Energy digital management method and system

Abstract

The invention discloses an energy digital management method and system, and the method comprises the steps: receiving a multi-source heterogeneous energy data collection signal, and generating an initial energy characteristic spectrum based on a space-time correlation feature extraction algorithm; performing graph neural network coding on the initial energy feature graph, constructing an energy knowledge graph vector set, and decomposing a high-dimensional vector in the knowledge graph vector set into a plurality of subspace vectors; performing real-time scheduling decision on the subspace vectors by using an energy flow optimization model driven by reinforcement learning to generate an energy distribution strategy set; a digital twinborn verification algorithm is applied to the energy distribution strategy set, physical-digital space consistency verification is carried out, and a trusted energy scheduling scheme is generated; and performing block chain evidence storage and intelligent contract execution on the credible energy scheduling scheme to generate a non-tampering energy transaction record. By utilizing the embodiment of the invention, the scheduling precision and the response speed of the energy system can be improved, and full-link credible management from decision-making to execution is realized.

Description

Technical Field

[0001] This invention belongs to the field of energy digitalization technology, specifically an energy digitalization management method and system. Background Technology

[0002] With the large-scale and intelligent development of energy systems, the collection and management of multi-source heterogeneous energy data has become a core challenge for the industry. Traditional energy management methods typically rely on a single data source or static models, making it difficult to cope with the uncertainties of intermittent energy sources such as wind, solar, and energy storage, and also unable to achieve dynamic collaborative optimization across devices and regions. In existing technologies, rule-based or offline optimization-based scheduling strategies lack real-time performance, while simple data aggregation methods cannot fully exploit the spatiotemporal correlation characteristics of energy data, resulting in limited energy efficiency improvements. Summary of the Invention

[0003] The purpose of this invention is to provide a digital energy management method and system to address the shortcomings of existing technologies, improve the scheduling accuracy and response speed of energy systems, and achieve reliable management across the entire chain from decision-making to execution.

[0004] One embodiment of this application provides a digital energy management method, the method comprising: It receives multi-source heterogeneous energy data acquisition signals and generates an initial energy feature map based on a spatiotemporal correlation feature extraction algorithm, wherein each feature node is associated with the operating parameter dimension of the energy system; The initial energy feature map is encoded by a graph neural network to construct an energy knowledge graph vector set. The high-dimensional vectors in the knowledge graph vector set are decomposed into several subspace vectors to capture the dynamic characteristics of energy flow. Each subspace vector represents the multi-dimensional operating state of the corresponding energy device. A reinforcement learning-driven energy flow optimization model is used to make real-time scheduling decisions on the subspace vectors to generate a set of energy allocation strategies. The optimization model maximizes energy utilization efficiency through a Markov decision process. The digital twin verification algorithm is applied to the energy allocation strategy set to perform physical-digital space consistency verification and generate a reliable energy scheduling scheme. The digital twin verification algorithm ensures the feasibility of the strategy through a two-way mapping mechanism. The trusted energy dispatch scheme is implemented with blockchain notarization and smart contract execution to generate tamper-proof energy transaction records, which serve as the final execution basis for the digital management of the energy system.

[0005] Optionally, the receiving of multi-source heterogeneous energy data acquisition signals and the generation of an initial energy feature map based on a spatiotemporal correlation feature extraction algorithm, wherein each feature node is associated with the operating parameter dimension of the energy system, including: It receives multi-source heterogeneous energy data acquisition signals, adds a unified timestamp to each signal through high-precision clock synchronization technology, generates a spatiotemporal index label by combining the GPS coordinates installed on the device, filters out invalid acquisition points with signal strength below the strength threshold, and outputs a set of raw energy signals with spatiotemporal index. The original energy signal set with spatiotemporal index is preprocessed, and the signal components of different frequency bands are separated by wavelet packet decomposition algorithm. The noise components generated by electromagnetic interference are eliminated by adaptive threshold denoising. At the same time, the missing data is supplemented by interpolation method based on the similarity of adjacent time moments, and the denoised and completed energy data sequence is output. Based on the denoised and completed energy data sequence, the operating parameter dimensions of each energy device are extracted, the CFS feature selection algorithm is used to calculate the mutual information value between parameters, the core parameters that are strongly correlated with the energy system state are retained, and a parameter-device association table is generated. Based on the parameter-equipment association table, a spatiotemporal association rule mining algorithm is used to calculate the co-occurrence frequency and association strength of different equipment parameters. The equipment parameters are used as feature nodes and the association strength is used as edge weights to construct an initial energy feature map containing node attributes and edge attributes.

[0006] Optionally, the initial energy feature map is encoded using a graph neural network to construct an energy knowledge graph vector set. The high-dimensional vectors in the knowledge graph vector set are then decomposed into several subspace vectors to capture the dynamic characteristics of the energy flow. Each subspace vector represents the multi-dimensional operating state of the corresponding energy device, including: The importance of nodes in the initial energy feature map is evaluated. The PageRank algorithm is used to calculate the importance score of each feature node. The top 90% of nodes are retained to simplify the map structure. Weakly related edges with weights below the weight threshold are deleted, and a simplified energy feature map is output. The simplified energy feature map is input into the graph attention neural network. The dependencies between nodes are learned through the multi-head attention mechanism. The attribute features and associated features of each node are fused and encoded into a 128-dimensional high-dimensional vector. Each dimension of the vector corresponds to the abstract expression of the feature, and the energy knowledge graph vector set is output. Based on the dynamic characteristics of energy flow, the kernel principal component analysis algorithm is used to decompose the high-dimensional vectors in the energy knowledge graph vector set into subspaces. The decomposition dimension is set according to the energy transfer efficiency of each link, and a preliminary subspace vector set is output. The initial subspace vector set is processed by the Gram-Schmidt orthogonalization process to eliminate the coupling correlation between different vectors, so that each subspace vector independently represents the operating state of a single energy link, and the optimized subspace vector set is output.

