Pumped storage power station energy optimization dispatching system based on digital twinning

Abstract

The invention discloses an energy optimization dispatching system of a pumped storage power station based on digital twinning, particularly relates to the field of electric power energy optimization, and is used for solving the problem of hidden loss of an energy path and mismatch of output rhythm caused by structural dislocation of market time sequence and mechanism response in a traditional dispatching method. A baseline data set and a checkpoint scheme are generated by aligning a scheduling sequence and a mechanism state sequence through a unified time axis, a hydroelectric coupling directed graph is constructed for path search, a candidate path set and a sub-graph are generated through layered screening, elimination and compensation, and transition response simulation is executed in a digital twin body to calculate an energy arc, an energy arc loss ratio and transient dissipation specific energy. A response deviation vector sequence is formed, a discrimination coefficient optimal operation path is obtained through label propagation, finally, an instruction sequence is expanded, forward replacement is triggered according to a check point sequence, a correction version instruction sequence is output, and stable convergence of an energy path and intraday disturbance elastic landing are achieved.

Description

Energy Optimization and Dispatch System for Pumped Storage Power Stations Based on Digital Twin Technical Field

[0001] This invention relates to the field of power energy optimization, and more specifically, to an energy optimization and scheduling system for pumped storage power plants based on digital twins. Background Technology

[0002] Pumped storage power plants play a crucial role in energy transport and rhythm restoration within the new power structure. Their scheduling is guided by both market forces and the temporal constraints of hydraulic-electromechanical mechanisms. Digital twins are used to reconstruct the interrelationships between generating units, reservoirs, hydraulic channels, and grid constraints in virtual space, enabling strategy simulations of charging / discharging paths, start-up / shutdown transitions, and grid demand during the day-ahead phase. Echoing mainstream practices in integrated energy systems, the paper "Digital twins based day-ahead integrated energy system scheduling under load and renewable energy uncertainties" by M. You et al. in Applied Energy proposes mapping the physical system to a multi-energy-flow virtual entity. Day-ahead optimization is achieved through interaction with the physical entity, and data-driven models absorb prediction errors, forming cost-oriented scheduling schemes. This embodies the basic framework of "virtual-physical interaction—day-ahead optimization—operational implementation." This approach reveals a practical insight: digital twins can aggregate uncertainties and equipment constraints during the day-ahead phase, providing a transferable solution framework for energy path planning in pumped storage scenarios.

[0003] In pumped storage scenarios, the problem is not whether a usable day-ahead twin scheduling framework exists, but rather why, how, and how structural misalignments occur between market timing, mechanistic response, and safety boundaries. This is because pumped storage involves path-dependent head evolution, pump-machine duplex switching costs, and short-term hydraulic transients. When intraday disturbances and reservoir backwater feedback intertwine, scheduling variables and mechanistic variables are no longer equivalent to general energy storage abstractions. In terms of execution mechanisms, if the general twin solution approach for integrated energy systems is adopted, current strategies are mainly driven by load and electricity price forecasts. The mechanistic layer is treated as boundary conditions or static efficiency curves. Time-shifting effects and transition losses in the hydraulic process are difficult to jointly model within the same target domain, resulting in lags and backtracking between virtual strategies and physical responses in terms of rhythm. The consequences are reflected in three aspects: First, the strategy fluctuates frequently during short-cycle corrections, resulting in hidden losses in the energy path; second, there is a mismatch between reservoir capacity and output rhythm, creating a tug-of-war between water allocation and grid-side demand; and third, the transition process during start-up, shutdown, and ramp-up phases is underestimated, making it difficult to converge intraday repairs in a timely manner.

[0004] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an energy optimization scheduling system for pumped storage power stations based on digital twins. This system generates a baseline dataset and checkpoint scheme by aligning scheduling sequences and mechanistic state sequences along a unified time axis. It then constructs a hydroelectric coupling directed graph for path searching, generates candidate path sets and subgraphs through hierarchical screening, performs transient response simulation within the digital twin to calculate energy arcs, energy arc loss ratios, and transient dissipation specific energy, forming a response deviation vector sequence. Label propagation is used to obtain discrimination coefficients to optimize the operating path. Finally, the system expands the instruction sequence and triggers forward replacement according to the checkpoint order to output a revised instruction sequence, thereby solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] Time alignment module: Aligns scheduling sequences and mechanism state sequences on a unified time axis, solidifies them into a baseline dataset, and generates a checkpoint scheme;

[0008] Graph Construction Module: Constructs a directed graph of hydroelectric coupling based on the baseline dataset to form a path search structure;

[0009] Path filtering module: Filters paths in the directed graph according to the hierarchical decision order, removes and supplements graph edges to generate a candidate path set, generates a candidate path subgraph and binds checkpoint schemes;

[0010] Response simulation module: Performs transient response simulation on the candidate path set, outputs energy arc, energy arc loss ratio, transient dissipation specific energy, and generates response deviation vector sequence by slicing according to the check point scheme;

[0011] The optimal discrimination module: On the candidate path subgraph, the discrimination coefficient is obtained by performing tag propagation based on the energy arc loss ratio and transient dissipation ratio. The running path is obtained by reducing and rearranging according to the discrimination coefficient.

[0012] Instruction correction module: Expands the instruction sequence according to the running path and references the checkpoint scheme. It calls the response deviation vector sequence in the order of checkpoints to trigger the forward replacement strategy and outputs the corrected instruction sequence.

[0013] Furthermore, the time alignment module collects the timestamps and sampling periods of the scheduling sequence and the mechanism state sequence, unifies them into a single time axis, and resamples them according to a unified period. Missing measurement positions are filled forward, and the starting segment is maintained by proximity. Time consistency checks are performed on the resampled scheduling power, reservoir capacity status, head status, and start / stop status to form a baseline dataset containing a time index and four types of fields. A set of checkpoints is constructed based on the time of strategy change, the time of working condition switch, and the time of head inflection point. The time axis is divided into a set of non-overlapping adjacent small windows to generate a checkpoint scheme, and the start and end indexes of the windows and the time mapping table are recorded in the baseline dataset.

[0014] Furthermore, the graph construction module extracts reservoir capacity status intervals and unit operating condition sets from the baseline dataset. It constructs a node set using Cartesian combinations of reservoir capacity status intervals and unit operating condition sets, establishes power transition edges according to time order and physical reachability, and writes a transition cost set and constraint instructions on each power transition edge. The transition cost set includes start-stop switching cost, power ramp-up cost, and minimum maintenance time cost, while the constraint instructions include water conservation, upper and lower limits of reservoir capacity, operating condition restricted areas, minimum maintenance time, and window index mapping. It calculates the topology order and establishes the reference relationship between nodes and power transition edges to the time axis for subsequent path search and time consistency verification.

