Unit generation plan determination method and device, computer device and storage medium

Abstract

The application relates to a unit power generation plan determination method and device, computer equipment and a storage medium. The method comprises the following steps: acquiring multi-source power operation data and a power system topology graph, constructing a graph structured data set, constructing a plurality of independent information flow channels based on the graph structured data set, performing space-time joint coding on each information flow channel to obtain node level feature representation of each information flow channel, performing weighted fusion on each node level feature representation through a preset gating fusion mechanism to obtain a node fusion embedding vector, coupling the node fusion embedding vector and a preset historical output baseline through a preset double-gating residual operator to generate a regional energy power generation power scheme, performing parallel level reasoning on the regional energy power generation power scheme to obtain a unit power generation plan, and performing feasibility checking on the unit power generation plan to obtain a target unit power generation plan to be executed. The method can obtain a more accurate and reliable unit power generation plan, which is beneficial to power dispatching.

Description

Technical Field

[0001] This application relates to the field of power grid dispatching technology, and in particular to a method, apparatus, computer equipment, computer-readable storage medium, and computer program product for determining unit power generation plans. Background Technology

[0002] In the acquisition and standardization of multi-source time-series data, existing power dispatching systems typically rely on independent ETL processes to extract, correct, interpolate, and label information such as raw power output, meteorology, and market data in steps. In the generation of power allocation and dispatching schemes from the regional level to the unit level, preliminary allocation is often made based on historical proportions or contract rigidity, and then residuals are corrected through regularized piecewise functions or multiple rounds of iterative convergence.

[0003] However, in the above scheme, the subsequent model input needs to be transformed multiple times, which increases the complexity of data processing. The power allocation and scheduling scheme is also prone to the disconnect between residual mapping and weighting coefficients. The error will continue to be amplified in the process of multi-level aggregation and splitting. It can be seen that the current scheduling scheme is difficult to meet the requirements of efficient and accurate scheduling. Summary of the Invention

[0004] Therefore, it is necessary to provide an efficient and accurate method, apparatus, computer equipment, computer-readable storage medium, and computer program product for determining the power generation plan of generating units, in response to the above-mentioned technical problems.

[0005] In a first aspect, this application provides a method for determining a generating unit's power generation plan, including:

[0006] Acquire multi-source power operation data and power system topology maps, and construct a graph-structured dataset;

[0007] Based on the graph structured dataset, multiple independent information flow channels are constructed, and spatiotemporal joint coding is performed on each information flow channel to obtain node-level feature representations of each information flow channel. Each information flow channel corresponds to a type of business constraint that affects power generation capacity.

[0008] The node-level feature representations are weighted and fused using a gated fusion mechanism to obtain a node fusion embedding vector.

[0009] The node residuals are obtained by coupling the node fusion embedding vector and the preset historical output baseline through a preset dual-gated residual operator, and a regional energy power generation scheme is generated based on the node residuals.

[0010] Parallel hierarchical reasoning is performed on the regional energy power generation scheme to obtain the unit power generation plan. The parallel hierarchical reasoning includes regional total amount judgment and power plant power generation share refinement.

[0011] The feasibility of the unit's power generation plan is verified, and the target unit's power generation plan to be executed is determined based on the verification results.

[0012] Secondly, this application also provides a device for determining a generator unit's power generation plan, comprising:

[0013] The data acquisition module is used to acquire multi-source power operation data and power system topology maps, and to construct a graph-structured dataset;

[0014] The feature extraction module is used to construct multiple independent information flow channels based on the graph structured dataset, perform spatiotemporal joint encoding on each information flow channel, and obtain node-level feature representations of each information flow channel. Each information flow channel corresponds to a type of business constraint that affects power generation capacity.

[0015] The feature fusion module is used to perform weighted fusion of the node-level feature representations through a gated fusion mechanism to obtain a node fusion embedding vector.

[0016] The scheme determination module is used to couple the node fusion embedding vector and the preset historical output baseline through a preset dual-gated residual operator to obtain the node residual, and generate a regional energy power generation scheme based on the node residual;

[0017] The parallel reasoning module is used to perform parallel hierarchical reasoning on the regional energy power generation scheme to obtain the unit power generation plan. The parallel hierarchical reasoning includes regional total amount judgment and power plant power generation share refinement.

[0018] The data verification module is used to verify the feasibility of the unit's power generation plan and determine the target unit's power generation plan to be executed based on the verification results.

[0019] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method for determining the generator unit power generation plan.

[0020] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described method for determining generator unit power generation plans.

[0021] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps in the above-described method for determining generator unit power generation plans.

[0022] The aforementioned methods, devices, computer equipment, computer-readable storage media, and computer program products for determining unit power generation plans unify the modeling of multi-source heterogeneous power operation data and power system topology by constructing a graph-structured dataset, breaking down data silos and enabling refined spatiotemporal reasoning. Then, based on the graph-structured dataset, independent information flow channels are constructed for forward embedding and parallel processing of business constraints, decoupling complex business constraints and improving the model's interpretability and adaptability. Next, a learnable gating fusion mechanism is used to perform weighted fusion of node-level feature representations, and historical output baselines are incorporated into the model using a dual-gated residual operator. This preserves scheduling experience while enhancing the response to real-time fluctuations, generating regional energy power generation schemes. Further parallel hierarchical reasoning, including regional total quantity judgment and power plant / station power generation share refinement, is applied to these schemes, refining them down to the unit power generation plan, ensuring consistency between upper and lower levels and multi-objective collaborative optimization. Finally, feasibility verification of the unit power generation plan ensures the plan's legality, compliance, and executability, achieving parallel closed-loop backtracking verification. Overall, this scheme not only improves the accuracy and efficiency of power generation planning, but also achieves a fundamental shift from static models to an adaptive intelligent dispatching system, providing strong technical support for the intelligent management of modern power systems. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is an application environment diagram of the unit power generation plan determination method in one embodiment;

[0025] Figure 2 This is a flowchart illustrating a method for determining a generator unit's power generation plan in one embodiment;

[0026] Figure 3 This is a flowchart illustrating the steps involved in constructing a graph-structured dataset in one embodiment.

[0027] Figure 4 This is a flowchart illustrating the steps for extracting node-level feature representations in one embodiment;

[0028] Figure 5 This is a flowchart illustrating the steps for obtaining a regional energy power generation scheme in one embodiment;

[0029] Figure 6This is a flowchart illustrating the steps for determining the unit's power generation plan in one embodiment;

[0030] Figure 7 This is a flowchart illustrating the method for determining the unit power generation plan in another embodiment;

[0031] Figure 8 This is a structural block diagram of a generator unit power generation plan determination device in one embodiment;

[0032] Figure 9 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0034] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0035] The unit power generation plan determination method provided in this application embodiment can be applied to, for example, Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104, or it can be located in the cloud or on another network server.

[0036] Specifically, the operator can send a generator power generation plan scheduling message to the server 104 via terminal 102. The server 104 responds to the message, extracts multi-source power system operation data from multiple data sources, obtains a power system topology map, and further constructs a graph-structured dataset based on the multi-source power operation data and the power system topology map. Based on the graph-structured dataset, multiple independent information flow channels are constructed, and spatiotemporal joint encoding is performed on each information flow channel to obtain node-level feature representations of each information flow channel. Each information flow channel corresponds to a type of business constraint that affects power generation capacity. Then, a learnable gating fusion mechanism is used to perform weighted fusion of each node-level feature representation to obtain a node fusion embedding vector. Then, a preset dual-gating residual operator is used to couple the node fusion embedding vector with a preset historical output baseline to obtain node residuals. Based on the node residuals, a regional energy power generation scheme is generated. Further, parallel hierarchical reasoning is performed on the regional energy power generation scheme to obtain the generator power generation plan. Finally, the feasibility of the generator power generation plan is verified, and the target generator power generation plan to be executed is determined based on the verification results.