[0007] Optionally, the step of using a reinforcement learning-driven energy flow optimization model to make real-time scheduling decisions on the subspace vector and generate a set of energy allocation strategies, wherein the optimization model maximizes energy utilization efficiency through a Markov decision process, including: The optimized subspace vector set is used as the state input for reinforcement learning. The state space is defined as the combination of real-time operating parameters of each energy link, and the action space is the executable scheduling operation. Each action corresponds to a specific parameter adjustment range, and the state-action space definition table is output. A multi-objective reward function is designed, with maximizing energy utilization efficiency as the primary objective and minimizing equipment operating costs and carbon emission intensity as secondary objectives. The weight allocation is determined by the analytic hierarchy process (AHP). Real-time state parameters are substituted into the function to calculate the instantaneous reward value, and the reward function calculation model is output. The DDPG algorithm is used to train the energy flow optimization model. The state-action space definition table and the reward function calculation model are used as inputs. The Actor-Critic network is updated iteratively through Markov decision process. An experience replay mechanism is introduced to store and sample effective decision samples, and a mature energy flow optimization model is output. The subspace vectors collected in real time are input into the trained and mature energy flow optimization model to generate multiple sets of energy allocation strategies. Infeasible strategies are eliminated by checking the equipment capacity constraints. The comprehensive benefit score of each strategy is calculated, and the set of energy allocation strategies sorted by score is output.

[0008] Optionally, the application of a digital twin verification algorithm to the energy allocation strategy set performs physical-digital spatial consistency verification to generate a reliable energy scheduling scheme. The digital twin verification algorithm ensures strategy feasibility through a bidirectional mapping mechanism, including: A full-element digital twin model is constructed based on the physical parameters of the energy system. Each strategy in the energy allocation strategy set is converted into a control command that the model can recognize, and the input command set of the twin model is output. The simulation engine of the digital twin model is started, and each instruction in the input instruction set is dynamically simulated. The changes of key parameters during the simulation process are recorded, and digital space operation curves are generated. At the same time, real-time physical operation data of the energy system is collected through IoT sensors to generate physical space operation curves and output a set of digital-physical operation curves. The dynamic time warping algorithm is used to calculate the deviation value of the corresponding parameter in the digital-physical operation curve set. When the deviation value exceeds the deviation threshold, the deviation source is located through a two-way mapping mechanism, the corresponding parameter in the twin model is corrected, and the simulation is repeated until the deviation value is lower than the deviation threshold. A deviation correction report is then output. Preliminary energy allocation strategies are selected based on the deviation correction report. The strategies are then ranked comprehensively based on their execution costs and energy utilization efficiency. The top 20% of the optimal strategies are selected, and implementation steps and emergency backup plans are added to generate a reliable energy dispatch plan.

[0009] Optionally, the implementation of blockchain notarization and smart contract execution for the trusted energy dispatch scheme to generate tamper-proof energy transaction records, serving as the final execution basis for the digital management of the energy system, includes: The trusted energy dispatch scheme is structured and parsed to extract key information including scheme number, execution time interval, list of energy equipment involved, and energy allocation of each stage. The SHA-256 hash algorithm is used to calculate the unique hash value of the scheme. The key information is bound to the hash value to generate scheme summary information, ensuring that the summary information can uniquely identify the scheme content. The scheme summary and hash value are output. The solution summary and hash value are uploaded to the Energy Consortium Chain, and a consensus request is initiated. Each node participates in the consensus by verifying the completeness of the solution summary and the correctness of the hash value. When more than 2 / 3 of the nodes confirm that it is valid, the notarization is completed. The on-chain record includes the notarization time, node signature and hash value, and outputs the blockchain notarization certificate. Smart contracts are written based on trusted energy dispatch schemes. The contract terms include trigger conditions, execution actions, and settlement rules. The contract is compiled into executable code through an on-chain virtual machine, deployed to the consortium blockchain and associated with the evidence storage certificate, and the deployed smart contract is output. Using smart contracts, trigger conditions are automatically determined based on real-time collected energy system operation data. When the conditions are met, the corresponding scheduling operation is executed. Each completed operation generates a transaction record containing the operation time, execution result, and equipment status. The hash value of the transaction record is written into the blockchain and linked with the evidence to form an immutable energy transaction record.

[0010] Another embodiment of this application provides an energy digital management system, the system comprising: The receiving module is used to receive multi-source heterogeneous energy data acquisition signals and generate an initial energy feature map based on a spatiotemporal correlation feature extraction algorithm, wherein each feature node is associated with the operating parameter dimension of the energy system. The encoding module is used to encode the initial energy feature map using a graph neural network, construct an energy knowledge graph vector set, and decompose the high-dimensional vectors in the knowledge graph vector set into several subspace vectors to capture the dynamic characteristics of energy flow. Each subspace vector represents the multi-dimensional operating state of the corresponding energy device. The scheduling module is used to make real-time scheduling decisions on the subspace vector using a reinforcement learning-driven energy flow optimization model, and generate a set of energy allocation strategies, wherein the optimization model maximizes energy utilization efficiency through a Markov decision process. The application module is used to apply a digital twin verification algorithm to the energy allocation strategy set to perform physical-digital space consistency verification and generate a reliable energy scheduling scheme. The digital twin verification algorithm ensures the feasibility of the strategy through a two-way mapping mechanism. The execution module is used to implement blockchain notarization and smart contract execution for the trusted energy dispatch scheme, generating tamper-proof energy transaction records as the final execution basis for the digital management of the energy system.

[0011] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0012] Another embodiment of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method described in any of the preceding claims.

