edge weights, they are injected into the adjacency matrix to generate the environmental protection weighted matrix. The node time sequence matrix carries dynamic characteristics such as bus voltage fluctuations and load change rates. The time sequence length of the node time sequence matrix is ​​calculated based on the scheduling cycle (15 minutes / step) × prediction duration (24 hours) and embedded as a vertex attribute in the graph. In peak shaving scenarios, when the carbon flow density of a high-carbon line... >0.6 When a power source is classified as high-carbon emission and should be prioritized for avoidance during scheduling, the preset environmental weight coefficient α is automatically set to the upper limit of 0.5, which significantly increases the weight of that edge in the topology model and drives the scheduling algorithm to actively avoid high-carbon paths.

[0068] This application embeds environmental attributes into the power grid topology. By dynamically adjusting the coefficient α∈[0.1,0.5], carbon transmission intensity is quantified and embedded in electrical connections, significantly increasing the weight of high-carbon lines in the adjacency matrix. This design transforms abstract carbon emission constraints into computable spatial parameters, guiding the scheduling algorithm to proactively avoid high-carbon paths when generating strategies, effectively overcoming the core bottleneck of traditional scheduling's difficulty in spatially implementing environmental constraints. By transforming abstract carbon constraints into a weighted graph structure, the system constructs a foundation for the collaborative mapping of power flow and carbon emission, solving the key deficiency of traditional scheduling's inability to perceive carbon spatial distribution.

[0069] This application further proposes that the construction of the power grid diagram further includes:

[0070] For each line Constructing 3D edge feature vectors: ;

[0071] in, Carbon flux density; This refers to the per-unit value of the line impedance; The voltage phase difference between nodes i and j;

[0072] Generate edge weight association model ;

[0073] Where a, b, and c are adjustable coefficients, satisfying a + b + c = 1;

[0074] Environmental weighted adjacency matrix Construct a trend-environmental collaborative adjacency matrix:

[0075] .

[0076] Based on the environmental weighted matrix, this application further integrates electrical safety constraints to construct a three-dimensional edge feature vector of quantifiable risk. Among them, the per-unit impedance value Obtained from the SCADA historical database, reflecting line transmission capacity; voltage phase difference between nodes. The system stability margin is updated in milliseconds by the PMU device and is indicated in real time.

[0077] These three sets of heterogeneous physical quantities are obtained through a linear correlation model. Aggregates into a unified weight, linear correlation model Specifically, it can be expressed as:

[0078]

[0079] The adjustable coefficients a, b, and c are configured according to electrical operation rules: when the system is in the N-1 safety check state, (a, b, c) = (0.2, 0.7, 0.1) are used to strengthen safety constraints; during carbon emission control periods, (0.5, 0.3, 0.2) are used to emphasize environmental protection. This ultimately generates a power flow-environmental protection synergy matrix. .

[0080] This application utilizes the fusion of carbon flux density. (Environmental constraints), reciprocal of impedance (Transmission capability) and voltage phase difference (Stability margin) A linear correlation model was constructed using three heterogeneous physical quantities. The model possesses triple dynamic adjustment capabilities: when the system performs N-1 verification, the coefficient combination (0.2, 0.7, 0.1) prioritizes ensuring line safety margin; during carbon emission control periods, the weighted configuration (0.5, 0.3, 0.2) is switched to highlight environmental protection orientation; when... When the temperature exceeds 8°, the weight of the adjustable coefficient c is automatically increased to prevent voltage collapse. A multi-physics coupling mechanism enables power flow optimization to overcome the limitations of traditional electrical constraints.

[0081] This application further proposes that the specific steps for generating short-term power output forecasts and uncertainty indices based on meteorological data and renewable energy output data, and converting them into compensation factors, include:

[0082] Based on meteorological data and renewable energy output data, a time-series forecast model is used to output short-term output forecasts for the next 5-60 minutes. and uncertainty index ;

[0083] like Then a compensation factor is generated. Where Z is the preset fluctuation threshold. The preset coefficient is used; otherwise, a compensation factor is defined. ;

[0084] Compensation factor Used to decay the state values ​​of renewable energy nodes in a multi-temporal embedding vector.

[0085] The power output forecast average for the next 5-60 minutes is output through a time series forecasting model (LSTM network). and uncertainty index (standard deviation) When predicting the intensity of fluctuations When the value exceeds Z (preset fluctuation threshold), a dynamic compensation factor is generated. This compensation factor acts on the state values ​​of renewable energy nodes in the multi-temporal embedding vector, through the state decay equation. Adjust the node weights (k is a preset sensitivity coefficient, with a value range of [0.8, 1.2]). This represents the original state value of the node in the multi-temporal embedding vector. (This refers to the corrected node state value).