[0037] Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, drones, low-altitude aircraft, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, and projection equipment. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted displays. Head-mounted displays can be virtual reality (VR) devices, augmented reality (AR) devices, and smart glasses. Server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0038] In one exemplary embodiment, such as Figure 2 As shown, a method for determining a generator unit's power generation plan is provided, which can be applied to... Figure 1 Taking server 104 as an example, the explanation includes steps 100 to 600. Wherein:

[0039] Step 100: Obtain multi-source power operation data and power system topology map, and construct a graph structured dataset.

[0040] Multi-source power system operation data includes, but is not limited to, real-time or historical data from various power generation equipment, sensors, weather forecasting systems, etc. A power system topology diagram is a graphical representation used to describe the physical connections of a power network, including a graph showing the relationships between nodes (such as substations and power plants) and edges (transmission lines) in the power network. A graph-structured dataset is a collection of data organized in the form of a graph, where nodes represent entities and edges represent relationships between entities.

[0041] In practical applications, the server can execute batch retrieval commands for six categories of topics: thermal power output, water inflow, reservoir capacity, hydropower output, wind power, photovoltaic power, grid electricity price, maintenance records, and power restriction records, extracting multi-source heterogeneous power system operation data from different data sources. The power system topology map can be pre-constructed or built based on the power system operation data. Then, graph database technology is used to integrate these heterogeneous data into a unified framework, forming a graph-structured dataset that reflects the spatial relationships and time-series characteristics between the data.

[0042] Step 200: Based on the graph structured dataset, construct multiple independent information flow channels, perform spatiotemporal joint coding on each information flow channel to obtain the node-level feature representation of each information flow channel, and each information flow channel corresponds to a type of business constraint that affects power generation capacity.

[0043] Information flow channels refer to independent data flow paths constructed for different types of business constraints. Each information flow channel focuses on expressing a specific type of factor affecting power generation capacity, achieving alignment between business semantics and model structure. Spatiotemporal joint coding is a method that combines temporal and spatial dimensional information to capture the changing patterns of data in both time and space dimensions. Node-level feature representation is a high-dimensional vector representation extracted from each power equipment node (such as generator sets, power plants, and substations) in graph-structured data, reflecting its dynamic operating state and static attributes.

[0044] In practice, the server can define multiple independent information flow channels (hereinafter referred to as channels) on a graph-structured dataset according to business needs. Each channel is responsible for processing specific types of business constraint data. The server uses a preset encoding algorithm to perform spatiotemporal joint encoding on the data in each information flow channel to capture the changing patterns of nodes in time and space, and finally generate feature representations for each node, thereby obtaining node-level feature representations for each channel. For example, the feature representation of a wind farm node may encode information such as its historical output curve, real-time wind speed, operating status of surrounding units, and load demand in the area. Different information flow channels (such as meteorological flow and maintenance flow) will generate their own independent node-level feature representations, reflecting the node behavior patterns under specific business constraints.

[0045] For example, spatiotemporal encoding can be performed using deep learning models such as GCN, or it can be performed by combining CNN and RNN / LSTM networks to capture spatiotemporal features. Alternatively, information flow patterns can be dynamically identified, and the encoding process can be simplified using the Transformer architecture. Other encoding methods can also be used, as long as they can extract the changing patterns in the spatiotemporal dimension.

[0046] Step 300: The node-level feature representations are weighted and fused using a learnable gating fusion mechanism to obtain the node fusion embedding vector.

[0047] Learnable gated fusion is a learning method that allows the model to automatically adjust weight coefficients to optimize feature fusion performance. Node fusion embedding vectors refer to unified vector representations formed by integrating node-level feature representations from multiple information flow channels through a learnable weighted fusion mechanism.

[0048] In practice, the server uses a pre-trained gating fusion model to perform weighted fusion of the feature representations of each node, generating a node fusion embedding vector. This process involves complex mathematical operations and parameter updates to achieve the best fusion effect. For example, a GRU can be used as the gating unit to adjust the weights between different features and dynamically generate the node embedding vector through weighted fusion. Alternatively, an attention mechanism can be used to assign weights to each feature, emphasizing key information, and reinforcement learning methods can be combined to optimize the gating parameters and improve the fusion effect. Specifically, the fusion process is controlled by a "learnable gating mechanism," with gating weights dynamically adjusted according to real-time business status (such as power rationing in a certain area or unit maintenance), achieving "prioritization of important constraints." For example, when a thermal power unit is under maintenance, the feature weight of "installed availability flow" is automatically increased by the gating mechanism, while the impact of "market electricity price flow" is correspondingly reduced.

[0049] Step 400: The node residuals are obtained by coupling the node fusion embedding vector and the preset historical output baseline through the preset dual-gated residual operator, and a regional energy power generation scheme is generated based on the node residuals.

[0050] The dual-gated residual operator is a technique that combines two gating mechanisms to handle differences in input signals. The historical output baseline refers to the average output level of the power system over a past period. Node residuals, calculated using the dual-gated residual operator, are vectors or sequences reflecting the deviation between the current overall state of a node and its historical operating baseline, used to guide the direction and magnitude of adjustments to power generation plans. The regional energy power generation scheme is a draft power generation plan generated based on the aggregation of node residuals, divided by region and energy type (e.g., thermal power, hydropower, wind power, photovoltaic), serving as an intermediate output for subsequent refinement to the unit-level plan. Here, "region" can refer to a geographical or dispatching division within the power grid (e.g., province, city, control area).

[0051] Following the previous step, the server uses a dual-gated residual operator to calculate the difference between the node fusion embedding vector and the historical output baseline, i.e., the node residual. Based on these node residuals, the server can predict future power generation through a regression model and formulate a preliminary plan for regional energy power generation. The plan generation process may incorporate preprocessing strategies such as peak-valley smoothing and volatility control to obtain the final regional energy power generation plan. Alternatively, time series analysis (such as ARIMA) combined with residuals can be used for more accurate power generation prediction, thereby formulating a preliminary plan for regional energy power generation.

[0052] Step 500: Perform parallel hierarchical reasoning on the regional energy power generation scheme to obtain the unit power generation plan.

[0053] Parallel hierarchical reasoning models are models capable of handling decision problems at multiple levels simultaneously. A generator unit power generation plan is a power output arrangement for each specific generator unit (such as #1 coal-fired unit and #3 wind turbine) over a future period (e.g., the next 24 hours), including target output values, start / stop status, and ramp rate for each time period.

[0054] Specifically, the server can apply a parallel hierarchical reasoning model to reverse-engineer the unit power generation plan from the regional energy power generation plan, and refine the allocation by combining factors such as unit priority, historical output mode, and contracted power volume, thus refining the regional energy power generation plan to the scheduling plan at the level of individual generator units, and realizing the implementation from macro planning to micro execution.

[0055] Step 600: Perform a feasibility verification on the unit power generation plan, and determine the target unit power generation plan to be executed based on the verification results.

[0056] Feasibility verification refers to the process of verifying whether a power generation plan meets the actual operating conditions.

[0057] Following the previous step, after obtaining the unit power generation plan, the server performs a series of rule checks to verify whether the power generation plan for each unit meets preset constraints, such as safety standards, regulatory requirements, and technical limitations. Feasibility verification can also be performed using a preset rule engine. Only when all conditions are met will the plan be approved as the final execution version; otherwise, modification suggestions will be returned or the plan content will be directly adjusted. For example, the server can use simulation software to simulate the actual operating environment, test the feasibility and efficiency of the power generation plan, and then, based on the feedback from the simulation results, use a genetic algorithm to determine the optimal target unit power generation plan.