[0013] Compared with existing technologies, this invention provides a digital energy management method that receives multi-source heterogeneous energy data acquisition signals and generates an initial energy feature map based on a spatiotemporal correlation feature extraction algorithm. The initial energy feature map is then encoded using a graph neural network to construct an energy knowledge graph vector set. The high-dimensional vectors in the knowledge graph vector set are decomposed into several subspace vectors. A reinforcement learning-driven energy flow optimization model is used to make real-time scheduling decisions on the subspace vectors, generating an energy allocation strategy set. A digital twin verification algorithm is applied to the energy allocation strategy set to perform physical-digital space consistency verification, generating a trusted energy scheduling scheme. Finally, the trusted energy scheduling scheme is stored on blockchain and executed via smart contracts, generating tamper-proof energy transaction records. This improves the scheduling accuracy and response speed of the energy system, achieving end-to-end trusted management from decision-making to execution. Attached Figure Description

[0014] Figure 1 Hardware structure block diagram of a computer terminal for an energy digital management method provided in an embodiment of the present invention; Figure 2 A flowchart illustrating a digital energy management method provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an energy digital management system provided in an embodiment of the present invention. Detailed Implementation

[0015] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0016] This invention first provides a digital energy management method, which can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0017] The following detailed explanation uses a computer terminal as an example. Figure 1 This is a hardware structure block diagram of a computer terminal for an energy digital management method provided in an embodiment of the present invention. (See diagram below.) Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0018] Non-volatile storage media can store operating systems and computer programs. These computer programs include program instructions that, when executed, cause the processor to perform any energy digital management method.

[0019] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0020] Internal memory provides an environment for the execution of computer programs in non-volatile storage media, which, when executed by a processor, enable the processor to perform any energy digital management method.

[0021] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 1 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.

[0022] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0023] See Figure 2 The present invention provides an energy digital management method, which may include the following steps: S201 receives multi-source heterogeneous energy data acquisition signals and generates an initial energy feature map based on a spatiotemporal correlation feature extraction algorithm, wherein each feature node is associated with the operating parameter dimension of the energy system; Specifically, it can receive multi-source heterogeneous energy data acquisition signals, add a unified timestamp to each signal through high-precision clock synchronization technology, generate spatiotemporal index tags by combining the GPS coordinates installed on the device, filter out invalid acquisition points with signal strength below the strength threshold, and output the original energy signal set with spatiotemporal index; The system first receives multi-source, heterogeneous energy data acquisition signals from energy production ends (such as photovoltaic inverters and wind turbines), transmission ends (such as smart meters and ring network cabinet monitoring units), and consumption ends (such as energy consumption sensors in industrial production lines and building air conditioning controllers) through a distributed data acquisition interface. These signal types include analog quantities (such as voltage, current, and temperature), digital quantities (such as device on / off status), pulse quantities (such as pulse counting in electricity meters), and structured messages (such as IEC 61850 protocol data). Due to millisecond-level clock deviations between different devices, the system employs high-precision clock synchronization technology—specifically using the IEEE 1588 Precision Time Protocol (PTP). A master clock source (usually a GPS / BeiDou dual-mode timing module) broadcasts time synchronization messages to all devices in the network, controlling the time error of each acquired signal to within 100 microseconds (μs). After synchronization, a unique timestamp (TS) is appended to each data entry, in the format "year-month-day-hour-minute-second.millisecond" (e.g., 20250411-143025.358). Simultaneously, the device's GPS coordinates (e.g., longitude 118.779°E, latitude 32.061°N) are read from the device's embedded positioning module or a backend database. The timestamp and GPS coordinates are then concatenated to form a spatio-temporal index tag (SIT), in the format "longitude, latitude@timestamp" (e.g., 118.779,32.061@20250411-143025.358). This tag serves as a unique spatio-temporal identifier for the data.

[0024] To ensure data validity, the system performs signal strength filtering on the raw signals with spatiotemporal indexes. Each signal is associated with an intensity threshold (IT) based on its physical characteristics: for example, the threshold for current signals is 0.5 amperes (A), below which is considered sensor drift error; the threshold for temperature signals is -50 degrees Celsius (°C), below which is determined to be a disconnection fault. The system calculates the signal amplitude (SA) in real time and compares it with the preset IT: if SA is less than IT, it is marked as an invalid data point (IDP) and discarded. For example, if the current signal of a photovoltaic string has SA=0.3A (less than IT=0.5A), the data packet corresponding to its spatiotemporal index label will be filtered. This step preserves valid data to form an Indexed Raw Energy Signal Set (IRS), which is stored as a Measurement Point (MP) in a time series database (such as InfluxDB). Each MP contains three core fields: the spatiotemporal index label SIT, the signal type (ST), and the raw value (RV).

[0025] To improve subsequent processing efficiency, the system compresses the IRS data. A lossy-lossless hybrid compression strategy is employed: Linear Interpolation Compression (LIC) is used for slowly changing parameters (such as ambient temperature), retaining a complete data point every 5 seconds; Wavelet Transform Compression (WTC) is used for high-frequency fluctuating parameters (such as power grid frequency), retaining 98% of the energy coefficient. After compression, the data volume is reduced by 60%, but accurate spatiotemporal correlation is still maintained through the spatiotemporal index label (SIT). The final output IRS dataset is sorted by the SIT, forming a spatiotemporally continuous data stream, laying the foundation for subsequent processing.