[0015] Furthermore, the path selection module performs path search on the hydro-electric coupling directed graph in a hierarchical decision-making order. The first layer retains power transition edges that satisfy water conservation and reservoir capacity limits. The second layer retains only power transition edges that satisfy the restricted operating conditions and minimum maintenance time. The last layer compares the cumulative power trajectory of the path with the scheduled power trajectory hourly and removes power transition edges that do not meet time consistency. When a breakpoint occurs, a local replacement edge is selected within the current window. The local replacement edge is limited to a reachable edge whose starting point is located at the window's starting index and whose ending point is located at the window's ending index, and which simultaneously satisfies water conservation, reservoir capacity limits, restricted operating conditions, minimum maintenance time, and time consistency constraints. Based on this, the breakpoint is filled and a candidate path set is generated.

[0016] Furthermore, the path selection module performs edge guidance on the nodes and power transition edges appearing in the candidate path set, generates a candidate path subgraph, binds the check point scheme to the relevant power transition edge record of the candidate path subgraph, and outputs the candidate path to the window mapping table.

[0017] Furthermore, the response simulation module performs transient response simulations on each candidate path in the digital twin, outputting a simulated power sequence, loss power sequence, and flow sequence consistent with the time axis. It generates an energy arc by time integration of the simulated power sequence, calculates the energy arc loss ratio based on the ratio of the discrete second-order difference to the first-order difference, and uses it to characterize the bending strength of the energy transmission curve. Within the small window of the checkpoint scheme, it calculates the transient dissipation specific energy per unit water mass using the loss power sequence and flow sequence, and uses it to quantify the dissipation intensity during the transition phase.

[0018] Furthermore, the response simulation module arranges the energy arc loss ratio and transient dissipation ratio of each window into a response deviation vector according to the window order, forming a response deviation vector sequence. It then writes the response deviation vector back to the candidate path subgraph with a three-level index consisting of path identifier, power transition edge identifier, and window index. At the same time, it generates a path-level summary table of energy arc loss ratio and transient dissipation ratio.

[0019] Furthermore, the selection and discrimination module generates initial labels on the candidate path subgraph based on the path-level energy-arc loss ratio and transient dissipation ratio. Paths with low loss and low dissipation are mapped to high-confidence positive labels and their overlying power transition edges are marked as positive sources. Paths with high loss or high dissipation are mapped to low-confidence negative labels and their overlying power transition edges are marked as negative sources. Label propagation is performed according to the topological order of the candidate path subgraph until convergence, and the power transition edge level confidence is obtained.

[0020] Furthermore, the selection and discrimination module aggregates the discrimination coefficients of each candidate path according to the confidence level of the coverage power transition edge using a fixed lower quantile caliber, eliminates low-order paths according to the threshold order of the discrimination coefficients, and rearranges the remaining paths according to the discrimination coefficients from high to low, outputs the mapping table of running paths and candidate paths to discrimination coefficients, and marks the running paths back to the candidate path subgraph.

[0021] Furthermore, the instruction correction module unfolds the instruction sequence according to the time sequence based on the running path, and generates a mapping table from instruction fragments to window indices by referencing the checkpoint scheme. It reads the current window component of the response deviation vector sequence in the checkpoint order and compares it with the path-level summary table index of the same path. When an unfavorable increase occurs, a forward replacement is performed on the set of adjacent power transition edges within the same window in the candidate path subgraph. The replacement candidates must meet the constraints of water conservation, upper and lower limits of reservoir capacity, operating condition restricted areas, minimum maintenance time, and time consistency. After each replacement, the revised version of the running path, the instruction sequence to window mapping, and the time consistency record are updated immediately, and the replacement record is recorded. After completing the sequential scan of all windows, the revised instruction sequence and replacement record are output and synchronously written back to the execution area of ​​the baseline dataset.

[0022] The technical effects and advantages of the energy optimization scheduling system for pumped storage power stations based on digital twins in this invention are as follows:

[0023] This invention starts with a time-consistent baseline, locking the scheduling sequence and mechanism state onto the same reference plane. A candidate set is formed within the constraint domain defined by the coupled directed graph. Two complementary types of information, energy arc morphology deviation and transient dissipation intensity, are extracted by digital twin simulation. These are then transformed into discriminant coefficients through graph domain label propagation. The forward replacement is driven by the segmented deviation vector of the checkpoint, making the strategy selection and physical response rhythmically in sync, resulting in smoother start-stop transitions, non-accumulation of short-cycle corrections, suppression of latent losses during the generation stage, consistency between scheduling trajectory and water allocation, stable convergence of the running path, reduced need for repeated global re-solution, and the formation of a flexible landing capability for intraday disturbances. Attached Figure Description

[0024] Figure 1 is a schematic diagram of the structure of the energy optimization scheduling system for pumped storage power stations based on digital twins according to the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Example 1: Figure 1 shows the energy optimization scheduling system for pumped storage power stations based on digital twins according to the present invention, including:

[0027] Time alignment module: Aligns scheduling sequences and mechanism state sequences on a unified time axis, solidifies them into a baseline dataset, and generates a checkpoint scheme;

[0028] Graph Construction Module: Constructs a directed graph of hydroelectric coupling based on the baseline dataset to form a path search structure;

[0029] Path filtering module: Filters paths in the directed graph according to the hierarchical decision order, removes and supplements graph edges to generate a candidate path set, generates a candidate path subgraph and binds checkpoint schemes;

[0030] Response simulation module: Performs transient response simulation on the candidate path set, outputs energy arc, energy arc loss ratio, transient dissipation specific energy, and generates response deviation vector sequence by slicing according to the check point scheme;

[0031] The optimal discrimination module: On the candidate path subgraph, the discrimination coefficient is obtained by performing tag propagation based on the energy arc loss ratio and transient dissipation ratio. The running path is obtained by reducing and rearranging according to the discrimination coefficient.

[0032] Instruction correction module: Expands the instruction sequence according to the running path and references the checkpoint scheme. It calls the response deviation vector sequence in the order of checkpoints to trigger the forward replacement strategy and outputs the corrected instruction sequence.

[0033] The time alignment module (data alignment and checkpoint scheme generation) collects the timestamps and sampling periods of the scheduling sequence and mechanism state sequence, unifies them into a single time axis, and resamples them according to a unified period. Missing measurement positions are filled forward, and the starting segment is maintained at the nearest point. Cross-checkpoint window interpolation is prohibited. Time consistency is checked on the resampled scheduling power, reservoir capacity status, head status, and start / stop status to form a baseline dataset containing time indexes and four types of fields. Checkpoint sets are constructed based on the time of strategy change, operating condition switch, and head inflection point. The time axis is divided into a set of non-overlapping adjacent small windows to generate a checkpoint scheme. The start and end indices of the windows and the time mapping table are recorded in the baseline dataset.