[0058] Furthermore, the server can also execute the target unit's power generation plan and schedule the power system.

[0059] The aforementioned method for determining unit power generation plans unifies the modeling of multi-source heterogeneous power operation data and power system topology by constructing a graph-structured dataset, breaking down data silos and enabling refined spatiotemporal reasoning. Then, based on the graph-structured dataset, it decouples complex business constraints by constructing independent information flow channels for forward embedding and parallel processing, improving the model's interpretability and adaptability. Next, it utilizes a learnable gating fusion mechanism to perform weighted fusion of node-level feature representations and incorporates historical output baselines into consideration using a dual-gated residual operator. This enhances the responsiveness to real-time fluctuations while preserving scheduling experience, generating regional energy power generation schemes. Further parallel hierarchical reasoning, including regional total quantity judgment and power plant / station power generation share refinement, is applied to these schemes, refining them down to the unit power generation plan and ensuring consistency across all levels and multi-objective collaborative optimization. Finally, the feasibility of the unit power generation plan is verified, ensuring its legality, compliance, and executability, achieving parallel closed-loop backtracking verification. Overall, this scheme not only improves the accuracy and efficiency of power generation planning, but also achieves a fundamental shift from static models to an adaptive intelligent dispatching system, providing strong technical support for the intelligent management of modern power systems.

[0060] In one exemplary embodiment, such as Figure 3 As shown, step 100 includes:

[0061] Step 110: Obtain multi-source power system operation data, map the power equipment identifiers from different sources in the multi-source power system operation data to global node identifiers based on preset hierarchical coding rules, and construct a power system topology map in combination with preset power grid connection relationships.

[0062] Step 120: Perform time alignment and format standardization processing on the multi-source power system operation data to obtain a time-uniform time series dataset.

[0063] Step 130: Differentiate the missing values ​​in the time series dataset according to the preset data quality indicators, and simultaneously generate the data confidence vector for each data point.

[0064] Step 140: Bind the static attribute data in the multi-source power system operation data to the corresponding unit node.

[0065] Step 150: The time series data, data confidence vector, static attribute data and power system topology map are fused and written into the graph database to generate a graph structure data snapshot with version identifier and effective time interval. The graph structured dataset includes the graph structure data snapshot.

[0066] In practice, the server can set up a search window within a unified interface of the target system. Batch retrieval commands were executed on six categories of topics: thermal power output, water inflow, reservoir capacity, hydropower output, wind power, photovoltaic power, municipal electricity price, maintenance records, and power restriction records. The raw multi-source power system operation sequence data was retrieved, and then the retrieved multi-source power system operation sequence data was organized into a two-dimensional array. ,in Indicates a topic index. Indicates the source timestamp.

[0067] Next, for the two-dimensional array Search line by line for the longitude of the unit And query the time zone offset Refer to the historical daylight saving time switching table Double-check the timestamps to obtain the sequence. .

[0068] Intervals in natural hours Resampling to a unified index set Use length for segments that span multiple days or are out of order. The symmetrical smooth window is rearranged as follows:

[0069]

[0070] in, For the rearranged matrix, the first... Line number Column elements; For the corrected matrix In terms of thematic dimension Time Index The value at; To smooth out the window's half-width, here we set... This ensures that the timeline has no overlap and no gaps.

[0071] Next, a four-level mapping dictionary can be constructed based on preset coding rules, consisting of a region, a branch office, a power plant / site, and a generating unit. A fixed four-segment encoding is adopted. Replace all original names. The same unit in different systems is uniquely mapped to a node identifier. This eliminates heterogeneous naming differences and achieves globally consistent indexes.

[0072] Subsequently, regarding the rearranged Introduce a business threshold table Calculate confidence level Continuous gap length Perform linear interpolation at the specified time; The gap should be preserved and marked.

[0073]

[0074] in, For time index Confidence level at the location; This is a threshold test indicator; when the threshold is met, it is taken. Otherwise take ; For index Length of continuous gaps; A constant weight is used to balance the threshold test and the gap length; .

[0075] Then convert the confidence vector The gap length is written synchronously to the data quality field to provide an uncertainty reference for downstream models.

[0076] Subsequently, the new energy access ledger was accessed. The branch-level wind and solar power is broken down to the level of the power station and the generating unit, forming a detailed matrix. Simultaneously, the altitude of the aircraft position was extracted. Blade type Wheel hub height For static attribute data, the extracted static attribute data will be bound to the corresponding node. Expanding the dimensions of meteorological driving features.

[0077] Finally, based on the four-level coding Mapping results and processed sequences Confidence level and static properties Write to a unified graph structure Each node and edge is appended with a version number. Effective period This enables data and topology synchronization and traceability. Furthermore, it allows for the output of graph snapshots. Write to object repository Simultaneously, it generates a metadata list and a quality report. .

[0078] In this embodiment, the snapshot is directly used for encoding in the three-flow graph neural network, achieving seamless connection between the data layer and the model layer.

[0079] In this embodiment, a unified and high-quality data foundation is laid for subsequent data processing through one-time extraction of multi-source time series data, unified time base correction, four-level topology mapping, confidence quantification, and graph structure construction.

[0080] In an exemplary embodiment, the information flow channels include an installed availability flow, a weather-driven flow, and a market demand flow; spatiotemporal joint coding is performed on each information flow channel to obtain the node-level feature representation of each information flow channel, including:

[0081] Step 220: The installed availability stream, weather-driven stream, and market demand stream are feature-encoded by a combination of temporal convolution and graph convolution operators to obtain node-level feature representations of the installed availability stream, weather-driven stream, and market demand stream.

[0082] The installed capacity availability flow is a state information flow reflecting whether generator units are available for use, primarily driven by physical and operational constraints such as unit maintenance plans, fault outages, and minimum start-up and shutdown times. The meteorological drive flow is an information flow carrying the impact of environmental variables on renewable energy output, including meteorological forecast data such as wind speed, irradiance, and water inflow, directly affecting the output capacity of wind power, photovoltaic power, and hydropower. The market demand flow is an information flow characterizing the supply and demand relationship and economic incentive direction in the electricity market, containing signals such as market clearing prices, contracted power volumes, load forecasts, and peak-shaving compensation, guiding units to respond according to the principle of prioritizing economic efficiency. Temporal convolution is a neural network operation suitable for processing time-series data. By sliding convolution kernels, it extracts local time patterns (such as output fluctuation cycles and meteorological change trends), excelling at capturing short-term dynamic features. Graph convolution is an operation that aggregates information on a graph structure, achieving spatial correlation modeling by aggregating the features of a node itself and its neighboring nodes.

[0083] In practical implementation, the core contradiction in power generation dispatching scenarios—the heterogeneous coupling of installed capacity availability, weather volatility, and market demand—can be addressed by designing three information flow channels. Specifically, the server in Copy all nodes and edges, and attach the dynamic features of the installation to them respectively. Meteorological dynamic characteristics and market dynamics This forms three independent channels: the installed capacity availability flow, the weather-driven flow, and the market demand flow. The replication process maintains node ID consistency with the topology, ensuring that the propagation results of each flow are aligned in the same coordinate system. Each channel is bound to a corresponding data source field (such as a maintenance schedule, weather forecast interface, or market clearing results). It is understood that in other embodiments, in addition to the three types of business constraints mentioned above, corresponding information flow channels can be constructed for other business constraints, depending on actual business needs and circumstances, and are not limited here.