[0026] The original energy signal set with spatiotemporal index is preprocessed, and the signal components of different frequency bands are separated by wavelet packet decomposition algorithm. The noise components generated by electromagnetic interference are eliminated by adaptive threshold denoising. At the same time, the missing data is supplemented by interpolation method based on the similarity of adjacent time moments, and the denoised and completed energy data sequence is output. The preprocessing begins with decomposing the signal's frequency domain characteristics. Wavelet Packet Decomposition (WPD) is used to perform multi-scale analysis on each signal in the IRS: the Daubechies 4 (DB4) wavelet basis function is selected, with a decomposition depth of 8 levels (Level=8), splitting the original signal into 256 frequency band components (FBCs). For example, a current signal in the 0-1000 Hz range is decomposed into sub-band components such as 0-3.9 Hz, 3.9-7.8 Hz, ..., 992-1000 Hz. Each FBC contains a set of wavelet coefficients (WC) representing the energy distribution of that frequency band. This step separates the fundamental, harmonic, and interharmonic components in the signal, facilitating targeted processing.

[0027] Noise cancellation is implemented in two steps. First, noise components are identified: the energy of wavelet coefficients (EWC) of each FBC is statistically analyzed. If the EWC of a certain FBC is lower than 5% of the average energy across the entire frequency band (this percentage is an adaptive threshold), it is determined to be a noise component (NC). For example, the EWC of high-frequency bands (such as 750-1000Hz) is usually lower. Then, soft thresholding denoising (STD) is used: a threshold function is applied to the wavelet coefficients WC of the noise component NC. If |WC| < threshold T (T=σ√(2lnN), where σ is the noise standard deviation and N is the number of coefficients), then the WC value is set to zero; otherwise, the WC value is reduced. The processed wavelet coefficients are then reconstructed into a denoised signal. Meanwhile, for missing data (such as a 5-minute data gap caused by communication interruption), an interpolation method based on the similarity of adjacent moments is used: the dynamic time warping distance (DTWD) of 10 valid data points before and after the missing moment is calculated. If the DTWD is less than the similarity threshold (such as 0.1), linear weighted interpolation is used to fill it; otherwise, Long Short-Term Memory (LSTM) network prediction is used to fill it.

[0028] Reconstruction and Verification Phase: The denoised FBC wavelet coefficients are reconstructed into time-domain signals through the Inverse Wavelet Packet Transform (IWPT), and merged with the interpolated and completed data segments to form a Denoised and Completed Energy Data Sequence (DCES). To verify the processing effect, the system calculates two metrics: 1) Signal-to-Noise Ratio Improvement (SNRI), which is required to increase by at least 15 decibels (dB); 2) Reconstruction Error Rate (RER), which is required to be less than 0.5%. For example, if the SNRI of a certain voltage signal increases from 25 dB to 42 dB after processing and the RER = 0.3%, it is judged to be qualified. Finally, the DCES is stored according to the spatio-temporal index label SIT for use in the next step.

[0029] Based on the denoised and completed energy data sequence, extract the operational parameter dimensions of each energy device, calculate the mutual information values between parameters using the CFS feature selection algorithm, retain the core parameters strongly correlated with the energy system state, and generate a parameter-device association table; Extract the operational parameter dimension (OPD) of the energy device from the DCES. A set of key parameters is defined for each device: for example, a photovoltaic inverter includes 15 dimensions such as DC Voltage (Vdc), AC Power (Pac), and Efficiency (Eff); a transformer includes 12 dimensions such as Oil Temperature (OilTemp), Loading Rate (LR), and Total Harmonic Distortion (THD). The parameter dimensions are predefined through the Device Meta-Model (DMM), and the system automatically matches the corresponding OPD set according to the device type code (such as PV-Inverter-300kW). During extraction, the device number (Device ID, DID) is associated with the spatio-temporal index label SIT from the DCES, and the time-series values of all OPDs of the device are extracted to form a Device Parameter Matrix (DPM).

[0030] The Correlation-based Feature Selection (CFS) algorithm is used to screen core parameters. First, the mutual information value (MIV) between parameters is calculated: for any two parameter dimensions OPD_i and OPD_j, the MIV formula is ∑∑p(OPD_i,OPD_j)log[p(OPD_i,OPD_j) / p(OPD_i)p(OPD_j)]. In actual calculations, the probability distribution is estimated using a histogram method. For example, the MIV of transformer oil temperature (Toil) and load rate (LR) can reach 0.85 (with a maximum value of 1), while the MIV of Toil and ambient humidity (Hum) is only 0.2. The CFS algorithm defines the score for the parameter subset S: Merit_S = k·r_cf / √[k+k(k-1)r_ff], where k is the number of parameters in the subset, r_cf is the average correlation between the parameter and the system state (such as the "fault" label), and r_ff is the average mutual information between parameters. By using a forward search strategy, parameters that maximize Merit_S are added incrementally.

[0031] Parameter filtering and association table generation: A strong correlation criterion is set as Merit_S > 0.7, and the MIV of a single parameter and system state > 0.6. For example, in photovoltaic inverters, Pac (MIV=0.92) and Eff (MIV=0.88) are retained, while secondary parameters such as chassis temperature (MIV=0.35) are removed. Finally, a Parameter-DeviceAssociation Table (PDAT) is generated for each type of equipment, with a four-tuple structure (equipment type code, parameter dimension name, mutual information value MIV, whether it is a core parameter flag). For example, an entry ("WTG-2MW", "gearbox vibration value", 0.91, yes) indicates that the parameter is retained. This table is stored in a relational database (such as MySQL) as the metadata foundation for subsequent graph construction.

[0032] Based on the parameter-equipment association table, a spatiotemporal association rule mining algorithm is used to calculate the co-occurrence frequency and association strength of different equipment parameters. The equipment parameters are used as feature nodes and the association strength is used as edge weights to construct an initial energy feature map containing node attributes and edge attributes.