[0034] The graph construction module (construction of a directed graph for hydropower coupling) extracts reservoir capacity state intervals and unit operating condition sets from the baseline dataset. It constructs a node set using Cartesian combinations of reservoir capacity stages and unit operating conditions, and establishes power transition edges according to time order and physical reachability. It writes a transition cost set and constraint instructions to each edge. The transition cost set includes start-stop switching cost, power ramp-up cost, and minimum maintenance time cost. The constraint instructions include water conservation, upper and lower limits of reservoir capacity, operating condition restrictions, minimum maintenance time, and window index mapping. It calculates the topological order and establishes the reference relationship between nodes and edges to the time axis for subsequent path search and time consistency verification.

[0035] The path selection module (layered selection, elimination, and candidate path subgraph generation) performs path search on the hydropower coupled directed graph in a layered decision-making order. The first layer retains only edges that satisfy water conservation and reservoir capacity limits, the second layer retains only edges that satisfy operating condition restrictions and minimum maintenance time, and the last layer compares the cumulative power trajectory of the path with the scheduled power trajectory hourly, eliminating edges that do not meet time consistency requirements. When a breakpoint occurs, only local replacement edges are selected within the current window. Local replacement edges are limited to reachable edges whose starting point is located at the window's starting index and whose ending point is located at the window's ending index, and which simultaneously satisfy water conservation, reservoir capacity limits, operating condition restrictions, minimum maintenance time, and time consistency constraints. Based on this, the breakpoint is filled to generate a candidate path set. Edge guidance is implemented on the nodes and edges appearing in the candidate path set to generate a candidate path subgraph, and the checkpoint scheme is bound to the relevant edge records of the candidate path subgraph. At the same time, the candidate path is output to the window mapping table.

[0036] The response simulation module (transient response simulation and response deviation vector sequence generation) performs transient response simulation on each candidate path in the digital twin, outputting a simulated power sequence, loss power sequence, and flow sequence consistent with the time axis; it generates an energy arc by time integration of the simulated power sequence, and calculates the energy arc loss ratio based on the ratio of the discrete second-order difference to the first-order difference, which is used to characterize the bending strength of the energy transmission curve; within the small window of the checkpoint scheme, it calculates the transient dissipation specific energy per unit flow mass using the loss power sequence and flow sequence, which is used to quantify the dissipation intensity during the transition phase; it forms a response deviation vector by combining the two indicators of each window in window order, forming a response deviation vector sequence, and writes it back to the candidate path subgraph with a three-level index of "path identifier - edge identifier - window index", while generating a path-level summary table of the two indicators.

[0037] The optimal discrimination module (label propagation and discrimination coefficient generation, running path determination) generates initial labels on the candidate path subgraph based on the path-level energy-arc loss ratio and transient dissipation ratio. Paths with low loss and low dissipation are mapped to high-confidence positive labels and their covering edges are marked as positive sources. Paths with high loss or high dissipation are mapped to low-confidence negative labels and their covering edges are marked as negative sources. Label propagation is performed according to the topological order of the candidate path subgraph until convergence, and edge-level confidence is obtained. For each candidate path, discrimination coefficients are generated by aggregating them according to the confidence of the covering edges using a fixed lower quantile. Low-order paths are eliminated according to the threshold order of the discrimination coefficients, and the remaining paths are rearranged from high to low according to the discrimination coefficients. The running paths and the mapping table from candidate paths to discrimination coefficients are output, and the running paths are marked back on the candidate path subgraph.

[0038] The instruction correction module (instruction expansion and forward replacement execution) expands the instruction sequence in chronological order according to the running path and generates a mapping table from instruction fragments to window indices by referencing the checkpoint scheme; it reads the current window component of the response deviation vector sequence in checkpoint order and compares it with two path-level summary indicators of the same path. When an unfavorable increase occurs, it performs a forward replacement within the candidate path subgraph, limiting it to the set of adjacent edges of the same window. Replacement candidates must meet constraints such as water conservation, upper and lower limits of reservoir capacity, operating condition restrictions, minimum maintenance time, and time consistency; after each replacement, it immediately updates the corrected version of the running path, the instruction sequence to window mapping, and the time consistency record, and records the replacement; after completing the sequential scan of all windows, it outputs the corrected instruction sequence and replacement record, and synchronously writes it back to the execution area of ​​the baseline dataset.

[0039] Unfavorable increase is defined as the situation where the component value of the current window response deviation vector exceeds the path-level summary index of the same path, indicating that the deviation increases in the unfavorable direction.

[0040] In the energy optimization and scheduling of pumped storage power plants, digital twin technology enables day-ahead strategy simulation by virtually reconstructing the hydraulic-electric mechanical linkages of the physical system, thus addressing the complexity of market guidance and timing constraints. This approach emphasizes virtual-physical interaction to absorb uncertainty and ensure the cost-orientation and feasibility of the scheduling scheme. However, in practical applications, the scheduling sequence and the mechanistic state sequence often become disconnected due to sampling differences, leading to structural misalignments in path planning. To address this, the time alignment module focuses on data alignment and checkpoint generation to establish a unified time reference framework, ensuring that subsequent path search and simulation are conducted on a consistent baseline dataset.

[0041] 1.1 Timestamp and sampling period acquisition of scheduling sequences and mechanism state sequences, and unified time axis resampling:

[0042] The scheduling sequences of pumped storage power stations typically include power commands and start / stop signals, while the mechanistic state sequences encompass head, reservoir capacity, and unit operating conditions. These sequences may be collected using different timestamps and sampling periods, easily leading to timing misalignments. Therefore, this method first extracts the precise timestamp and corresponding sampling period for each data point from the scheduling sequence, while simultaneously extracting the same information from the mechanistic state sequences, converting all timestamps to Coordinated Universal Time (UTC) as a single time axis reference. Subsequently, the sampling period of the scheduling sequence is selected as the unified period for resampling of both sequences: during resampling, for the scheduling sequence, power and start / stop status are directly interpolated at each time point within the unified period; for the mechanistic state sequences, head, reservoir capacity, and operating condition data are similarly interpolated within the unified period. For missing data points encountered during resampling, a forward-filling method is used, filling the current missing data point with the previous valid data point; for missing data points at the beginning of the sequence, a nearest-neighbor hold method is used, copying the nearest valid data point until the beginning of the sequence. Specifically, interpolation across windows is prohibited between subsequently defined checkpoint windows to avoid introducing spurious continuity across time periods. Finally, aligned versions of the two sequences on a single time axis are obtained, providing a basis for time consistency verification and ensuring data synchronization across time.