[0084] Next, the server can extract short-term dynamic features from each information flow channel—installation availability flow, weather-driven flow, and market demand flow—using one-dimensional temporal convolution operations. Then, it simultaneously uses graph convolution to aggregate these features, achieving spatial correlation modeling. Finally, it performs spatiotemporal fusion processing on the outputs of temporal and graph convolutions, outputting a node-level feature representation matrix with consistent dimensions for each information flow channel. Each row corresponds to a high-dimensional embedding vector of a unit node in the current time slice. It is understandable that the above processing can flexibly adopt different structures (such as temporal first then graph, graph first then temporal, parallel dual-branch, alternating execution, etc.), and attention mechanisms can be introduced to enhance the weights of key nodes, but the core logic remains "independent encoding for each flow, preserving semantic isolation."

[0085] In this embodiment, by constructing three semantically clear information flow channels and using a combination of temporal convolution and graph convolution operators for independent encoding, a refined and structured model of the factors influencing power generation capacity is achieved. During processing, the server transforms the raw data into a high-dimensional semantic representation, which not only improves the model's expressive power and interpretability but also lays a high-quality feature foundation for subsequent gating fusion and residual generation.

[0086] There are no restrictions on the spatiotemporal joint coding method. For example... Figure 4 As shown, in an exemplary embodiment, step 220 includes:

[0087] Step 222: For the installed capacity availability flow, mark the shutdown or de-processing status according to the unit maintenance plan, and recursively determine the dispatchable capacity margin along the progressive direction of the unit, power plant and region to obtain the time-varying capacity matrix.

[0088] Step 224: For the meteorological driving flow, load meteorological forecast data to supplement the wind speed, irradiance and inflow prediction sequences, and propagate the meteorological fluctuation signal through the geographical adjacency edge.

[0089] Step 226: For market demand flow, convert market electricity price and contracted electricity volume into demand pull tensor, and propagate it in reverse from the load center to the generation-side node along the transmission network.

[0090] Step 228: For the installed capacity availability flow, weather-driven flow, and market demand flow, short-term dynamic features are extracted based on temporal convolution, and spatial coupling effects are captured based on graph convolution. The extracted dynamic features and spatial coupling effects are spatiotemporally fused to obtain node-level feature representations of the installed capacity availability flow, weather-driven flow, and market demand flow.

[0091] In this embodiment, the server may target the installed available stream. Read the maintenance plan and at the nodes The status of shutdown or reduced output is marked; this status is recursively calculated along the direction of unit → power plant → region, and the dispatchable capacity margin of the upper-level node is deducted in real time to form a time-varying capacity matrix. Based on the time-varying capacity matrix, the network can explicitly identify the spatial transmission effect of capacity shrinkage during propagation.

[0092] For meteorological driving flow The system loads hourly forecast data to complete the wind speed, irradiance, and water inflow sequences, and propagates them to adjacent nodes through geographical adjacency edges. Once a threshold change rate is detected, the fluctuation flag is immediately written to the affected node, thus exposing the synchronous risk of large-scale weather changes at a spatial scale in advance.

[0093] Targeting market demand flow Mapping electricity price series and contracted electricity volume to a demand-pull tensor The positive gradient propagates in reverse from the load center to the generation-side nodes along the transmission topology; high-price nodes generate a larger positive gradient, thereby strengthening the attraction of this region to the generating capacity in subsequent convolution calculations.

[0094] Next, the temporal convolution → graph convolution combination operator is executed sequentially on the above three information flow channels, as follows:

[0095] Stream tag For indexing, a one-dimensional temporal convolution is performed on the node-level feature representation to extract short-term dynamics. Subsequently, at the same time step, graph convolution is performed along the power system topology to capture spatial coupling.

[0096]

[0097] in, For nodes At any moment flow Fusion characterization; It is a non-linear activation function; For nodes The set of neighbors; For the edge Normalized weights; For nodes At any moment The original flow characteristics; This represents the length of the temporal convolution window; These are the parameters of the temporal convolution kernel; The graph is a convolution weight matrix.

[0098] Next, the temporal convolution and spatial convolution are embedded in the same summation symbol to complete the spatiotemporal feature fusion in one go, avoiding repeated storage and redundant propagation.

[0099] Construct a cross-flow gating layer to block maintenance windows. Scheduling priority Emission reduction weight The three types of constraints are mapped to learnable gating vectors, and the weighted fusion of the three information flow channels (hereinafter referred to as the three flows) is completed by the following formula:

[0100]

[0101] in, For nodes At any moment fused embedding vector; For flow The gating coefficient; Element-wise multiplication; This is a gating weight hyperparameter.

[0102] Gating injects business rules into the network in the form of tensors, enabling the algorithm to learn the dynamic adjustment of the rules to the three-stream weights during the training phase, rather than hard-correcting them after inference.

[0103] Finally, embed the fusion. Write data to the graph database, create a time index and attach a version number to achieve persistence of node-level three-flow representation; at the same time, write gating weights into the parameters to provide fine-grained input for regional power inference that includes physical, meteorological and market constraints, and achieve seamless connection from data to model.

[0104] In this embodiment, based on the improved temporal convolution-graph convolution joint operator and cross-stream gating layer, business constraints are directly embedded into the feature fusion process, and finally node-level three-stream embedding vectors are output, providing interpretable and differentiable fine input for regional power inference.

[0105] like Figure 5 As shown, in an exemplary embodiment, step 400 includes:

[0106] Step 410: By coupling the node fusion embedding vector with the preset dual-gated residual operator and the preset historical output baseline, the node residual that takes into account both real-time fluctuations and historical inertia is obtained.

[0107] Step 420: Sum the absolute values ​​of the node residuals according to the time dimension to obtain the residual modulus of each unit node in the region.

[0108] Step 430: Generate adaptive weighting coefficients based on the residual modulus of each unit node in the region, and weight and aggregate the output of units of the same type based on the adaptive weighting coefficients to generate a draft of regional energy power generation capacity.

[0109] Step 440: Perform peak-valley smoothing on the draft regional energy power generation capacity to obtain the regional energy power generation capacity scheme.

[0110] In this embodiment, it can be based on the node fusion embedding vector obtained above. By using a four-stage progressive approach—node residual, adaptive weighting, regional aggregation, and peak-valley smoothing—the representation of the unit at the granularity level is transformed into a draft of power generation capacity at the granularity level of region-energy type (i.e., regional energy power generation scheme), replacing the static rule of traditional installed capacity × utilization hours, and providing dynamic and schedulable input for regional total inference.

[0111] Specifically, the server could read all node fusion embedding vectors from the graph database. (Hereinafter referred to as node embedding), the node set is based on the unit tag. Divided into thermal power units Hydropower collection New energy collection Next, the actual power output curves for the same month over the past three years (historical power output baseline) are retrieved, and each node is analyzed. Calculate the daily arithmetic mean to obtain the monthly static baseline. ,in This represents the number of calendar days in the current month. The historical output baseline is considered as the inertia level formed by historical utilization hours and is used to measure the deviation of the current node's embedding from historical operation.

[0112] In this embodiment, a hierarchical dual-gated residual combination operator is proposed. Addressing the technical pain point of excessive historical inertia and insufficient response to real-time fluctuations due to long-term internal planning-performance imbalances within the same month, the operator embeds nodes... Compared with static baseline Coupled with a dual-gating mechanism, the node residuals are output all at once. The operator incorporates both the immediate reliability of dynamic features and the steady-state weights of the baseline into the mapping process, ensuring that the residuals retain historical interpretability while explicitly quantifying real-time increments.