[0033] The map construction begins with the mining of spatiotemporal association rules. Based on the core parameters in PDAT, parameter time series are extracted from DCES and grouped by spatiotemporal index label SIT (the same geographical area at the same time is grouped together). For each group of data, the improved Apriori algorithm is used to calculate the co-occurrence frequency (CF): if parameter A and parameter B both exceed the normal threshold range within a time window Δt = 5 minutes (e.g., A > upper limit, B < lower limit), it is recorded as a co-occurrence event. The number of co-occurrences within all groups, N_AB, is counted, then CF = N_AB / total number of time groups. For example, in a certain wind farm, the CF of "gearbox vibration > 7 mm / s" and "generator temperature > 120℃" is 0.15 (i.e., co-occurring within 15% of the time period).

[0034] The association strength (AS) is calculated as the edge weight. A symmetric uncertainty (SU) index is used: SU = 2 × MIV / [H(OPD_i) + H(OPD_j)], where H is the information entropy. The SU value range is [0,1], with larger values ​​indicating stronger causal relationships between parameters. For example, the SU for transformer "oil temperature > 85℃" and "load rate > 90%" is 0.82, while the SU for "oil temperature" and "ambient temperature" is 0.45. Spatiotemporal constraints are defined: the association strength is calculated only when the spatial distance between two parameters is less than 1 kilometer (within the same substation) and the time lag is less than 10 seconds.

[0035] Constructing the Initial Energy Feature Graph (IEFG): Node: Each core parameter is a feature node, and its attributes include parameter type (e.g., "current"), unit (e.g., "ampere"), and device affiliation (e.g., "inverter #07"). Edge: If the SU of parameters A and B > 0.3 (association strength threshold), then create an undirected edge with weight W = SU value and attributes including co-occurrence frequency CF and spatiotemporal constraints. Visualization: The graph is presented using a force-oriented layout algorithm (Force-Atlas 2), with strongly correlated nodes automatically clustered. For example, photovoltaic parameters form a cluster (Pac, Vdc, Eff are interconnected), while distribution parameters form another cluster (voltage, current, power factor are interconnected). The final IEFG is stored as a set of nodes and relationships in a graph database (such as Neo4j) for subsequent processing by a graph neural network.

[0036] By integrating energy data from various devices and systems, including real-time monitoring data of multiple energy forms such as electricity, gas, and heat, and employing spatiotemporal correlation algorithms, scattered data points are transformed into topologically structured feature maps. Each node represents an energy device or parameter, and the connections between nodes reflect energy flow paths and dependencies. Data synchronization technology ensures that signals from different sources are aligned in time and space, resolving the heterogeneity problem of multi-source energy data and providing structured input for subsequent analysis. The spatiotemporal feature maps can intuitively display the overall operating status of the energy system, helping to identify key parameter nodes and energy flow bottlenecks, laying a data foundation for optimization decisions.

[0037] S202, the initial energy feature map is encoded by a graph neural network to construct an energy knowledge graph vector set. The high-dimensional vector in the knowledge graph vector set is decomposed into several subspace vectors to capture the dynamic characteristics of energy flow. Each subspace vector represents the multi-dimensional operating state of the corresponding energy device. Specifically, the importance of nodes in the initial energy feature map can be evaluated. The PageRank algorithm is used to calculate the importance score of each feature node. The top 90% of nodes are retained to simplify the map structure. Weakly related edges with weights below the weight threshold are deleted, and a simplified energy feature map is output. Node Importance Evaluation and Application of PageRank Algorithm The initial energy feature graph consists of thousands of feature nodes (such as photovoltaic inverter output power, energy storage battery SOC, and transformer winding temperature) and edges (representing the correlation strength between parameters). To identify key nodes, the PageRank algorithm (a graph-based method for quantifying node influence) is used. The algorithm simulates the transmission process of "virtual energy flow" in the graph: each node distributes its importance score (initially 1.0) to its neighboring nodes according to the weight of its outgoing edges, with the distribution formula calculated iteratively. After 20 iterations, the node scores tend to stabilize. For example, the voltage stability node of a regional power grid hub substation scores 0.85 (out of 1.0), while the edge lighting energy consumption node scores only 0.03. An importance score threshold is set at the top 90% of the nodes (if there are 1000 nodes in total, nodes with scores ≥ 100 are retained). Nodes with scores below this threshold (such as secondary office temperature control parameters) are considered redundant and deleted.

[0038] Weakly related edge filtering and graph simplification The edge weight threshold is set to 0.15 (an empirical value), determined through analysis of historical operation and maintenance data. For example, the correlation weight between transformer oil temperature and ambient humidity is only 0.08 (weak correlation), while the weight model for photovoltaic array output and irradiance sensor is 0.92 (strong correlation). The system automatically scans all edges and deletes edges with a weight < 0.15 (e.g., deleting the "air conditioner start / stop - elevator energy consumption" edge). Simultaneously, topology optimization is performed: if a node becomes isolated due to edge deletion (e.g., a backup generator node has no effective connection), it is removed. The final graph node count is reduced to 60-70% of its original size, and the edge count is reduced by 40%, forming a simplified energy feature graph focusing on core energy flow paths.

[0039] Dynamic weight threshold adjustment mechanism The weighted threshold of 0.15 is not a fixed value. When a peak load period is detected (such as a summer afternoon), the system automatically raises the threshold to 0.25 to further compress the graph, retaining only strongly correlated edges to handle emergency dispatch; during periods of stable load, it reverts to 0.15. This mechanism is triggered by data from the Real-time Load Rate Sensor (RLRS), activating the enhanced filtering mode when the load rate is >85%.

[0040] The simplified energy feature map is input into the graph attention neural network. The dependencies between nodes are learned through the multi-head attention mechanism. The attribute features and associated features of each node are fused and encoded into a 128-dimensional high-dimensional vector. Each dimension of the vector corresponds to the abstract expression of the feature, and the energy knowledge graph vector set is output. Graph Attention Neural Network Architecture Design A Multi-head Graph Attention Network (GAT) architecture is adopted, comprising eight parallel attention heads. Each attention head independently calculates the dependency strength between nodes: for the target node (e.g., the #32 energy storage battery), its attention coefficient with neighboring nodes (e.g., the #15 photovoltaic array and the #67 grid interface) is calculated. This coefficient is jointly determined by a three-layer perceptron (node ​​attribute similarity, edge weights, and topological distance). For example, when the photovoltaic output suddenly increases, its attention coefficient with the energy storage node jumps from 0.3 to 0.9, representing a real-time increase in dependency.