[0043] 1.2 Time consistency verification of scheduling power, reservoir capacity status, head status, and start / stop status after resampling, and formation of baseline dataset:

[0044] Based on the aligned sequences obtained from the aforementioned resampling, the dispatch power of pumped storage power stations often reflects market demand, while reservoir capacity status, head status, and start-up / shutdown status reflect the physical mechanism response. These fields must be strictly consistent in time to avoid strategy deviations. Therefore, the resampled dispatch power sequence is checked point-by-point to ensure its time index matches a single time axis. Similarly, the reservoir capacity status sequence is verified point-by-point to confirm its correspondence with the single time axis; the same verification is performed on the head status sequence to confirm no time shift; and the start-up / shutdown status sequence is also checked point-by-point to detect any potential temporal inconsistencies. Subsequently, the verified dispatch power, reservoir capacity status, head status, and start-up / shutdown status are integrated into a structured table to form a baseline dataset. This dataset uses the time index of a single time axis as the primary key and contains four types of fields: the dispatch power field records the power command value, the reservoir capacity status field records the reservoir volume or water level, the head status field records the effective head height, and the start-up / shutdown status field records the unit operating mode. This formation process ensures that the baseline dataset becomes a reliable starting point for subsequent path construction, eliminating the impact of temporal misalignment on mechanistic constraints.

[0045] 1.3 Construction of checkpoint sets and generation of checkpoint schemes by dividing small windows at strategy change time, operating condition switching time, and head inflection point time:

[0046] Based on the established baseline dataset, the timing of strategy changes in pumped storage power stations marks significant adjustments in dispatched power, the timing of operating condition switching indicates the transition between start-up and shutdown states, and the timing of head inflection points captures the inflection points in head state to reflect hydraulic dynamics. To construct the checkpoint set, the dispatched power field of the baseline dataset is first scanned to identify strategy change times, i.e., points where power values ​​change beyond a preset threshold. Then, the start-up / shutdown state field is scanned to mark operating condition switching times, i.e., points where the state changes from pumping to power generation or vice versa. Next, the head state field is scanned, and the finite difference method is used to detect head inflection points, i.e., points where the sign of the first-order difference changes. These times are merged and deduplicated to form the checkpoint set. Subsequently, using the checkpoint set as a boundary, a single time axis is divided into a set of non-overlapping and adjacent small windows. Each small window starts from one checkpoint and ends at the next, thus generating a checkpoint scheme. This scheme includes the start and end times of the small windows and the corresponding checkpoint list. Finally, the start and end indices and time mapping table for each small window are recorded in the additional columns of the baseline dataset to ensure that the checkpoint scheme is closely bound to the baseline dataset. This generation process provides a time-segmentation framework for subsequent hierarchical decision-making and deviation assessment, strengthening the rhythmic matching between scheduling and mechanism.

[0047] The time alignment module successfully integrated the scheduling sequence and the mechanism state sequence into a baseline dataset and generated a checkpoint scheme by collecting and unifying timestamps and sampling periods, resampling data, verifying time consistency, and constructing a checkpoint set. This process eliminated temporal misalignment, ensured data synchronization on a single time axis, provided a solid starting point for energy path planning of pumped storage power plants, and avoided deviations and losses caused by data inconsistencies in subsequent steps.

[0048] The time alignment module has aligned the scheduling sequence and the mechanism state sequence through a unified time axis, generating a baseline dataset and checkpoint scheme, providing a time-consistent reference basis for subsequent processing. This baseline dataset includes fields such as scheduling power, reservoir capacity status, head status, and start-up / shutdown status, ensuring the synchronization of scheduling and mechanism. However, in the pumped storage scenario, the path dependency of reservoir capacity evolution and unit operating conditions requires the transformation of this data into a structured model to capture the linkage between hydroelectric machinery. To this end, step S2 constructs a hydroelectric coupling directed graph from the baseline dataset, forming a path search structure. By designing nodes and edges, it integrates time sequence and physical reachability, providing a constraint domain for subsequent path selection and simulation, ensuring that the structural misalignment between market timing and mechanism response in energy path planning is alleviated.

[0049] 2.1 Extract the storage capacity status range and unit operating condition set:

[0050] Starting with the baseline dataset generated by the time alignment module, extracting reservoir capacity status intervals and unit operating condition sets is the core starting point for constructing the graph model. This is because reservoir capacity status intervals reflect the continuous range of water volume changes, while unit operating condition sets capture discrete patterns such as charging / discharging and start-up / shutdown of pumped storage. The time index field of the baseline dataset is used to traverse all records, scanning the reservoir capacity status field in chronological order, aggregating consecutive and similar reservoir capacity values ​​into intervals. The interval boundaries are defined by abrupt changes in the reservoir capacity status field, which are determined by calculating the difference in reservoir capacity values ​​between adjacent records and comparing it with a preset threshold. Similarly, the unit operating condition field is scanned, grouping consecutive records with the same operating condition into sets, with each set recording the start and end time indices. The extraction results form a list of reservoir capacity status intervals and a list of unit operating condition sets. Each list item includes the time span, numerical range, and corresponding baseline dataset index, ensuring data traceability.

[0051] 2.2 Constructing a set of nodes:

[0052] The Cartesian combination of reservoir capacity state intervals and unit operating condition sets comprehensively covers the possible state space of pumped storage, avoiding path omissions caused by a single dimension. For the list of reservoir capacity state intervals and unit operating condition sets extracted in step S2.1, a Cartesian product operation is performed, pairing each reservoir capacity state interval with each unit operating condition set to form a node. Each node is represented by a tuple, where the first element is the reservoir capacity state interval and the second element is the unit operating condition set. The total number of nodes is the product of the lengths of the two lists. Each node is appended with a time sequence label, inherited from the time index field of the baseline dataset, ensuring implicit temporal relationships between nodes. The node set is output as a list for easy reference during subsequent edge construction.

[0053] 2.3 Establishing a power transition edge:

[0054] Connections between nodes must reflect temporal order and physical reachability to limit invalid paths and reflect actual operational constraints. Nodes are sorted by time-sequence labels from the node set. Physical reachability is determined by checking whether the reservoir capacity state intervals between adjacent nodes transition continuously and whether the unit operating conditions allow for switching. Continuous transition is defined as the starting value of the reservoir capacity state interval of the subsequent node not being lower than the ending value of the preceding node minus the allowable fluctuation range. Allowable switching is defined as the unit operating conditions changing from pumping to power generation, or vice versa, meeting the minimum maintenance time requirement. Directed edges are established between node pairs that meet the conditions, with the edge direction following the time-sequence labels from smallest to largest to ensure the graph's acyclicity. The power transition edge list records the starting node, ending node, and transition type, categorized as power ramp-up or start-stop switching.