[0086] For example, sudden changes in wind speed lead to When the fluctuation exceeds the preset fluctuation threshold Z, the system automatically calculates... =0.85, which causes the renewable energy node's state value to decay by 15%, reducing the node's priority in the scheduling strategy.

[0087] This scheme generates compensation factors by quantifying uncertainty indicators, and corrects the state weights of renewable energy nodes in multi-temporal and spatial embedding vectors in real time, thus addressing the core deficiency of traditional scheduling in its inability to respond to second-level output fluctuations. When sudden changes in wind speed lead to a significant increase in the prediction standard deviation... When the threshold is exceeded, the compensation factor automatically decays the node state value, guiding the scheduling algorithm to prioritize the use of energy storage or thermal power as backup, thus avoiding power flow over-limit accidents caused by volatility.

[0088] This application further proposes to output short-term power output forecasts for the next 5-60 minutes using a time-series forecasting model. and uncertainty index The specific steps include:

[0089] The short-term output forecast is generated using a multi-step rolling forecasting method, which executes the output forecast for the next 60 minutes in 5-minute increments.

[0090] After each prediction step in the time-series forecasting model, Kalman filtering and meteorological data are used to correct the model, yielding the final short-term power output forecast. .

[0091] This application uses a 5-minute time unit to perform rolling forecasts of renewable energy output for the next 60 minutes. The core forecasting algorithm employs a recursive least squares (RLS) algorithm with a forgetting factor, utilizing a state-space model. The predicted sequence is updated recursively, where, Let be the state vector at time t-1. Here is the state transition matrix. , Autocorrelation coefficient contributing to historical performance; Vector of historical contribution , For real-time wind speed, Let B be the irradiance, and B be the input control matrix. , This refers to the wind speed-to-power conversion factor. For process noise, ∼N(0,R).

[0092] After completing a single-step prediction, the system receives second-level wind speed / irradiance data from the weather radar in real time, and then uses the Kalman gain matrix... The formula for dynamically correcting the predicted value is as follows:

[0093] ;

[0094] The observation matrix H maps meteorological variables to the output space, and the parameters of the Kalman gain formula are defined as follows:

[0095]

[0096] In high-gain scenarios, sudden weather changes lead to A sudden increase (decreased predictive reliability) triggers automatic system upgrades. This assigns higher weight to real-time observations. In low-gain scenarios, when sensor noise R increases (such as when sand and dust cover photovoltaic panels), the system reduces Kt to suppress noise interference.

[0097] In one embodiment, the technical effectiveness is verified through a provincial power grid case study, as shown in the table below:

[0098]

[0099] pass Co-tuning with R, Kalman gain In predicting reliability ( The dynamic trade-off between renewable energy output forecasting and observation reliability (R) ensures that renewable energy output forecasts maintain an accuracy of ±10% even under extreme weather conditions, providing a stable input for dispatch decisions.

[0100] This application adjusts photovoltaic (PV) forecasts downward based on real-time visibility data, proactively capturing actual output declines and keeping forecast errors within the confidence band. Through tight coupling of rolling forecasts and Kalman correction, renewable energy fluctuations caused by weather disturbances are transformed into quantifiable factors. Indicators, and then through compensation factors Changing node state values ​​ultimately affects the direction of deep reinforcement learning strategy generation. This application uses Kalman filtering to fuse real-time meteorological observation data and dynamically corrects multi-step rolling prediction results, solving the technical problem of traditional prediction models' lagging response to sudden weather changes. In typhoon scenarios, the system corrects wind power prediction curves based on radar precipitation data in a short time, effectively controlling the error of prediction points and providing highly reliable input for compensation factor generation.

[0101] like Figure 4 As shown, this application further proposes to generate a scheduling strategy by using a deep reinforcement learning model to integrate multi-temporal embedding vectors, compensation factors, and dynamic weights, and to conduct feasibility verification, which includes:

[0102] The feasible region is calculated based on power flow safety, equipment capacity, reserve capacity, and carbon emission scheduling strategies, and is defined as a four-dimensional constraint space. :

[0103] ;

[0104] If equipment capacity Exceeding limits but with spare capacity It is permissible to sacrifice reserve capacity to prioritize power supply;

[0105] If carbon emissions CI exceeds However, if the load rate of the power flow safety line is less than 70%, the carbon emission transfer compensation mechanism will be activated.