[0041] 128-dimensional high-dimensional vector encoding process The original attributes of each node (such as 12-dimensional parameters like voltage, current, and temperature) are concatenated with the associated features (8-dimensional) output from the 8 attention heads to form a 20-dimensional intermediate vector. This vector is then used for feature abstraction via a three-layer fully connected neural network (layer widths 256-128-128). The first layer extracts device-level state features (such as encoding current fluctuation patterns as "harmonic distortion index"). The second layer integrates cross-device coupling features (such as encoding "photovoltaic-energy storage charge-discharge synergy efficiency"). The third layer outputs a 128-dimensional vector, with each dimension corresponding to an abstract feature label (e.g., dimension 37 represents "feeder voltage stability margin" and dimension 89 represents "transformer thermal aging risk").

[0042] Energy knowledge graph vector set generation The 128-dimensional vectors of all nodes constitute the vector set of the energy knowledge graph. This set is stored in matrix form (rows = device nodes, columns = feature dimensions), for example: row 5 (photovoltaic array node): [0.87, 0.12, ..., 0.45] (128 dimensions); row 32 (energy storage node): [0.02, 0.95, ..., 0.31] (128 dimensions); the cosine similarity between vectors directly reflects the correlation of device status (e.g., the similarity between photovoltaic and energy storage vectors is >0.7, while that with lighting systems is <0.1).

[0043] Based on the dynamic characteristics of energy flow, the kernel principal component analysis algorithm is used to decompose the high-dimensional vectors in the energy knowledge graph vector set into subspaces. The decomposition dimension is set according to the energy transfer efficiency of each link, and a preliminary subspace vector set is output. Dynamic Characteristic Analysis and Kernel Function Selection The dynamic characteristics of energy flow are manifested in three modes: transmission loss characteristics (such as voltage attenuation caused by cable resistance); conversion efficiency characteristics (such as the AC / DC conversion efficiency curve of the inverter); and time delay characteristics (such as the heat transfer delay in heating network pipelines).

[0044] To address the nonlinear characteristics, kernel principal component analysis (KPCA) using the radial basis function kernel (RBFKernel) was selected. The kernel parameter σ was set to 0.5 and determined through cross-validation to accommodate the efficiency fluctuation range of most energy devices.

[0045] Dimensional allocation driven by energy transfer efficiency The decomposition dimensions are set based on energy transfer efficiency (ETE): High-efficiency components (ETE > 90%, such as transformers): 2 dimensions are allocated (to capture the characteristics of high-efficiency steady-state operation); Medium-efficiency components (75% ≤ ETE ≤ 90%, such as gas turbines): 3 dimensions are allocated (efficiency fluctuations need to be monitored); Inefficient components (ETE < 75%, such as old water pumps): 5 dimensions are allocated (focus on analyzing the root causes of losses).

[0046] For example, the subspace decomposition of a photovoltaic power station is as follows: photovoltaic panel (ETE=92%) → 2-dimensional vector (irradiance utilization efficiency, panel temperature influence coefficient); inverter (ETE=85%) → 3-dimensional vector (conversion loss, harmonic components, load response speed).

[0047] Kernel PCA Decomposition and Vector Generation Input the 128-dimensional vector into the kernel PCA model: Calculate the kernel matrix: compare the similarity of all vector pairs based on the RBF kernel; solve the characteristic equation: extract the first k principal components (k is determined by the dimension allocation rule); project to subspace: transform the original vectors into k-dimensional subspace vectors.

[0048] Output a preliminary subspace vector set (e.g., the node vector of a transformer is compressed from 128 dimensions to a 3-dimensional vector [1.2, -0.3, 0.8] representing "copper loss-iron loss-heat dissipation efficiency").

[0049] The initial subspace vector set is processed by the Gram-Schmidt orthogonalization process to eliminate the coupling correlation between different vectors, so that each subspace vector independently represents the operating state of a single energy link, and the optimized subspace vector set is output.

[0050] Location of Coupling and Correlation Problems The initial subspace vectors exhibit parameter coupling (e.g., a vector simultaneously contains "generator speed" and "coolant temperature," which are strongly correlated under fault modes). Covariance matrix analysis is used to detect the degree of coupling: if the absolute value of the covariance between the two dimensions is >0.7 (e.g., the covariance between speed and coolant temperature = -0.82), it is determined to be strongly coupled and requires decoupling.

[0051] Gram-Schmidt Orthogonalization Execution Vectors are grouped by energy source (e.g., devices on the same feeder are grouped together): Step 1: Select a basis vector (e.g., select photovoltaic vector V1=[0.9, 0.2]); Step 2: Calculate the next vector projection (e.g., the projection component of the energy storage vector V2=[0.4, 0.6] onto V1: Proj=(V2·V1) / ||V1||). 2 × V1); Step 3: Generate orthogonal vectors V2' = V2 - Proj (e.g., obtaining [-0.12, 0.56]). Iterate layer by layer until all vectors within the group are orthogonal (inner product < 10). -6 ).

[0052] Independent state representation optimization After orthogonalization, each dimension represents only a single physical meaning: Dimension 1 (photovoltaics): pure irradiance response coefficient (decoupled from energy storage); Dimension 2 (energy storage): pure charge / discharge efficiency (decoupled from grid frequency); the final output is an optimized subspace vector set, used as input for reinforcement learning. For example, the decoupling vector received by a microgrid controller: [Photovoltaic vector: 0.82, 0, 0]; [Storage vector: 0, 0.74, 0]; [Diesel engine vector: 0, 0, -0.15].