[0055] 2.4 Writing the Transition Cost Set and Constraint Instructions:

[0056] Each power transition edge needs to embed a transition cost set and constraint instructions to quantify energy loss and safety boundaries, ensuring that mechanistic responses are considered during path search. The transition cost set includes start-stop switching cost, power ramp-up cost, and minimum sustaining time cost. The start-stop switching cost is estimated by calculating the energy consumption difference between the operating conditions of the units at both ends of the edge. The power ramp-up cost is calculated based on the power change rate of the edge span. The minimum sustaining time cost is the penalty value for violating the minimum sustaining time constraint. Constraint instructions include water conservation, reservoir capacity upper and lower limits, operating condition exclusion zones, minimum sustaining time, and window index mapping. Water conservation requires that the difference between the reservoir capacity state intervals at both ends of the edge equals the flow integral. Reservoir capacity upper and lower limits check whether the interval exceeds the preset limit. Operating condition exclusion zones exclude prohibited operating conditions. The window index mapping inherits the time window correspondence from the baseline dataset. The write process iterates through the power transition edge list one by one, attaching these sets and instructions as attributes to the edges.

[0057] 2.5 Calculate the topological order and establish reference relationships:

[0058] Calculating the topological order ensures the orderliness of path search, while the reference relationships between nodes and edges to the time axis maintain the accuracy of time consistency checks. For the power transition edge list and node set, a depth-first search algorithm is used to calculate the topological order, recursively traversing reachable nodes from the starting node, recording the visit order, and forming a topological order list.

[0059] The referencing relationships are established as follows: each node is associated with the time index of its reservoir capacity status interval and unit operating condition set; each edge is associated with the time span of its starting and ending nodes, ensuring alignment with the time axis of the baseline dataset. The output is a directed hydroelectric coupling graph, containing a set of nodes, a list of power transition edges, a list of topological order, and a referencing relationship mapping.

[0060] Step S2 extracts reservoir capacity state intervals and unit operating condition sets from the baseline dataset, constructs a node set and power transition edges, and embeds transition cost sets and constraint instructions to ultimately form a hydro-electric coupling directed graph. This process addresses the path dependence and time-shifting effects of pumped storage, achieving a structured integration of scheduling sequences and mechanistic states, ensuring efficient subsequent path selection within the constraint domain, and avoiding implicit energy path losses and rhythm mismatches.

[0061] Step S2 has constructed a hydro-electric coupling directed graph from the baseline dataset, integrating reservoir capacity state intervals, unit operating condition sets, power transition edges, transition cost sets, and constraint instructions, providing a structured constraint domain for path search. This hydro-electric coupling directed graph ensures the unity of temporal order and physical reachability, laying a reliable foundation for energy path planning. However, under the intraday disturbances and reservoir backwater feedback environment of pumped storage, a simple graph structure cannot directly generate executable paths. It is necessary to eliminate invalid edges through hierarchical screening and elimination mechanisms, and construct subgraphs to bind checkpoint schemes, thereby eliminating structural misalignments between market timing and mechanism response, and ensuring the practicality and rhythmic consistency of the candidate path set. To this end, step S3 performs path screening according to the hierarchical decision-making order, eliminates and supplements graph edges to generate a candidate path set, and finally forms a candidate path subgraph, enhancing the flexible implementation capability of the scheduling strategy.

[0062] 3.1 Hierarchical Decision Sequence Path Search and Filtering:

[0063] Path search in hydroelectric coupled directed graphs should follow a hierarchical decision-making order to apply constraints at each level, avoiding computational complexity and path invalidity caused by processing all at once.

[0064] First, a path search is performed on the hydroelectric coupling directed graph according to the first-level constraints. The first level retains only power transition edges that satisfy water conservation and reservoir capacity limits. Water conservation is verified by checking whether the difference between the reservoir capacity state intervals at both ends of the edge is equal to the flow integral. The reservoir capacity limits are compared by checking whether the interval values ​​are within preset boundaries. Edges that meet the conditions are marked as valid edges in the first level, forming a subset of first-level paths.

[0065] Next, we move to the secondary constraint layer. Only edges that satisfy the restricted operating conditions and minimum maintenance time are retained based on the valid edges of the primary layer. The restricted operating conditions check whether the unit operating conditions associated with the edge avoid the prohibited area, and the minimum maintenance time verifies whether the edge span exceeds the specified duration. After secondary filtering, the path subset is updated.

[0066] The final layer compares the cumulative power trajectory of the paths with the scheduling power trajectory hourly. The cumulative power trajectory of the paths is generated from the power transition types on the edges, while the scheduling power trajectory is extracted directly from the scheduling power field of the baseline dataset. The hourly comparison is achieved by calculating the power difference at each time point and accumulating the absolute difference. Edges whose cumulative difference exceeds a threshold are removed to ensure temporal consistency. The hierarchical order ensures that the constraints are gradually tightened, generating a filtered set of paths, providing efficient input for subsequent elimination and supplementation.

[0067] 3.2 Generating a candidate path set by edge removal and patching:

[0068] The filtered path set may contain breakpoints. These breakpoints arise from missing edges caused by constraint filtering and require local replacement to fill in the gaps and ensure the integrity of the candidate path set. When a breakpoint appears in the path set, local replacement edges are selected within the current window. The current window is obtained from the set of small windows of the checkpoint scheme. The starting point of the local replacement edge is located at the starting index of the window, and the ending point is located at the ending index of the window.

[0069] The selection process checks whether the alternative edges simultaneously satisfy constraints related to water conservation, reservoir capacity limits, restricted operating conditions, minimum maintenance time, and time consistency. Water conservation is verified by verifying that the interval difference equals the flow integral. Reservoir capacity limits are compared against interval boundaries. Restricted operating conditions are avoided. Minimum maintenance time checks the span. Time consistency is achieved by matching the time index with the baseline dataset. Eligible alternative edges are inserted at breakpoint locations to fill the path and form a continuous path.

[0070] Repeat this process until all breakpoints are eliminated, generating a candidate path set. For each path, record the list of power transition edges it covers and the window mapping, ensuring that the path set covers the entire time axis.

[0071] 3.3 Generate candidate path subgraphs and bind checkpoint schemes:

[0072] The candidate path set needs to be transformed into a subgraph to support subsequent label propagation and deviation vector processing, while binding checkpoint schemes to maintain window-level constraints. For nodes and power transition edges appearing in the candidate path set, edge-induced subgraph generation is implemented. Edge induction constructs a subgraph by retaining all edges covered by the candidate paths and their connected nodes. The subgraph inherits the topological order of the hydroelectric coupling directed graph.

[0073] The binding process maps the set of small windows of the checkpoint scheme to the relevant edges of the subgraph. Specifically, it records the start and end indices of the window that each edge crosses and generates a candidate path to a window mapping table. The mapping table uses the path identifier as the key and the value as a list of window indices.

[0074] Output a candidate path subgraph, including a subset of nodes, a subset of edges, a topological order, and binding records, to ensure seamless integration of the subgraph with the checkpoint scheme.

[0075] Step S3 filters paths and removes breakpoints through a hierarchical decision-making sequence to generate a candidate path set, and constructs a candidate path subgraph binding checkpoint scheme, thus achieving effective refinement of the hydroelectric coupling directed graph. This process addresses the path dependence and intraday disturbances of pumped storage, ensuring that the candidate path set is complete and rhythmically consistent within the constraint domain, significantly reducing implicit losses and output mismatch, and providing efficient structural support for subsequent transient response simulation.