[0106] in, The maximum transmission capacity of the line is calculated dynamically based on the line's thermal stability limit and ambient temperature. and To maintain stable minimum and maximum power for the generator unit, This is the reserve capacity threshold. This represents the regional carbon emission target.

[0107] The carbon emission transfer compensation mechanism prioritizes increasing the output of low-carbon units such as photovoltaic / wind power and reducing the output of high-carbon units such as coal power, with the transfer amount not exceeding 20% ​​of the power flow of the transmission lines.

[0108] Among them, four-dimensional constrained space The parameter definition table is as follows:

[0109]

[0110] Current safety constraints Ensure that the line power does not exceed the thermal stability limit (conductor overheating) and static stability limit (voltage collapse threshold) to prevent line overload and meltdown or voltage collapse. For example, during typhoon weather. The temperature needs to be dynamically adjusted downwards based on the ambient temperature (e.g., at +40℃). (Reduction of 15%).

[0111] Equipment capacity constraints Ensure that the generator set operates under safe conditions, such as a turbine vibration margin of >30%, a boiler water-cooled wall temperature of <450℃, and a coal-fired unit of [300,600]MW.

[0112] Reserve capacity constraints Meeting the N-1 fault criterion (available power ≥ 5% of maximum load within 10 minutes), when When triggered, non-critical loads are cut off, and the gas turbine is quickly put into standby mode.

[0113] Under carbon emission constraints Comply with carbon quota policies (per unit of electricity generated) (Emissions ≤ Regional Annual Target), where, It can be dynamically adjusted. The data is updated monthly from the carbon trading platform API, as detailed in the table below:

[0114]

[0115] This application constructs a decision-making security barrier based on a four-dimensional constraint space D, dynamically coupling four constraint types: when the device capacity exceeds the limit ( When the standby capacity is available, If this occurs, the priority power supply protocol is activated, temporarily increasing the power supply capacity by reducing the reserve margin. Upper limit (not exceeding 110%) This also triggers carbon emission transfer compensation.

[0116] Match transmission channels with a line load rate of <70% using online hash indexing (load rate = ( ), transferring excess carbon emissions to low-carbon corridors.

[0117] This application integrates four-dimensional constraints into a unified mathematical expression. This system addresses the strategy oscillation problem caused by traditional step-by-step verification in dispatching. The four-dimensional constraint space D achieves a flexible balance between safety, economy, and environmental protection goals through a dynamic rule base, pioneering a coupled compensation mechanism for equipment capacity and reserve capacity, and a mathematical quantification model for carbon emission transfer. Compared to traditional static constraint verification, it improves efficiency in resolving conflicts between renewable energy consumption and line overload, ensuring stable operation of 220kV and above sections.

[0118] This application further proposes that, when the verification fails, the specific steps for correcting the action sequence output by the scheduling strategy using the Euclidean projection algorithm are as follows:

[0119] Acquire real-time temperature data, determine the health status of energy storage equipment through equipment health indicators, and calculate the rate of change of power adjustment on the lifespan loss of energy storage equipment based on the health status of energy storage equipment and real-time temperature data.

[0120] The action correction direction of the action sequence is determined based on minimizing lifetime loss. The action correction direction is the negative gradient direction where the rate of change decreases the fastest.

[0121] The original action sequence output by the scheduling strategy is adjusted along the negative gradient direction to satisfy the boundary conditions of the feasible region.

[0122] The rate of change of power adjustment with respect to the lifetime loss of energy storage equipment is calculated using the Euclidean projection algorithm to calculate the lifetime loss gradient, and the formula is as follows:

[0123] ;

[0124] The parameters are defined as follows:

[0125]

[0126] The ∇L mechanism is transformed from the chemical aging mechanism of the battery into an executable scheduling control law, and the average annual capacity decay rate of energy storage is reduced by using gradient projection.

[0127] This solution overcomes the limitation of traditional projection algorithms that ignore equipment aging, and for the first time transforms the electrochemical mechanism (temperature exponential term + capacity decay derivative) into the basis for gradient direction decision-making. Through SOH adaptive step size control, it improves energy storage life by 130% while ensuring scheduling flexibility, thus resolving the fundamental contradiction between frequent adjustments and equipment health.

[0128] When lifetime gradient projection triggers carbon emission transfer compensation, the system automatically records the quantitative relationship between ∇L and CI, generating an empirical matrix of compensation coefficients to update the historical parameter database. Empirical studies in the Jiangsu power grid demonstrate that this linkage reduces equipment lifespan loss under wind power fluctuation scenarios while maintaining carbon emission intensity, achieving cross-dimensional synergistic optimization of safety, lifespan, and environmental protection.