[0053] Zero values ​​indicate that dimensions are completely independent, ensuring that scheduling decisions accurately correspond to the specific device status.

[0054] Graph neural networks are used to deeply encode feature maps, transforming complex energy network relationships into high-dimensional vector representations. Through subspace decomposition techniques, the overall energy system is broken down into multiple interconnected subsystems. Each subspace vector focuses on the operational characteristics of specific energy equipment, such as dynamic parameters like load rate and efficiency curves. This enables modular analysis of the energy system, allowing for both a grasp of the overall operational status and a deep understanding of the operational details of individual devices. Vectorized representation facilitates computer processing and provides standardized input for intelligent decision-making.

[0055] S203, using a reinforcement learning-driven energy flow optimization model to make real-time scheduling decisions on the subspace vector, generating a set of energy allocation strategies, wherein the optimization model maximizes energy utilization efficiency through a Markov decision process; Specifically, the optimized subspace vector set can be used as the state input for reinforcement learning. The state space is defined as the combination of real-time operating parameters of each energy link, and the action space is the executable scheduling operation. Each action corresponds to a specific parameter adjustment range, and the state-action space definition table is output. The optimized subspace vector set (each vector independently representing the operating state of a single energy link) is input into the reinforcement learning system in real time. The state space is constructed based on these vectors: each subspace vector contains multi-dimensional operating parameters of a specific energy link (such as photovoltaic arrays, energy storage battery packs, and gas turbines). For example, the subspace vector of a photovoltaic array may contain three core parameters: real-time power generation (kW), module temperature (°C), and irradiance (watts per square meter); the vector of an energy storage battery pack contains parameters such as state of charge (SOC, percentage), charging and discharging current (Amperes), and internal temperature (°C). The state space is an ordered combination of all these link parameters, forming a high-dimensional vector. The system sets a reasonable value range for each parameter (e.g., photovoltaic power 0-5000 kW, SOC 20%-95%) to ensure that the state space covers all possible operating conditions while avoiding invalid regions. Simultaneously, the action space is defined as the set of specific scheduling operation instructions that the energy management system can execute. Each action corresponds to the adjustment of one or more equipment parameters. For example, Action A1: Increase the charging power of the energy storage battery by 100 kW; Action A2: Reduce the output of the gas turbine by 50 kW and start the standby diesel generator; Action A3: Sell 200 kW of excess photovoltaic power to the grid. Each action is associated with a specific, quantified range of parameter adjustments (such as a charging power adjustment step of 50-500 kW and an output adjustment accuracy of 1 kW) to ensure that the instructions can be executed accurately by the equipment.

[0056] The key to defining the state-action space lies in precise mapping and constraint setting. The system maintains a parameter-operation mapping dictionary, explicitly recording which actions (such as energy storage charging and electricity sales) can affect each state parameter (e.g., photovoltaic power), as well as the strength and direction of the influence (increase / decrease). More importantly, physical constraints are strictly embedded when defining the action space. For example, the power limit of the action of charging the battery is limited by the battery's current maximum allowed charging current (C-rate) and the converter capacity; the action of increasing the gas turbine output is limited by its ramp rate (e.g., a maximum increase of 5% of rated power per minute) and maximum output limit. These constraints are encoded as hard boundary conditions for action selection, ensuring that any generated scheduling instructions are physically executable and will not cause equipment overload or damage.

[0057] The final output State-Action Space Definition Table is a structured document or database. It details: 1) A list of state dimensions: categorized by energy stage, clearly describing the name (e.g., "PV_Real-time Power"), data type (floating-point), unit (kW), minimum value, maximum value, and normal range for each state parameter; 2) A list of actions: a unique code for each action (e.g., ACT_101), description (e.g., "Increase Energy Storage Charging Power"), affected parameters (e.g., "Energy Storage_Charging Power"), adjustment amount (e.g., +Delta kW, Delta range 50-200), and associated physical device (e.g., "Energy Storage Converter No. 1"); 3) Constraint rules: clearly defining the boundary conditions for each action (e.g., when executing ACT_101, the new value of "Energy Storage_Charging Power" <= Min(Energy Storage_Maximum Allowable Charging Power, Converter_Rated Power)). This table forms the foundational framework for subsequent reinforcement learning algorithms, ensuring that the algorithm explores the optimal strategy within a legal and safe range.

[0058] A multi-objective reward function is designed, with maximizing energy utilization efficiency as the primary objective and minimizing equipment operating costs and carbon emission intensity as secondary objectives. The weight allocation is determined by the analytic hierarchy process (AHP). Real-time state parameters are substituted into the function to calculate the instantaneous reward value, and the reward function calculation model is output. The reward function is the core driving force for the optimization of reinforcement learning models. This method designs a multi-objective reward function that simultaneously optimizes three key objectives: the primary objective is to maximize energy utilization efficiency, defined as the ratio of the system's effective output energy (supplying loads or storage) to total input energy (including fuel, purchased electricity, and renewable energy); the secondary objectives include minimizing operational costs (covering fuel costs, maintenance costs, and the price difference between the grid and retail prices) and minimizing carbon emission intensity (the carbon dioxide equivalent emissions per unit of energy output). To integrate these three objectives, which have different dimensions and varying importance, into a quantifiable immediate reward value, the system employs the Analytic Hierarchy Process (AHP). The AHP process invites energy experts to conduct pairwise importance comparisons of the three objectives (e.g., "How much more important is efficiency improvement than cost savings?", "How much more important is carbon emission reduction than cost savings?"), constructing a judgment matrix, and finally determining the weight coefficients for each objective through eigenvector calculation. For example, the final weight allocation might be: energy efficiency weight Weff = 0.6, operating cost weight Wcost = 0.25, and carbon emission weight Wcarbon = 0.15. This reflects a management strategy that prioritizes efficient energy use.