[0113]

[0114] in, For nodes The three-stream fusion embedding matrix, dimension ; Energy type The mapping weight matrix, dimension ; For the corresponding bias column vector, dimension ; This is the Sigmoid activation function, used to generate a confidence gate in the 0-1 interval; Element-wise multiplication; For nodes Monthly static baseline, dimensions ; The baseline amplification factor scalar is estimated offline from the internal system target using hourly thresholds. The hyperbolic tangent function is used to... Perform smooth compression; The output node residual curves, dimension .

[0115] The parentheses on the left side of the above equation For dynamic channels, weights are used. Map the embedding to the output space; For trust-based gating, utilize After extracting features related to data quality, they are converted to 0-1 weights using a Sigmoid function, and adaptive compression is then applied to the dynamic channels; (The text in parentheses on the right is missing.) For baseline calibration, multiply the historical baseline by This allows for flexible scaling up or down. After combining these three parts with dynamic gating and baseline, a residual is obtained that balances real-time fluctuations and historical inertia. .

[0116] Next, regarding the above-obtained... The residual modulus is obtained by taking the absolute value and summing over the time dimension. A larger residual within a region indicates that the nodes rely more on real-time information than on historical inertia. Standardization to adaptive weighting coefficients ,in, Indicates the region Internal energy type A set of nodes. For the same region Internal energy type The nodes, according to the coefficient Mapping output force By performing a weighted summation, the regional energy power generation scheme is obtained. :

[0117]

[0118] in, For the region Internal energy type The initial power generation scheme curve; For nodes At any moment The absolute value of the residual; For nodes The output prediction curve.

[0119] Then, the weight calculation and power aggregation are embedded within a single summation symbol, reducing two-stage error propagation and ensuring that weights and power are updated synchronously. The resulting region matrix is ​​used to perform a peak-valley smoothing algorithm. If the new energy column... If the difference between adjacent days exceeds the scheduling threshold, the thermal power residual pool will be dynamically invoked to move the thermal power train... The remaining weights are fed back to the new energy peak to ensure that the initial power generation scheme curve meets the start-up and shutdown flexibility and ramp-up constraints. Next, the smoothed regional energy power generation scheme, node residuals, and adaptive coefficients are written into the draft library, assigned a version number and generation timestamp, to provide a versioned input that can be directly referenced for subsequent total inference.

[0120] In this embodiment, a four-stage progressive approach—node-level residual, adaptive weighting, regional aggregation, and peak-valley smoothing—transforms the unit-level representation into a region-energy type-level power generation draft, replacing the traditional static rule of installed capacity × utilization hours, and providing a dynamic and schedulable input for regional total inference.

[0121] like Figure 6 As shown, in an exemplary embodiment, step 500 includes:

[0122] Step 520: A parallel branch structure is used to perform regional total power generation inference on the regional energy power generation scheme. The first branch structure extracts the slow variation trend of the regional energy power generation scheme, and the second branch structure captures the high-frequency oscillation of the regional energy power generation scheme. The outputs of the first and second branch structures are dynamically fused based on the confidence level of weather forecasts and the penetration rate of new energy sources to obtain regional supply forecast data.

[0123] Step 540: Based on the regional residual between regional supply forecast data and regional energy power generation scheme, an initial unit power generation plan is generated by combining historical operating modes and contractual constraints. The physical feasibility and demand pull of the initial unit power generation plan are coupled through a cross-attention mechanism. After multiple rounds of residual convergence iteration, the converged unit power generation plan is output.

[0124] Parallel branching structure refers to a model architecture that sets up two or more independent sub-network paths to process different feature dimensions of the same input data (such as low-frequency trends and high-frequency fluctuations), and finally fuses the outputs of each branch. In this embodiment, the parallel branching structure is used to decompose and model the time series signal of "regional energy power generation scheme": the first branch structure focuses on extracting slow-changing trends, such as long-term evolution features like overall daytime power output level, load base changes, and seasonal power output patterns. The second branch structure focuses on capturing high-frequency oscillations, such as minute-level fluctuations in new energy power output, sudden weather disturbances, and short-term load jumps.

[0125] Slow-change trend refers to the steady change trend of regional power generation over a longer time scale (such as hourly or daily). High-frequency oscillation refers to the sharp fluctuations in regional power generation over a short time scale (such as minute or 15-minute time scale), mainly caused by random factors such as sudden changes in wind speed, cloud cover, and instantaneous load fluctuations. Weather forecast confidence level indicates the reliability of current weather forecast results. New energy penetration rate refers to the proportion of fluctuating renewable energy output such as wind power and photovoltaic power in the total load or total installed capacity within a region. Regional supply forecast data refers to the prediction results of the region's total power generation capacity over a future period, generated after parallel branch inference and dynamic fusion. It is in the form of a time series curve and is used to guide the total constraints and resource allocation of subsequent unit-level plans.

[0126] Regional residuals refer to the discrepancy between regional supply forecast data and regional energy generation capacity plans, reflecting the inconsistency between current forecasts and preliminary plans. Historical operating patterns refer to the output behavior of generating units under similar past operating conditions. Contractual constraints refer to the commercial or dispatch agreement restrictions that generating units must comply with.

[0127] After completing the regional energy power draft, the focus shifted to the core challenge of monthly supply and demand matching within the group's regions. This involved collaborative modeling of high-frequency fluctuations in renewable energy and steady-state assurance for thermal power. Based on a dual-branch, single-gated structure, slow-varying trends and high-frequency oscillations were extracted in parallel. Dynamic fusion was then implemented based on weather forecast confidence and renewable energy penetration rates, outputting a regional-level rolling 35-day supply and demand curve and gap range, providing an upper-level boundary for detailed reasoning at the power plant / station level.

[0128] In practice, parallel hierarchical reasoning includes two layers. The first layer is the dynamic judgment of the total amount in the region:

[0129] It could be a server based on a region-based energy power draft. By energy type In the region Perform summation within the region to obtain the regional supply benchmark sequence. For the region At any moment The supply benchmark; The area output by the aforementioned steps Energy type The power draft. The baseline sequence serves as the sole supply starting point for subsequent window inference.

[0130] Next, retrieve the load data for the same period in the past three years. Holiday markings Latest demand-side response booking volume According to the time dimension and The input tensors are concatenated to form a multi-feature input tensor. The tensor maintains feature alignment during subsequent window partitioning, ensuring that supply and demand characteristics can be learned simultaneously.

[0131] Furthermore, Divided into lengths by natural day sliding method window Each window is first fed into the Long Short-Term Memory branch, where a slowly changing trend is extracted through a gated loop, and the output is a region-specific slowly changing vector. This branch focuses on capturing low-frequency characteristics such as monthly periodicity and load seasonality. Simultaneously, it applies the same window in parallel. The vectors are fed into a multi-head attention transformer branch, where global dependencies are established among features after locking the temporal order based on position encoding, and the output region high-frequency vectors are generated. The branch enhances its ability to capture random oscillations and load spikes in new energy sources.

[0132] Subsequently, a single-gated fusion layer was constructed, and weather forecast confidence was introduced. With the penetration rate of new energy The gating coefficients are generated by the Sigmoid transform. By dynamically weighting the outputs of the two branches, regional supply forecasts are obtained. for:

[0133]

[0134] in, For the region time The supply forecast curve; , For the time-by-time output of the Transformer and LSTM branches; To integrate the gating coefficient, the value is selected as follows: ; The confidence level of the weather forecast is obtained by weighting the inverse exponential sum of the forecast-measurement error over the past three days. This refers to the penetration rate of new energy sources, which is the proportion of new energy installed capacity in a region. The gating weights are to be trained.