[0129] like Figure 5 As shown, this application further proposes that the specific steps for issuing the execution after verification, sending back the result, and updating the scheduling policy online include:

[0130] If the action sequence generated by the scheduling strategy satisfies the boundary constraints of the four-dimensional constraint space D, then it passes the verification.

[0131] Collect the actual execution results of the scheduling strategy, including generator output, energy storage charging and discharging status, line power flow and carbon emissions;

[0132] The actual execution results are stored in the experience replay pool, and data batches are extracted to calculate the strategy gradient.

[0133] The scheduling strategy is updated using a small step parameter η;

[0134] Verify the carbon emission change rate ΔE and load shedding rate change rate ΔL of the updated scheduling strategy in the historical scenario set;

[0135] If ΔE > 5% or ΔL > 3%, a safety rollback mechanism is triggered, rolling back to the previous safe scheduling strategy and triggering offline training.

[0136] The safe rollback mechanism also includes:

[0137] If two consecutive updates trigger a rollback, pause the online learning of the scheduling strategy and generate a device maintenance alarm;

[0138] After the rollback, the scheduling strategy was reinforced and trained for three cycles using historical safety strategy data.

[0139] This application constructs a dynamic evolution and security protection system for scheduling strategies, the core of which includes a triple protection mechanism: first, boundary verification of the four-dimensional constraint space D is performed; after the verification passes, actual operation data is collected. Stored in a distributed experience replay pool; key data batches are extracted through priority sampling, and the policy gradient is calculated. The strategy parameters are updated using a small step size parameter η.

[0140] ;

[0141] The updated scheduling strategy needs to undergo dual security verification in the historical scenario database: calculating the rate of change in carbon emissions. ( This is a predicted value for carbon emissions. (actual carbon emissions) and rate of change of load shedding rate ( This is the predicted load shedding rate. (This refers to the actual value of the load shedding rate). A Level 4 response is triggered when ΔE > 5% or ΔL > 3%.

[0142] Immediately roll back to the previous security policy version. ;

[0143] Mark abnormal update events and generate alarm codes;

[0144] If a rollback is triggered twice in a row, online learning will be paused and the equipment maintenance process will be initiated.

[0145] Inject historical security data to enhance training.

[0146] This application achieves safe evolution of the scheduling strategy. Small-step updates ensure learning stability, dual change rate thresholds construct a safe circuit breaker mechanism, and historical data reinforcement training quickly restores a reliable state, significantly shortening the strategy update failure recovery time.

[0147] This application further proposes that, in the scheduling strategy using a small step size parameter η, the small step size parameter η satisfies:

[0148] ;

[0149] in, The initial step size, ∈[0.005,0.02], where S is the risk score. λ is the adjustment coefficient, which adjusts the sensitivity of the risk score S to the decay of the small step parameter η, where λ∈[0.1,0.3].

[0150] The risk score S∈[0.2,1.0] comes from the fusion calculation of equipment health and meteorological risk (S<0.2 is set to S=0.2), and λ∈[0.1,0.3] is the sensitivity adjuster.

[0151] The risk score S is calculated using a multi-factor fusion weighted model, and the specific calculation formula is as follows:

[0152]

[0153] in, The equipment health sub-score is calculated from the Health Indicator (HI) of critical equipment, and its value ranges from [0, 1.0]. The meteorological risk sub-score is derived from real-time meteorological data and its warning level, and its value ranges from [0, 1.0]. Other sub-scores, such as grid structure risk and load forecast deviation risk, are used to calculate the compensation risk score S.

[0154] , Let be the weighting coefficient, satisfying + = 1. This value is determined based on engineering experience and regression analysis of historical failure data, and is typically set to 1. = 0.6, = 0.3, This is to focus more on reflecting the actual operating status of power grid equipment.

[0155] The small step size parameter η is implemented using a three-level control logic:

[0156] 1. Risk-based response:

[0157] Low risk (S<0.4): λ=0.1, maintaining η≈η0 accelerates convergence;

[0158] Medium risk (0.4≤S<0.7): λ=0.2, step size decay of 50-70% to prevent oscillation;

[0159] High risk (S≥0.7): λ=0.3, with a maximum decay to η. min =0.001;

[0160] 2. Hot-swap adjustment: For example, when a typhoon warning triggers a jump in the S value, η is adaptively reduced in order within 5 seconds by exponential decay exp(-λ·S);

[0161] 3. Equipment protection linkage: When the transformer oil temperature is >85℃, S is forced to 1.0, and η is reduced to the minimum value to protect the equipment.