[0076] Step S3 has already screened paths according to the hierarchical decision-making order and generated a candidate path set by eliminating and supplementing them. Simultaneously, it constructs a checkpoint scheme for binding candidate path subgraphs, providing a constrained and refined path framework for subsequent simulations. This candidate path set ensures the feasibility of paths in terms of water conservation and time consistency, avoiding wasted computational effort. However, under the short-term hydraulic transients and transient losses of pumped storage, static paths cannot directly assess the mechanistic response. Dynamic indicators need to be extracted through simulation within a digital twin to quantify energy arc morphology deviations and transient dissipation intensity, thereby bridging structural misalignments between market timing and safety boundaries. Therefore, step S4 performs transient response simulations on the candidate path set, outputting energy arcs, energy arc loss ratios, and transient dissipation specific energy. It also generates a response deviation vector sequence by slicing according to the checkpoint scheme, enhancing the physical response verification capability of the strategy.

[0077] In this invention, digital twin technology serves as a pre-defined auxiliary tool, used solely to simulate the virtual environment of a pumped-storage power station to support the execution of transition response simulations. It is not the core innovation or focus of this invention, but rather considered part of the existing technical framework for extracting data such as simulated power sequences, loss power sequences, and flow sequences. The main focus of this invention is on the construction of a directed graph of hydroelectric coupling based on a baseline dataset, hierarchical path selection, label propagation discrimination, and an optimized scheduling mechanism for forward replacement strategies. These mechanisms aim to address the structural misalignment between market timing and mechanistic response. The digital twin only appears as a simulation carrier in step S4, without involving details of its internal construction or improvement.

[0078] 4.1 Perform transient response simulation:

[0079] Each path in the candidate path set represents a potential scheduling scheme, but it needs to be simulated in a digital twin to capture dynamic responses. The candidate path set is traversed path by path, and a digital twin is invoked for simulation of each path. The digital twin is a virtually reconstructed model of the generating unit, reservoir area, hydraulic channel, and network side. The simulation input is a list of power transition edges covered by the path, including the transition cost set and constraint instructions, ensuring that the simulation process adheres to water conservation and no-go zones. The simulation outputs a simulation power sequence, a loss power sequence, and a flow sequence consistent with the time axis. The simulation power sequence records the output power at each time point, the loss power sequence captures energy losses during transitions, and the flow sequence tracks water volume changes. The entire simulation uses the finite element method to simulate hydraulic transients, ensuring that the sequence length matches the time index of the baseline dataset.

[0080] 4.2 Generating the energy arc and calculating the energy arc loss ratio:

[0081] The simulated power sequence provides time-series data of energy transfer. Energy curves are generated through integration to evaluate the overall efficiency of the path. For each path's simulated power sequence, time integration is performed to calculate the energy curve. The integration proceeds from the start to the end of the time axis, accumulating the power at each time point multiplied by the time interval to obtain the cumulative energy curve. Next, the energy curve deflection ratio is calculated to characterize the bending strength of the energy transfer curve. The energy curve deflection ratio is calculated based on the ratio of the discrete second-order difference to the first-order difference. The first-order difference is obtained by subtracting adjacent integral values, and the second-order difference is obtained by subtracting the first-order difference. The ratio reflects the degree of curve bending. The formula is: ,in This represents the first-order difference energy change. This represents the second-order difference energy change.

[0082] 4.3 Calculation of transient dissipation specific energy:

[0083] Within the small window of the checkpoint scheme, the dissipation intensity of the transition phase needs to be quantified to identify hidden losses. For each small window, data is extracted from the loss power sequence and flow rate sequence. The small window is defined by the start and end indices of the checkpoint scheme. The transient dissipation specific energy is calculated as the energy dissipation per unit flow mass. The process involves accumulating the loss power sequence within the window, multiplying by the time interval to obtain the total dissipated energy, and dividing by the accumulated flow rate sequence multiplied by the time interval to obtain the total flow mass. The formula for transient dissipation specific energy is: ,in To reduce power loss, For traffic, For time intervals.

[0084] 4.4 Generate the response bias vector sequence and write it back to the subgraph:

[0085] The response deviation vector sequence integrates window-level indicators to provide segmented deviation information for subsequent discrimination. Following the window order of the checkpoint scheme, the energy-arc loss ratio and transient dissipation ratio of each window are combined to form a response deviation vector. The vector is a binary tuple, with the first element being the energy-arc loss ratio and the second element being the transient dissipation ratio. The sequence length equals the number of small windows, and each vector is associated with a window index. After forming the response deviation vector sequence, a three-level index—path identifier, edge identifier, and window index—is written back to the candidate path subgraph. The path identifier is inherited from the candidate path set, the edge identifier corresponds to the edge in the subgraph, and the window index is obtained from the bound record. Simultaneously, a path-level summary table is generated. By summing the elements of all vectors in the sequence, the total path loss and total dissipation are obtained.

[0086] Step S4 generates energy arcs, energy arc loss ratios, and transient dissipation specific energy by performing transient response simulations on the candidate path set. These are then sliced ​​according to the checkpoint scheme to form a response deviation vector sequence, achieving dynamic path verification. This process addresses the path dependence and transient losses of pumped storage, quantifying energy arc morphology deviations and transient dissipation intensity. It ensures that the strategy and physical response are in sync, reducing implicit losses and intraday repair needs, and providing deviation information support for subsequent operational path determination.

[0087] Step S4 has performed transient response simulation on the candidate path set, generating energy arcs, energy arc loss ratios, and transient dissipation specific energy, and forming a response deviation vector sequence, providing quantitative indicators for path optimization. This response deviation vector sequence captures the dynamic deviations during the transition phase, ensuring accurate evaluation of the mechanistic response. However, under the path dependence and short-cycle correction environment of pumped storage, these indicators need to be transformed into global discrimination signals through a graph domain propagation mechanism to avoid frequent fluctuations and hidden losses. To this end, step S5 performs label propagation on the candidate path subgraph based on the energy arc loss ratio and transient dissipation specific energy to obtain discrimination coefficients, and then reduces and rearranges the paths based on the discrimination coefficients to obtain the running paths, thus enhancing the stable convergence capability of the strategy.

[0088] 5.1 Generate initial tags:

[0089] Path-level metrics on the candidate path subgraph need to be mapped to labels to initialize the propagation process.

[0090] The path-level summary table is traversed. For each path, its arc loss ratio and transient dissipation specific energy are checked. If both are below a preset threshold, it is mapped to a high-confidence positive label, and the power transition edges it covers are marked as positive sources. If either indicator is above the threshold, it is mapped to a low-confidence negative label, and the power transition edges it covers are marked as negative sources. The threshold is dynamically determined from the distribution in the path-level summary table to ensure that the label assignment reflects the overall deviation level. After the initial labels are generated, the edge attribute fields of the candidate path subgraph are updated to record the label value and source type.