[0135] In the formula, the weight of the first term increases with the increase of meteorological confidence and new energy penetration rate, making high-penetration-high-confidence areas more dependent on high-frequency branches and low-penetration-low-confidence areas more dependent on slow-changing branches, thus meeting the heterogeneous scheduling needs of multiple regions.

[0136] Finally, With demand vector Differences in supply and demand are used to identify supply and demand gaps. And calculate the 95% confidence interval. The gap curve along with the gating weights Stored in the regional inference library, it provides upper-level boundaries and uncertainty scales for detailed inference of power plant / station power generation share.

[0137] The second layer of parallel hierarchical reasoning is the refinement of power plant / station power generation share:

[0138] In this embodiment, a closed-loop drill-down approach is used to calculate the daily executable power of the unit based on the regional residual, focusing on whether the regional power allocation is fair and whether the unit curve falls within the physically feasible region.

[0139] First, receive the regional monthly demand residuals generated from the above steps. Based on the power generation ratio of the same period last year With the monthly long-term contract quota Calculate the initial share of the power plant :

[0140]

[0141] in, For power plants In the month The initial charge; For the region Internal power plant cluster; For power plants The percentage compared to the same period last year; For power plants Current long-term contract volume; This represents the residual electrical quantity in the region.

[0142] Next, the calculation results and the power plant ID are packaged into a message and pushed to the message queue to ensure that the starting point of power allocation is synchronized with historical and contractual boundaries. When the message arrives, the regional residuals need to be... The allocation was precisely broken down to individual power plants. To balance historical contributions with contractual obligations, a trade-off factor was introduced. By decoupling and weighting the proportion from the same period last year with the current month's long-term contract quota, a more intuitive formula for calculating the initial share is obtained:

[0143]

[0144] in, For power plants In the month The initial share of electricity; For the region In the month The supply and demand residual electricity volume; The value range is a weighting factor. , The larger the value, the more it leans towards historical proportions; For power plants The actual power generation percentage in the same month last year; For power plants Long-term contract electricity volume for the current month; For the region A set of indices for all power plants in the system.

[0145] Next, normalize the historical percentage and long-term contract quota respectively, and then... Perform a convex combination, and finally multiply by the regional residual. Generate initial shares at the power plant level; thereby providing a clear quantitative basis for historical inertia and contractual rigidity, and ensuring that the weights of the two can be flexibly configured according to management regulations.

[0146] This embodiment proposes a dual-cross attention combination function, with one branch outputting a physical availability vector. The output demand vector of the two branches Inter-tower coupling, gated compression, and power normalization are achieved in one step through cross-gated combination functions:

[0147]

[0148] in, For the reasoning window The unit power vector; Activated by Sigmoid, outputting a 0–1 range gate; Energy type The combined weight matrix, size ; Vector concatenation operator; Element-wise multiplication; , The physical tower and the demand tower are respectively located in Japan. The hidden vector, dimension ; A single vector, dimension .

[0149] Next, second-order interactions are captured in the latent space through concatenation and element-wise multiplication. Then, the same gating system simultaneously compresses both physical quantities and demand quantities, ensuring that the allocation result reflects both the real-time capacity of the generating units and adheres to regional demand intensity. When the node label is for new energy, a meteorological fluctuation coefficient is loaded. Bundle Projection to normal distribution After taking the median path As a suggestion, efforts are made and confidence intervals are recorded for subsequent risk mitigation. When the node label is thermal power, verification is performed. Does it trigger the minimum boot time period? With climbing rate If the limit is exceeded, the start / stop corrector will... The section is linearly pulled back to push the out-of-limit values ​​back into the feasible region, and the physical compliance curve is output. .

[0150] On the other hand, the server sums the curves of all power plants to obtain the regional actual delivery. ; and demand sequence Difference to generate the first wheel residual The proposed segmented update function-driven multi-round refinement is as follows:

[0151]

[0152] in, For the first Wheel area residuals; Adjustment step size for the entire unit; For power plants No. Total daily curve volume; For power plants The adjustable upper limit can be further increased; A dedicated adjustment step size for thermal power plants; It is a collection of thermal power plants within the region; For power plants Maximum climbing ability; The threshold values ​​are for two levels of residuals.

[0153] When the absolute value of the residual is greater than The entire unit will coordinate and scale; falling into Special thermal power regulation is selected during the interval; below When convergence is achieved, the value is set to zero to avoid unnecessary iterations, and the converged unit power generation plan is output.

[0154] when At that time, lock the power plant version number. The final daily curve of the unit With confidence level Archive the data for subsequent physical backtracking and scheduling interactions, completing the subdivision from region to unit.

[0155] In this embodiment, high-frequency fluctuations of new energy sources and steady-state protection of thermal power are modeled collaboratively. Based on a dual-branch-single-gated structure, slow-varying trends and high-frequency oscillations are extracted in parallel. Then, dynamic fusion is performed based on the confidence level of weather forecasts and the penetration rate of new energy sources to output a regional rolling 35-day supply and demand curve and gap range. This provides an upper-level boundary for refined reasoning at the power plant / station level, thereby enabling more precise unit power generation plans.

[0156] like Figure 7 As shown, in an exemplary embodiment, step 600 includes:

[0157] Step 620: Based on the equipment physical parameters and preset external compliance constraints, perform a dual verification of the unit's power generation plan using both equipment physical limits and compliance operation constraints to obtain the verification results. Based on the verification results, determine the target unit's power generation plan to be executed.

[0158] Equipment physical parameters refer to the inherent technical performance indicators of the generator set itself, including but not limited to maximum and minimum output, gradeability (the maximum increase or decrease in output per unit time), minimum start-up and shutdown time (the shortest duration of continuous operation or shutdown), and parameters such as cold and hot start-up time and energy consumption. External compliance constraints refer to mandatory operational requirements from grid dispatch rules, environmental management regulations, market mechanisms, cross-regional agreements, etc., which are not inherent to the equipment itself; these are also known as "institutional constraints" or "normative boundaries."

[0159] In this embodiment, the dual verification of the unit's power generation plan by checking both the physical limits of the equipment and the compliance constraints of operation refers to checking whether the unit's power generation plan exceeds its physical capacity and whether the unit's power generation plan meets the external supervision and dispatch requirements. This forms a dual check mechanism for the technical feasibility and institutional compliance of the power generation plan, ensuring that the plan is both "operable" and "executable".

[0160] In practice, it can be based on the daily output scheme of the S5 unit. With physical limits of installed capacity, upper limits of environmental emissions, and inter-provincial transaction coefficients as hard constraints, and electricity prices as the basis for these constraints. Emission factors With energy efficiency coefficient As a soft criterion, a closed-loop verification chain is formed, consisting of detection, adjustment, conversion, and reverse writing, to ensure that the monthly plan is both physically feasible and meets the standard priority.

[0161] Specifically, the server first aggregates all unit curves output from the above steps, and then sorts them according to natural days. Stacked into a three-dimensional regional power matrix ,in For regional indexes, The hour is used as the identifier. The matrix serves as the unique view for all subsequent verification processes, including version stamps. With generation time All of these are solidified together to ensure traceability.