[0162] This application achieves dynamic adjustment of the step size η according to the system stability requirements through coupled control of the risk score S and sensitivity λ. The exponential decay model solves the problem of policy oscillation in high-fluctuation scenarios, improving the step size control accuracy by 40 times and ensuring zero divergence accidents during training.

[0163] The following is a specific implementation of a power system scheduling optimization method based on deep reinforcement learning:

[0164] A provincial power grid dispatch center is monitoring a 220kV regional power grid with a high proportion of wind power (35% of installed capacity) and photovoltaic power (20% of installed capacity). This region includes 5 wind farms (total capacity 1.2GW), 3 photovoltaic power plants (total capacity 0.8GW), and 8 thermal power units (total capacity 2.5GW). Historically, the dispatch center has experienced high curtailment rates (average daily 12%) due to fluctuations in renewable energy output, and frequent equipment overload risks under extreme weather conditions.

[0165] The scheduling system deploys multi-source data acquisition terminals:

[0166] Real-time operating data: The bus voltage of each substation (e.g., the fluctuation range of 220kV bus voltage is 1.01-1.05pu) and the power flow of the lines (e.g., the average power flow of L1-L8 lines is 320MW, with a fluctuation range of ±50MW) are collected through PMU (synchronous phasor measurement device, sampling rate 50Hz).

[0167] Meteorological data: The wind farm is equipped with ultrasonic anemometers (sampling rate 1Hz, measurement range 0-60m / s) and irradiance sensors (accuracy ±5W / m²) to upload wind speed (e.g., 10-minute moving average wind speed 8.2m / s), temperature (25℃) and cloud cover (30%) in real time.

[0168] Equipment health indicators: The transformer is equipped with a fiber optic grating sensor to monitor the top oil temperature (68℃) and the degree of polymerization of the insulating paper (DP value 2800). The transmission line is equipped with a micro-wind vibration monitoring device (amplitude 0.12mm).

[0169] All data is synchronized in time via the IEEE 1588 PTP protocol (error < 1ms), generating a unified dataset with timestamps (e.g., "2025-08-20 14:00:00, wind speed at wind farm A 8.2m / s, power flow at line L3 315MW").

[0170] Construct a power grid map based on power grid GIS topology data (including 500 nodes and 720 lines):

[0171] Topology adjacency matrix: A 500×500 adjacency matrix is ​​generated by exporting node-line connection relationships through the EMS system. ;

[0172] Node time series matrix: Extract the dynamic features of each node over the past 24 hours (96 time steps × 12 features, such as load, voltage, and current). For example, the time series matrix of node 3 (wind farm grid connection point) shows that the output dropped from 450MW to 380MW in the past hour.

[0173] Environmental weighted adjacency matrix: Calculating line carbon flux density (coal-fired power units) =0.8 Line L3 =0.52 A weighted adjacency matrix is ​​generated by combining α=0.3 (environmental weight coefficient). = • (1+0.3×0.52), strengthen the weight of high-carbon lines in the topology model.

[0174] The power grid graph is encoded into a multi-temporal embedding vector (500×256 dimensions) using ST-GCN, which integrates node dynamic features and topological relationships.

[0175] An LSTM time-series forecasting model (input features: wind speed, irradiance, temperature; 128 hidden nodes) is used to predict renewable energy output over the next 60 minutes: Short-term output forecast for wind farm A over the next 15 minutes. =400MW, uncertainty index =25MW (preset fluctuation threshold Z=20MW);

[0176] because =25>Z=20, generate compensation factor =0.02×25=0.5 (k=0.02 is a preset coefficient), through the state decay equation Reduce the weight of wind farm nodes in the embedding vector to weaken their impact on the scheduling strategy.

[0177] The equipment health indicators are: transformer oil temperature 68℃ (threshold 85℃) corresponding to a health index of 0.8, and line vibration amplitude 0.12mm (threshold 0.2mm) corresponding to 0.9; the meteorological risk indicator is the 2-hour forecast of thunderstorms (level 3, weight 0.3); by integrating the equipment health indicators and meteorological data, the risk score S can be obtained: S=0.6×(1-0.8)+0.3×0.3+0.1×0.2=0.23 (equipment health accounts for 0.6, meteorological risk for 0.3, and others for 0.1). Since S<0.4 (low risk), the dynamic weights are set as follows: economy 0.6, reliability 0.3, and environmental protection 0.1.