[0091] 5.2 Perform tag propagation:

[0092] Label propagation utilizes subgraph topology to transmit confidence signals, enabling edge-level evaluation.

[0093] Following the topological order of the candidate path subgraph, the label propagation algorithm is executed starting from positive and negative sources until the confidence converges. During propagation, the label values ​​of adjacent edges are collected for each edge, and the current edge label is updated using a majority voting mechanism, with voting limited to adjacent edges covered by the same path. The convergence condition is that the change in confidence is less than a threshold in consecutive iterations. After completion, edge-level confidence scores are obtained, with each confidence score being a value between 0 and 1, representing the edge reliability.

[0094] 5.3 Generating discriminant coefficients:

[0095] Edge-level confidence scores need to be aggregated into path-level metrics to support pruning operations. For each candidate path, extract the confidence list of the power transition edges it covers, and aggregate them using a fixed lower quantile to generate discriminant coefficients. The lower quantile is the value at the bottom 25% of the list, ensuring that path weaknesses are captured. The discriminant coefficients range from 0 to 1, with higher values ​​indicating overall path reliability.

[0096] 5.4 The running path is obtained by deletion and rearrangement:

[0097] Discriminant coefficients are used to optimize the path set for efficient selection. Low-order paths are eliminated based on the threshold order of the discriminant coefficients, with the threshold order being the coefficients sorted from smallest to largest; paths below the median are removed. The remaining paths are then rearranged in descending order of discriminant coefficients to form a list of running paths. Simultaneously, a mapping table between candidate paths and discriminant coefficients is output, and the edge identifiers of the running paths are annotated back to the corresponding records in the candidate path subgraph.

[0098] Step S5 achieves optimal selection on the candidate path subgraph by generating initial labels, performing label propagation, aggregating discrimination coefficients, and pruning and rearranging, thus obtaining the operating path. Addressing the implicit energy losses and output mismatch in pumped storage, the deviation index is transformed using graph domain propagation as a discrimination signal to ensure the rhythmic stability and convergence of the operating path, providing reliable path support for subsequent command deployment.

[0099] Step S5 generates discriminant coefficients through label propagation and uses these coefficients to prune and rearrange the operational path, providing an optimal solution for final scheduling. This operational path integrates energy arc morphology deviation and transient dissipation intensity, ensuring path reliability and low loss. However, under the intraday disturbances and backwater feedback environment of pumped storage, the operational path still needs to be expanded and segmented to cope with adverse deviations and avoid strategy oscillations and output mismatch. To this end, step S6 expands the instruction sequence based on the operational path and references the checkpoint scheme. It calls the response deviation vector sequence in the checkpoint order to trigger the forward replacement strategy and output a revised instruction sequence, achieving short-cycle adaptive optimization.

[0100] 6.1 Expanding the instruction sequence:

[0101] The optimized execution path needs to be transformed into an executable sequence of instructions to guide actual scheduling. First, the power transition edge list of the execution path is traversed chronologically, with each edge containing a transition type and constraint instructions. Starting from the originating edge, the power ramp-up cost and start / stop switching cost are extracted from its transition cost set. These are then combined with the water conservation constraint instructions to generate specific instructions, such as the power target value and duration. Edge-by-edge accumulation ensures that the time index of the instruction sequence is aligned with the baseline dataset, avoiding time drift. The instruction sequence is output as a list, with each element recording the time point, power instruction, and operating condition switching flag.

[0102] 6.2 Generate a mapping table from instruction fragments to window indices:

[0103] The instruction sequence needs to be associated with a checkpoint scheme to support segmented processing. A set of small windows referencing the checkpoint scheme is used to divide the instruction sequence into segments based on the window's start and end indices. Each segment corresponds to a subsequence of instructions within a single window. The mapping table uses window indices as keys and lists of instruction segments as values, ensuring that segment boundaries do not cross windows. The generation process verifies that each segment meets the minimum sustaining time constraint; if violated, the boundary time points are adjusted.

[0104] If the segment duration is less than the minimum maintenance time constraint, the adjustment rule is to extend backward from the end boundary time of the segment, borrowing part of the start time of the subsequent window, until the segment duration reaches or exceeds the minimum maintenance time, while verifying that the extended part does not violate the water conservation and no-go zone constraints; if the extension causes the subsequent window duration to be insufficient, some edges of the subsequent window are merged into the current segment, and the window index in the mapping table is updated to reflect the adjusted boundary; during the adjustment process, time consistency is maintained first, if it cannot be satisfied, backtracking to the end boundary of the previous window for forward shortening, but the shortening amount does not exceed 10% of the original segment duration to avoid excessive perturbation of the overall sequence.

[0105] 6.3 Trigger forward replacement by invoking the response bias vector sequence:

[0106] The response deviation vector sequence provides window-level deviation information to trigger local optimization. The response deviation vector of the current window is read in the order of the checkpoint scheme, and its components of energy arc loss ratio and transient dissipation specific energy are compared with the corresponding indicators in the path-level summary table. If an unfavorable increase occurs, i.e., the component value exceeds the summary indicator, a forward replacement is performed on the set of adjacent edges within the same window in the candidate path subgraph. Replacement candidates are selected from adjacent edges and must satisfy the following conditions: water conservation interval difference equals the flow integral, reservoir capacity upper and lower limits are within the boundaries, prohibited operating conditions are avoided, maintenance time span is sufficient, and time consistency matches the baseline dataset. The candidate with the lowest discriminant coefficient is selected to replace the current edge, and the revised running path and instruction sequence are updated to the window mapping.

[0107] 6.4 Output the revised instruction sequence:

[0108] After each replacement, the relevant records are updated immediately, and the entire window is scanned before the output is summarized. Updates include adjusting the edge list of the revised execution path, refreshing the fragment mapping from instruction sequences to windows, re-verifying time consistency records, and recording the replacement records, which include the replacement window, original edge, and new edge identifiers. After the scan is complete, the revised instruction sequence and replacement records are output, and the execution area field of the baseline dataset is written back synchronously. The execution area records the revised power and state sequences.

[0109] Step S6 ensures the flexible adjustment of the operating path under intraday disturbances by expanding the instruction sequence, generating a mapping table, triggering forward replacement, and outputting a revised version. It addresses the underestimation and convergence issues in the transition process of pumped storage, achieving rhythmic in-phase operation and suppressing latent losses, thus providing a stable and practical solution for energy optimization.

[0110] Specifically, the above description is only a preferred embodiment of this application and is not intended to limit this application.