[0162] Next, regarding Daily traversal is performed if the element value exceeds the machine's assembly limit. or below the minimum technical output Immediately mark with a red label At the same time, the regional emission red line Hourly cumulative emissions Time slices exceeding the threshold also fall into The red markers, linked together, form a violation vector, driving subsequent adjustment processes. For existing... Japanese films, calculating peak-to-valley differences .like If there is a shortage of energy during a valley period, then check the energy storage list. The upper limit reserves charging windows for pumped storage units; if overload occurs during peak periods, the ramping capacity list will be updated accordingly. Select scalable thermal power plants for pre-emptive peak shaving. After adjustments at both ends, write back again. If the peak-valley transition is not closed, the residual redistribution of the preferential pricing, priority scheduling, and energy efficiency will be initiated based on the electricity price tensor. Arrange in descending order, prioritizing allocation of higher-priced time periods; if the emission factor of higher-priced units... Exceeding the average threshold If the emissions are low, the process will proceed to the next unit, prioritizing lower emissions; if there are still unallocated residuals, they will be allocated according to the energy efficiency coefficient. Replenish from high to low until the residual. The convergence is 0. The entire process forms a monotonically decreasing sequence, ensuring that the redistribution will not disrupt the resolved peaks and valleys.

[0163] Next, the latest inter-provincial transaction conversion factor is applied to the adjusted curve. Converted to electricity consumption for accounting purposes, and based on contract boundaries. With regional emission limits Double verification: If any metric still exceeds the limit, the most recent allocation record will be rolled back. And trigger the rollback retry flag until the matrix Full compliance. Then extract the net increase / decrease at each node. Mapped to the new gradient of the third-flow gating weights , , These correspond to capacity, peak shaving, and emission channels, respectively. Gradient writing parameters It serves as an explicit physical prior during model fine-tuning in the next cycle, thereby improving the prediction convergence speed.

[0164] Finally, the final executable curve will be... Encapsulated as a three-level JSON object: region-power plant-unit, with object keys containing... Along with the version number Pushed to the scheduling system And start real-time deviation monitoring: recover actual output every hour during runtime. Market transactions and actual meteorological measurements ; to deviation Store in the supervised sample table For use in daily closed-loop training.

[0165] In this embodiment, through the above-mentioned dual verification, the consistency of the reading power generation plan can be verified from the physical constraints of the unit, the regional emission targets and the cross-provincial transaction boundaries, and it is stored in the form of versioned JSON, realizing a closed loop of full-link execution.

[0166] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0167] Based on the same inventive concept, this application also provides a generator set generation plan determination device for implementing the generator set generation plan determination method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the generator set generation plan determination device provided below can be found in the limitations of the generator set generation plan determination method described above, and will not be repeated here.

[0168] In one exemplary embodiment, such as Figure 8 As shown, a generator unit power generation plan determination device 800 is provided, including: a data acquisition module 810, a feature extraction module 820, a feature fusion module 830, a scheme determination module 840, a parallel inference module 850, and a data verification module 860, wherein:

[0169] The data acquisition module 810 is used to acquire multi-source power operation data and power system topology maps to construct a graph-structured dataset.

[0170] The feature extraction module 820 is used to construct multiple independent information flow channels based on the graph structured dataset, perform spatiotemporal joint encoding on each information flow channel, and obtain the node-level feature representation of each information flow channel. Each information flow channel corresponds to a type of business constraint that affects power generation capacity.

[0171] The feature fusion module 830 is used to perform weighted fusion of the feature representations of each node through a gating fusion mechanism to obtain the node fusion embedding vector.

[0172] The scheme determination module 840 is used to obtain the node residual by coupling the node fusion embedding vector and the preset historical output baseline through a preset dual-gated residual operator, and to generate a regional energy power generation scheme based on the node residual.

[0173] The parallel reasoning module 850 is used to perform parallel hierarchical reasoning on the regional energy power generation scheme to obtain the unit power generation plan. The parallel hierarchical reasoning includes regional total amount judgment and power plant power generation share refinement.

[0174] The data verification module 860 is used to verify the feasibility of the unit's power generation plan and determine the target unit's power generation plan to be executed based on the verification results.

[0175] In an exemplary embodiment, the data acquisition module 810 is used to acquire multi-source power system operation data, map the identifiers of power equipment from different sources in the multi-source power system operation data to global node identifiers based on preset hierarchical coding rules, and construct a power system topology map in combination with preset grid connection relationships. The multi-source power system operation data is then processed for time alignment and format standardization to obtain a time-uniform time-series dataset. Missing values ​​in the time-series dataset are differentiated according to preset data quality indicators, and a data confidence vector for each data point is generated synchronously. Static attribute data in the multi-source power system operation data is bound to the corresponding unit nodes. The time-series data, data confidence vectors, static attribute data, and power system topology map are fused and written into a graph database to generate a graph structure data snapshot with version identifiers and effective time intervals. The graph structured dataset includes the graph structure data snapshot.

[0176] In an exemplary embodiment, the information flow channel includes an installation availability flow, a weather-driven flow, and a market demand flow; the feature extraction module 820 is used to perform feature encoding on the installation availability flow, the weather-driven flow, and the market demand flow respectively through a combination of temporal convolution and graph convolution operators to obtain node-level feature representations of the installation availability flow, the weather-driven flow, and the market demand flow.

[0177] In one exemplary embodiment, the feature extraction module 820 is further configured to:

[0178] For the installed capacity availability flow, the shutdown or de-processing status is marked according to the unit maintenance plan, and the dispatchable capacity margin is determined recursively along the progressive direction of the unit, power plant and region to obtain the time-varying capacity matrix;

[0179] For meteorological driven flows, meteorological forecast data is loaded to supplement the wind speed, irradiance and inflow prediction sequences, and meteorological fluctuation signals are propagated through geographical adjacency edges;

[0180] In the market demand flow, market electricity prices and contracted electricity volumes are converted into demand pull tensors and propagated backward from the load center to the generation-side nodes along the transmission network.

[0181] For the installation availability flow, weather-driven flow, and market demand flow, short-term dynamic features are extracted using temporal convolution, and spatial coupling effects are captured using graph convolution. The extracted dynamic features and spatial coupling effects are then fused spatiotemporally to obtain node-level feature representations for the installation availability flow, weather-driven flow, and market demand flow.

[0182] In an exemplary embodiment, the feature fusion module 830 is further configured to map the maintenance blockade window, scheduling priority, and emission reduction weight into learnable gating vectors, and to perform weighted fusion of the node-level feature representations by element-wise multiplication based on the learnable gating vectors to obtain a node fusion embedding vector.

[0183] In an exemplary embodiment, the scheme determination module 840 is further configured to couple the node fusion embedding vector with the preset dual-gated residual operator and the preset historical output baseline to obtain the node residual that takes into account both real-time fluctuations and historical inertia, sum the absolute values ​​of the node residuals according to the time dimension to obtain the residual modulus of each unit node in the region, generate adaptive weighting coefficients based on the residual modulus of each unit node in the region, and perform weighted aggregation of the output of the same type of unit based on the adaptive weighting coefficients to generate a regional energy power generation draft, and perform peak-valley smoothing on the regional energy power generation draft to obtain the regional energy power generation scheme.

[0184] In an exemplary embodiment, the parallel inference module 850 is further configured to perform regional total inference on the regional energy power generation scheme using a parallel branch structure. The first branch structure extracts the slow-changing trend of the regional energy power generation scheme, and the second branch structure captures the high-frequency oscillation of the regional energy power generation scheme. The outputs of the first and second branch structures are dynamically fused based on the weather forecast confidence level and the new energy penetration rate to obtain regional supply forecast data. Based on the regional residual between the regional supply forecast data and the regional energy power generation scheme, an initial unit power generation plan is generated by combining historical operating modes and contractual constraints. The physical feasibility and demand pull of the initial unit power generation plan are coupled through a cross-attention mechanism. After multiple rounds of residual convergence iteration, the converged unit power generation plan is output.

[0185] In an exemplary embodiment, the data verification module 860 is further configured to perform dual verification of the unit power generation plan based on the equipment physical parameters and preset external compliance constraints, thereby obtaining verification results and determining the target unit power generation plan to be executed based on the verification results.