[0178] Multi-temporal embedding vectors, compensation factors ( =0.5) and a dynamic weighted input deep reinforcement learning model (3-layer fully connected network, 256 hidden nodes):

[0179] The initial dispatch strategy is output, including the output of thermal power unit G1 +50MW, energy storage charging 20MW, and controllable load reduction 10MW;

[0180] Based on four-dimensional constrained space Feasibility verification was conducted, and it was found that the power flow of line L3 reached 340MW (maximum capacity 350MW, load rate 97%), which is close to the safety threshold.

[0181] Perform Euclidean projection correction and calculate lifetime loss gradient. Adjust the output of G1 to +30MW along the negative gradient direction, increase the energy storage charging to 30MW, and reduce the power flow of the line to 320MW (load rate 91%), thus meeting the constraints.

[0182] After the scheduling strategy was executed, the actual results were collected (G1 output 305MW, energy storage SOC 62%, line power flow 318MW, carbon intensity CI=0.58). Store in the experience replay pool:

[0183] Using a small step size parameter η= The model parameters were updated by updating exp(-λ·S)=0.01·exp(-0.1·0.23)≈0.0098. Historical scenario set tests showed that the carbon emission change rate ΔE=3% and the load shedding rate change rate ΔL=1.2%, neither of which triggered a rollback. After running continuously for 72 hours, the system adaptively reduced the load rate of high-carbon lines (from an average of 85% to 75%), the curtailment rate dropped to 6%, and the response time of dispatch instructions under typhoon weather was shortened to within 2 minutes.

[0184] This embodiment achieves dynamic scheduling optimization of a high-proportion renewable energy power grid through multi-source data fusion and deep reinforcement learning. During thunderstorms, the system adjusts energy storage charging and discharging plans 15 minutes in advance to avoid line overload; it dynamically decays the weight of renewable energy nodes through compensation factors to reduce the interference of prediction errors on the strategy; and Euclidean projection correction ensures that equipment meets safety constraints while minimizing lifespan losses. The overall solution synergistically improves the economy and safety of power grid operation, providing an integrated solution for renewable energy consumption and equipment health management.

[0185] Example 2:

[0186] like Figure 6 As shown, a power system dispatch optimization system based on deep reinforcement learning, using the aforementioned power system dispatch optimization method based on deep reinforcement learning, includes:

[0187] The data acquisition module is used to receive multi-source data and transmit it synchronously according to timestamps; the multi-source data includes real-time operation data, meteorological data, equipment health indicators and renewable energy output data;

[0188] The graph modeling module is used to combine multi-source data with the node and line information of the power grid to construct a power grid graph, and generate a node time series matrix as the node attribute of the power grid. The power grid graph and node attributes are encoded into multi-temporal embedding vectors.

[0189] The forecasting and compensation module is used to generate short-term power output forecasts and uncertainty indices based on meteorological data and renewable energy power output data, and convert them into compensation factors;

[0190] The risk assessment module is used to calculate risk scores based on equipment health indicators and meteorological data and map them to dynamic weights;

[0191] The scheduling decision module is used to generate scheduling strategies by using multi-temporal embedding vectors, compensation factors and dynamic weights through a deep reinforcement learning model, and to verify the feasibility of these strategies.

[0192] The execution and update module is used to correct the action sequence output by the scheduling strategy using the Euclidean projection algorithm when the verification fails; when the verification passes, it is executed and the result is sent back and the scheduling strategy is updated online.

[0193] The data acquisition module consists of a data acquisition unit, a communication interface, and a clock synchronization device. It is used to receive multi-source data from monitoring systems, meteorological monitoring equipment, equipment health monitoring terminals, and renewable energy forecasting devices, and to synchronize timestamps through a unified clock.

[0194] The graph modeling module consists of a computing processing unit and a memory. It is used to combine multi-source data with the node and line information of the power grid to construct a power grid graph, generate a node time series matrix, and obtain multi-temporal embedding vectors through an embedding coding algorithm.

[0195] The forecasting and compensation module includes a short-term forecasting engine and an uncertainty modeling unit, which are used to generate output forecasts based on meteorological and renewable energy data, calculate forecast uncertainties, and then convert them into compensation factors to be input into the scheduling model.

[0196] The risk assessment module consists of a data analysis unit and a risk calculation unit, which are used to integrate equipment health indicators and meteorological data, calculate equipment risk scores, and generate corresponding dynamic weights.

[0197] The scheduling decision module includes a deep reinforcement learning model processor and a constraint verification unit, used to generate scheduling policies and perform feasibility verification. The DRL (Deep Reinforcement Learning) model adopts a 3-layer fully connected network with 256 hidden nodes and ReLU activation.