[0111] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0112] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A pumped-storage power station energy optimization dispatch system based on digital twins, characterized in that, include: Time alignment module: Aligns scheduling sequences and mechanism state sequences on a unified time axis, solidifies them into a baseline dataset, and generates a checkpoint scheme; Graph Construction Module: Constructs a directed graph of hydroelectric coupling based on the baseline dataset to form a path search structure; Path Selection Module: Selects paths in the directed graph according to the hierarchical decision order, removes and supplements graph edges to generate a candidate path set, generates a candidate path subgraph and binds checkpoint schemes; Response Simulation Module: Performs transient response simulation on the candidate path set, outputs energy arc, energy arc loss ratio, and transient dissipation specific energy, and generates a response deviation vector sequence by slicing according to the checkpoint scheme; Optimization and Discrimination Module: Performs label propagation on the candidate path subgraph based on the energy arc loss ratio and transient dissipation specific energy to obtain discrimination coefficients, and obtains the running path by reducing and rearranging according to the discrimination coefficients; Instruction Correction Module: Expands the instruction sequence according to the running path and references the checkpoint scheme, calls the response deviation vector sequence according to the checkpoint order to trigger the forward replacement strategy and outputs the corrected instruction sequence.

2. The energy optimization scheduling system for pumped storage power stations based on digital twins according to claim 1, characterized in that: The time alignment module collects the timestamps and sampling periods of the scheduling sequence and the mechanism state sequence, unifies them into a single time axis, and resamples them according to a unified period. Missing measurement positions are filled forward, and the starting segment is maintained by proximity. Time consistency checks are performed on the resampled scheduling power, reservoir capacity status, head status, and start / stop status to form a baseline dataset containing time index and four types of fields. A set of checkpoints is constructed based on the time of strategy change, the time of working condition switch, and the time of head inflection point. The time axis is divided into a set of non-overlapping adjacent small windows to generate a checkpoint scheme. The start and end indexes of the windows and the time mapping table are recorded in the baseline dataset.

3. The energy optimization scheduling system for pumped storage power stations based on digital twins according to claim 2, characterized in that: The graph construction module extracts reservoir capacity status intervals and unit operating condition sets from the baseline dataset. It constructs a node set using Cartesian combinations of reservoir capacity status intervals and unit operating condition sets, and establishes power transition edges according to time order and physical reachability. It writes a transition cost set and constraint instructions to each power transition edge. The transition cost set includes start-stop switching cost, power ramp-up cost, and minimum maintenance time cost. The constraint instructions include water conservation, upper and lower limits of reservoir capacity, operating condition restricted areas, minimum maintenance time, and window index mapping. It calculates the topology order and establishes the reference relationship between nodes and power transition edges to the time axis for subsequent path search and time consistency verification.

4. The energy optimization scheduling system for pumped storage power stations based on digital twins according to claim 3, characterized in that: The path selection module performs path search on the hydroelectric coupled directed graph in a hierarchical decision-making order. The first layer retains power transition edges that satisfy water conservation and reservoir capacity limits. The second layer retains only power transition edges that satisfy the restricted operating conditions and minimum maintenance time. The last layer compares the cumulative power trajectory of the path with the scheduled power trajectory hourly and removes power transition edges that do not meet time consistency. When a breakpoint occurs, a local replacement edge is selected within the current window. The local replacement edge is limited to a reachable edge whose starting point is located at the window's starting index and whose ending point is located at the window's ending index, and which simultaneously satisfies water conservation, reservoir capacity limits, restricted operating conditions, minimum maintenance time, and time consistency constraints. Based on this, the breakpoint is filled and a candidate path set is generated.

5. The energy optimization scheduling system for pumped storage power stations based on digital twins according to claim 4, characterized in that: The path selection module performs edge guidance on the nodes and power transition edges appearing in the candidate path set, generates a candidate path subgraph, binds the check point scheme to the relevant power transition edge record of the candidate path subgraph, and outputs the candidate paths to the window mapping table.

6. The energy optimization scheduling system for pumped storage power stations based on digital twins according to claim 5, characterized in that: The response simulation module performs transient response simulations on each candidate path in the digital twin, outputting a simulated power sequence, loss power sequence, and flow sequence consistent with the time axis. It generates an energy arc by time integration of the simulated power sequence and calculates the energy arc loss ratio based on the ratio of the discrete second-order difference to the first-order difference, which is used to characterize the bending strength of the energy transmission curve. Within the small window of the checkpoint scheme, it calculates the transient dissipation specific energy per unit water mass using the loss power sequence and flow sequence, which is used to quantify the dissipation intensity during the transition phase.

7. The energy optimization scheduling system for pumped storage power stations based on digital twins according to claim 6, characterized in that: The response simulation module arranges the energy arc loss ratio and transient dissipation ratio of each window into a response deviation vector according to the window order, forming a response deviation vector sequence. It then writes the response deviation vector back to the candidate path subgraph with a three-level index of path identifier, power transition edge identifier, and window index. At the same time, it generates a path-level summary table of energy arc loss ratio and transient dissipation ratio.

8. The energy optimization scheduling system for pumped storage power stations based on digital twins according to claim 7, characterized in that: The optimal discrimination module generates initial labels on the candidate path subgraph based on the path-level energy-arc loss ratio and transient dissipation ratio. Paths with low loss and low dissipation are mapped to high-confidence positive labels and their overlying power transition edges are marked as positive sources. Paths with high loss or high dissipation are mapped to low-confidence negative labels and their overlying power transition edges are marked as negative sources. Label propagation is performed according to the topological order of the candidate path subgraph until convergence, and the power transition edge level confidence is obtained.

9. The energy optimization scheduling system for pumped storage power stations based on digital twins according to claim 8, characterized in that: The preferred discrimination module aggregates the discrimination coefficients of each candidate path using a fixed lower quantile caliber based on the confidence level of the coverage power transition edge. It then eliminates low-order paths according to the threshold order of the discrimination coefficients and rearranges the remaining paths from high to low according to the discrimination coefficients. The module outputs a mapping table from the running path and candidate path to the discrimination coefficients and marks the running path back to the candidate path subgraph.

10. The energy optimization scheduling system for pumped storage power stations based on digital twins according to claim 9, characterized in that: The instruction correction module unfolds the instruction sequence according to the time sequence based on the running path, and generates a mapping table from instruction fragments to window indices by referencing the checkpoint scheme. It reads the current window component of the response deviation vector sequence in the checkpoint order and compares it with the path-level summary table index of the same path. When an unfavorable increase occurs, it performs a forward replacement on the set of adjacent power transition edges within the same window in the candidate path subgraph. The replacement candidates must meet the constraints of water conservation, upper and lower limits of reservoir capacity, operating condition restricted areas, minimum maintenance time, and time consistency. After each replacement, the revised version of the running path, the instruction sequence to window mapping, and the time consistency record are updated immediately, and the replacement record is recorded. After completing the sequential scan of all windows, the revised instruction sequence and replacement record are output and synchronously written back to the execution area of ​​the baseline dataset.

Images