[0186] The modules in the aforementioned generator set power generation plan determination device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0187] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores power system operation data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for determining a generator unit's power generation plan.

[0188] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0189] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in any of the above embodiments of the unit power generation plan determination method.

[0190] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in any of the above embodiments of the unit power generation plan determination method.

[0191] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the unit power generation plan determination method.

[0192] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to multi-source power system operation data used for analysis, stored data such as power generation schemes, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with relevant regulations.

[0193] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0194] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0195] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for determining a generator unit's power generation plan, characterized in that, The method includes: Acquire multi-source power operation data and power system topology maps, and construct a graph-structured dataset; Based on the graph structured dataset, multiple independent information flow channels are constructed, and spatiotemporal joint coding is performed on each information flow channel to obtain node-level feature representations of each information flow channel. Each information flow channel corresponds to a type of business constraint that affects power generation capacity. The node-level feature representations are weighted and fused using a gated fusion mechanism to obtain a node fusion embedding vector. The node residuals are obtained by coupling the node fusion embedding vector and the preset historical output baseline through a preset dual-gated residual operator, and a regional energy power generation scheme is generated based on the node residuals. Parallel hierarchical reasoning is performed on the regional energy power generation scheme to obtain the unit power generation plan. The parallel hierarchical reasoning includes regional total amount judgment and power plant power generation share refinement. The feasibility of the unit's power generation plan is verified, and the target unit's power generation plan to be executed is determined based on the verification results.

2. The method according to claim 1, characterized in that, The process of acquiring multi-source power operation data and power system topology maps, and constructing a graph-structured dataset with version identifiers includes: The system acquires multi-source power system operation data, maps the identifiers of power equipment from different sources in the multi-source power system operation data to global node identifiers based on preset hierarchical coding rules, and constructs a power system topology map in combination with preset power grid connection relationships. The multi-source power system operation data is time-aligned and format-standardized to obtain a time-uniform time-series dataset. The missing values ​​in the time series dataset are differentiated according to the preset data quality indicators, and the data confidence vector of each data point is generated synchronously. Bind static attribute data in the multi-source power system operation data to the corresponding unit nodes; The time-series data, the data confidence vector, the static attribute data, and the power system topology map are fused and written into the graph database to generate a graph structure data snapshot with version identifier and effective time interval. The graph structured dataset includes the graph structure data snapshot.

3. The method according to claim 1, characterized in that, The information flow channels include installation availability flow, weather-driven flow, and market demand flow. The step of performing spatiotemporal joint coding on each of the information flow channels to obtain the node-level feature representation of each of the information flow channels includes: By combining temporal convolution and graph convolution operators, feature encoding is performed on the installed capacity availability stream, the weather-driven stream, and the market demand stream, respectively, to obtain node-level feature representations of the installed capacity availability stream, the weather-driven stream, and the market demand stream.

4. The method according to claim 3, characterized in that, The method of using a combination of temporal convolution and graph convolution operators to perform feature encoding on the installed capacity availability stream, the weather-driven stream, and the market demand stream, respectively, to obtain node-level feature representations of the installed capacity availability stream, the weather-driven stream, and the market demand stream, includes: For the installed capacity availability flow, the shutdown or de-processing status is marked according to the unit maintenance plan, and the dispatchable capacity margin is determined recursively along the progressive direction of the unit, power plant and region to obtain the time-varying capacity matrix; For the meteorological driving flow, meteorological forecast data is loaded to supplement the wind speed, irradiance and inflow prediction sequences, and meteorological fluctuation signals are propagated through geographical adjacency edges; In the market demand flow, the market electricity price and contracted electricity volume are converted into a demand pull tensor and propagated backward from the load center to the generation-side nodes along the transmission network; For the installed capacity availability stream, the weather-driven stream, and the market demand stream, short-term dynamic features are extracted using temporal convolution, and spatial coupling effects are captured using graph convolution. The extracted dynamic features and spatial coupling effects are then fused spatiotemporally to obtain node-level feature representations of the installed capacity availability stream, the weather-driven stream, and the market demand stream.

5. The method according to any one of claims 1 to 4, characterized in that, The node-level feature representations are weighted and fused using a gated fusion mechanism to obtain a node fusion embedding vector, including: The maintenance blockade window, scheduling priority, and emission reduction weight are mapped to learnable gating vectors, and the node-level feature representations are weighted and fused by element-wise multiplication based on the learnable gating vectors to obtain the node fusion embedding vector.

6. The method according to any one of claims 1 to 4, characterized in that, The process involves coupling the node fusion embedding vector with a preset dual-gated residual operator to obtain node residuals, and generating a regional energy power generation scheme based on these residuals, including: By coupling the node fusion embedding vector with the preset historical output baseline through a preset dual-gated residual operator, a node residual that takes into account both real-time fluctuations and historical inertia is obtained. The residual modulus of each unit node in the region is obtained by summing the absolute values ​​of the node residuals along the time dimension. An adaptive weighting coefficient is generated based on the residual modulus of each unit node in the region, and the output of units of the same type is weighted and aggregated based on the adaptive weighting coefficient to generate a draft of regional energy power generation capacity. The draft regional energy power generation capacity is subjected to peak-valley smoothing to obtain the regional energy power generation capacity scheme.

7. The method according to claim 6, characterized in that, The parallel hierarchical reasoning of the regional energy power generation scheme to obtain the unit power generation plan includes: A parallel branch structure is used to perform regional total power inference on the regional energy power generation scheme. The first branch structure extracts the slow variation trend of the regional energy power generation scheme, and the second branch structure captures the high frequency oscillation of the regional energy power generation scheme. The outputs of the first branch structure and the second branch structure are dynamically fused based on the confidence level of the weather forecast and the penetration rate of new energy sources to obtain regional supply forecast data. Based on the regional residual between the regional supply forecast data and the regional energy power generation scheme, an initial unit power generation plan is generated by combining historical operating modes and contractual constraints. The physical feasibility and demand pull of the initial unit power generation plan are coupled through a cross-attention mechanism. After multiple rounds of residual convergence iteration, the converged unit power generation plan is output.

8. The method according to any one of claims 1 to 4, characterized in that, The step of performing a feasibility verification on the unit's power generation plan and determining the target unit's power generation plan to be executed based on the verification results includes: Based on the equipment's physical parameters and preset external compliance constraints, the unit's power generation plan is subjected to dual verification of equipment physical limits and compliance operation constraints, and the verification results are obtained. Based on the verification results, the target unit power generation plan to be executed is determined.

9. A device for determining a generator unit's power generation plan, characterized in that, The device includes: The data acquisition module is used to acquire multi-source power operation data and power system topology maps, and to construct a graph-structured dataset; The feature extraction module is used to construct multiple independent information flow channels based on the graph structured dataset, perform spatiotemporal joint encoding on each information flow channel, and obtain node-level feature representations of each information flow channel. Each information flow channel corresponds to a type of business constraint that affects power generation capacity. The feature fusion module is used to perform weighted fusion of the node-level feature representations through a gated fusion mechanism to obtain a node fusion embedding vector. The scheme determination module is used to couple the node fusion embedding vector and the preset historical output baseline through a preset dual-gated residual operator to obtain the node residual, and generate a regional energy power generation scheme based on the node residual; The parallel reasoning module is used to perform parallel hierarchical reasoning on the regional energy power generation scheme to obtain the unit power generation plan. The parallel hierarchical reasoning includes regional total amount judgment and power plant power generation share refinement. The data verification module is used to verify the feasibility of the unit's power generation plan and determine the target unit's power generation plan to be executed based on the verification results.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images