[0198] The execution and update module consists of an execution controller, a communication interface, and an online update engine. It is used to call the Euclidean projection algorithm to correct the scheduling action when the verification fails; and to issue scheduling instructions to the execution end when the verification passes, and to send the execution feedback back to update the model parameters.

[0199] In one specific embodiment, the system collects grid operation data through a PMU (Phasor Measurement Unit) and receives wind speed, temperature, and irradiance data from meteorological monitoring stations, acquiring health status indicators of key transformers and transmission lines uploaded by equipment sensors. After the data acquisition module adds a unified timestamp, the data enters the graph modeling module to generate a grid graph and node time-series matrix. Subsequently, in the prediction and compensation module, renewable energy output is predicted based on wind speed and irradiance, and uncertainty is calculated. The risk assessment module generates risk scores and forms dynamic weights based on transformer aging indicators and temperature conditions. Finally, the scheduling decision module uses deep reinforcement learning to generate scheduling strategies. When current exceedances are detected, the Euclidean projection algorithm corrects the action to ensure the result meets grid constraints. This scheduling instruction is sent to the scheduling center via the execution and update module, and the feedback data is used for online updates of the reinforcement learning model.

[0200] This invention realizes intelligent optimization of power system dispatch based on multi-source data, which not only ensures the feasibility of dispatch strategies, but also enhances the ability to cope with renewable energy fluctuations and equipment risks, and significantly improves the safety and economy of power grid operation.

[0201] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A power system dispatch optimization method based on deep reinforcement learning, characterized in that: Includes the following steps: Receive multi-source data and transmit it synchronously according to timestamps; the multi-source data includes real-time operation data, meteorological data, equipment health indicators and renewable energy output data; The multi-source data is combined with the node and line information of the power grid to construct a power grid map, and a node time series matrix is ​​generated as the node attribute of the power grid. The power grid map and the node attribute are encoded into a multi-temporal embedding vector. Based on the meteorological data and the renewable energy output data, short-term output forecasts and uncertainty indices are generated and converted into compensation factors; A risk score is calculated based on the equipment health indicators and the meteorological data, and then mapped to a dynamic weight. The state values ​​of renewable energy nodes in the multi-temporal embedding vector are attenuated using the compensation factor to obtain a corrected multi-temporal embedding vector. The corrected multi-temporal embedding vector and the dynamic weights are input into a deep reinforcement learning model, wherein the dynamic weights are used to adjust the weight ratio of economic, reliability and environmental protection objectives in the reward function of the deep reinforcement learning model to generate a scheduling strategy and perform feasibility verification. If the verification fails, the action sequence output by the scheduling strategy is corrected using the Euclidean projection algorithm; if the verification passes, the action is executed and the result is sent back, and the scheduling strategy is updated online. The specific steps for correcting the action sequence output by the scheduling strategy using the Euclidean projection algorithm are as follows: Real-time temperature data is acquired, and the health status of the energy storage device is determined through the device health indicators. Based on the health status of the energy storage device and the real-time temperature data, the rate of change of power adjustment on the lifespan loss of the energy storage device is calculated. The action correction direction of the action sequence is determined based on minimizing lifetime loss, and the action correction direction is the negative gradient direction in which the rate of change decreases the fastest. The original action sequence output by the scheduling strategy is adjusted along the negative gradient direction to satisfy the boundary conditions of the feasible region.

2. The power system dispatch optimization method based on deep reinforcement learning according to claim 1, characterized in that: The specific steps for constructing a power grid map by combining the multi-source data with the node and line information of the power grid, and generating a node time series matrix as the node attribute of the power grid, include: Constructing the adjacency matrix of the power grid topology This identifies the connection relationship between node i and node j of the power grid; The multi-source data also includes the unit's carbon emission coefficient. Obtain the carbon emission coefficient of the unit. And calculate the line carbon flow density of the power grid. ; in For the unit On the line The power components are obtained through a power flow tracing algorithm; For line power flow; G represents the set of all generators in the system; Based on the power grid topology adjacency matrix A and the line carbon flux density Generate an environmentally friendly weighted adjacency matrix The node time series matrix is ​​used as a vertex attribute, and the carbon flux density is... The edge attribute is embedded in the power grid graph, where α is a preset environmental protection weight coefficient. .

3. The power system dispatch optimization method based on deep reinforcement learning according to claim 2, characterized in that: The construction of the power grid diagram further includes: For each line Constructing 3D edge feature vectors: ; in, The carbon flux density; This refers to the per-unit value of the line impedance; The voltage phase difference between nodes i and j; Generate edge weight association model ; Where a, b, and c are adjustable coefficients, satisfying a + b + c = 1; The environmental weighted adjacency matrix Construct a trend-environmental collaborative adjacency matrix: 。 4. The power system dispatch optimization method based on deep reinforcement learning according to claim 1, characterized in that: Feasibility verification includes: The feasible region of the scheduling strategy is calculated based on power flow safety, equipment capacity, reserve capacity, and carbon emissions, and the feasible region is defined as a four-dimensional constraint space. : ； If the equipment capacity Exceeding the limit but the aforementioned reserve capacity This allows for the sacrifice of the reserve capacity to prioritize power supply; If the carbon emissions CI exceed However, if the line load rate for the power flow safety is less than 70%, the carbon emission transfer compensation mechanism will be activated. in, Let be the active power transmitted from line i to line j. The maximum transmission capacity of the line is calculated dynamically based on the line's thermal stability limit and ambient temperature. and To maintain stable minimum and maximum power for the generator unit, This is the reserve capacity threshold. This represents the regional carbon emission target.

5. The power system dispatch optimization method based on deep reinforcement learning according to claim 4, characterized in that: The specific steps for issuing the execution order after the verification passes, sending the result back, and updating the scheduling policy online include: When the action sequence generated by the scheduling strategy satisfies the four-dimensional constraint space When there are boundary constraints, the verification is passed; Collect the actual execution results of the scheduling strategy, including generator output, energy storage charging and discharging status, line power flow, and carbon emissions; The actual execution results are stored in the experience replay pool, and data batches are extracted to calculate the strategy gradient. The scheduling strategy is updated using a small step size parameter η; Verify the carbon emission change rate ΔE and load shedding rate change rate ΔL of the updated scheduling strategy in the historical scenario set; If ΔE > 5% or ΔL > 3%, a safety rollback mechanism is triggered, rolling back to the previous safe scheduling strategy and triggering offline training.

6. The power system dispatch optimization method based on deep reinforcement learning according to claim 5, characterized in that: In updating the scheduling strategy using a small step size parameter η, the small step size parameter η satisfies: ； in, The initial step size, ∈[0.005,0.02], where S is the risk score. λ is an adjustment coefficient that adjusts the sensitivity of the risk score S to the decay of the small step parameter η, where λ∈[0.1,0.3].

7. A power system dispatch optimization system based on deep reinforcement learning, characterized in that: The power system scheduling optimization method based on deep reinforcement learning as described in any one of claims 1 to 6 includes: The data acquisition module is used to receive multi-source data and transmit it synchronously according to timestamps; the multi-source data includes real-time operation data, meteorological data, equipment health indicators and renewable energy output data; The graph modeling module is used to combine the multi-source data with the node and line information of the power grid to construct a power grid graph, generate a node time series matrix as the node attribute of the power grid, and encode the power grid graph and the node attribute into a multi-temporal embedding vector. The prediction and compensation module is used to generate short-term output prediction values ​​and uncertainty indicators based on the meteorological data and the renewable energy output data, and convert them into compensation factors; The risk assessment module is used to calculate a risk score based on the equipment health indicators and the meteorological data and map it to a dynamic weight. The scheduling decision module is used to attenuate the state values ​​of renewable energy nodes in the multi-temporal embedding vector using the compensation factor to obtain a corrected multi-temporal embedding vector; the corrected multi-temporal embedding vector and the dynamic weights are input into a deep reinforcement learning model, wherein the dynamic weights are used to adjust the weight ratio of economic, reliability and environmental protection objectives in the reward function of the deep reinforcement learning model to generate a scheduling strategy and perform feasibility verification. The execution and update module is used to correct the action sequence output by the scheduling strategy using the Euclidean projection algorithm when the verification fails; and to send out the execution and send back the result and update the scheduling strategy online when the verification passes. The specific steps for correcting the action sequence output by the scheduling strategy using the Euclidean projection algorithm are as follows: Real-time temperature data is acquired, and the health status of the energy storage device is determined through the device health indicators. Based on the health status of the energy storage device and the real-time temperature data, the rate of change of power adjustment on the lifespan loss of the energy storage device is calculated. The action correction direction of the action sequence is determined based on minimizing lifetime loss, and the action correction direction is the negative gradient direction in which the rate of change decreases the fastest. The original action sequence output by the scheduling strategy is adjusted along the negative gradient direction to satisfy the boundary conditions of the feasible region.

Images