[0059] The reward function transforms each objective into a standardized benefit or penalty term. 1) Energy Efficiency (Reff): Calculates the current system's overall energy efficiency η (e.g., η=82%) and compares it to a baseline energy efficiency η_base (e.g., historical average 80%). If η > η_base, a positive benefit is obtained (the benefit value is proportional to the improvement); otherwise, a penalty is imposed. 2) Operating Cost (Rcost): Calculates the current total operating cost C (unit: yuan) and compares it to the baseline cost C_base. If C < C_base, a positive benefit is obtained (the benefit value is inversely proportional to the cost savings); otherwise, a penalty is imposed. 3) Carbon Emission (Rcarbon): Calculates the current carbon emission intensity I (unit: kgCO2e / kWh) and compares it to the baseline intensity I_base. If I < I_base, a positive benefit is obtained; otherwise, a penalty is imposed. The final instantaneous reward value Rt is calculated as: Rt = Weff * Reff + Wcost * Rcost + Wcarbon * Rcarbon. All baseline values ​​(η_base, C_base, I_base) are typically dynamically updated rolling averages to ensure that the reward function adapts to the long-term operating trend of the system.

[0060] The output reward function calculation model is a configurable software module. It receives real-time state parameters (such as photovoltaic power generation, gas turbine gas consumption, grid electricity purchases, load demand, equipment operating hours, real-time electricity price, and fuel carbon intensity coefficient) as input. Internally, the module: 1) calculates the current energy efficiency η, total operating cost C, and carbon emission intensity I according to predefined formulas; 2) retrieves or calculates the current dynamic baseline values ​​η_base, C_base, and I_base from the database; 3) calculates Reff, Rcost, and Rcarbon based on the configured weight coefficients Weff, Wcost, and Wcarbon and the preset reward / penalty calculation rules; 4) performs a weighted summation to obtain the final immediate reward value Rt. The key parameters of this model (weights, baseline update frequency, and reward / penalty coefficients) can be adjusted through the management interface to adapt to different strategy preferences in different scenarios (such as extreme focus on cost or low carbon). Each time a state transition occurs (the system enters a new state after an action is performed), the model is called to calculate Rt, guiding the learning direction of the reinforcement learning model.

[0061] The DDPG algorithm is used to train the energy flow optimization model. The state-action space definition table and the reward function calculation model are used as inputs. The Actor-Critic network is updated iteratively through Markov decision process. An experience replay mechanism is introduced to store and sample effective decision samples, and a mature energy flow optimization model is output. The Deep Deterministic Policy Gradient (DDPG) algorithm was chosen as the core for training the energy flow optimization model. DDPG is a model-free reinforcement learning algorithm that combines deep neural networks (DNNs) and deterministic policy gradients, making it particularly suitable for problems involving continuous states and continuous action spaces (such as continuous power regulation in energy scheduling). The training process strictly follows the Markov Decision Process (MDP) framework: at each discrete time step t, the agent observes the current state St (i.e., the state represented by the subspace vector set), selects an action At (scheduling instruction) according to its policy, and the environment (i.e., the energy system) executes At, transitions to the new state St+1, and returns an immediate reward Rt (calculated by the reward function model). The core of DDPG is maintaining two key neural networks: the Actor Network and the Critic Network, as well as copies of their target networks (Actor_Target, Critic_Target). The Actor network is responsible for learning the policy. Its input is the state St, and its output is a definite action At (e.g., "increase charging power by 150.3 kW"). The Critic network is responsible for evaluating the value Q(St, At) of the state-action pair, which is the expected cumulative reward.

[0062] The core of the training iteration process is experience replay and network updates. The system maintains a replay buffer, which is a database storing historical experience tuples, each containing (St, At, Rt, St+1). As the agent explores the environment (e.g., performing random actions or noisy actor actions), these experiences are continuously added to the buffer. The buffer typically has a fixed capacity (e.g., 1 million entries) and is updated using a first-in, first-out (FIFO) strategy. During each training iteration, a small batch (e.g., 128 entries) of experience samples is randomly sampled from the buffer. This breaks the temporal correlation of the data and improves learning stability. The network is updated using the sampled samples: 1) Update the Critic network: Calculate the target Q-value for each sample: y = Rt + γ * Q_Target(St+1, Actor_Target(St+1)), where γ is the discount factor (e.g., 0.95), representing the importance of future rewards. Then, minimize the mean squared error between the Critic network's predicted value Q(St, At) and the target value y. 2) Update the Actor network: Using the gradient information provided by the Critic network, update the Actor network parameters through policy gradient ascent, with the goal of maximizing the Q-value Q(St, Actor(St)) evaluated by the Critic network. 3) Softly update the target network: Slowly update the parameters of the Actor and Critic networks to their target network (e.g., update 0.1% per iteration) to ensure learning stability.

[0063] Training is typically conducted offline in environments such as high-precision energy system simulation platforms, using historical running data or large amounts of simulated scenario data. Key training parameters include the learning rate (e.g., 0.001 for the Critic network, 0.0001 for the Actor network), discount factor γ, the magnitude of exploration noise (e.g., the Ornstein-Uhlenbeck noise process), the target network update rate (τ=0.001), the replay buffer size, and the batch size. Training continues until the model performance converges and stabilizes on the validation set (e.g., the average cumulative reward no longer significantly increases over N consecutive validation cycles). The final output, a mature trained energy flow optimization model, contains stable and high-performance Actor network parameters. This Actor network is the module used for real-time decision-making during actual deployment: it takes the current state St as input and outputs the optimal action At (scheduling instruction).

[0064] The subspace vectors collected in real time are input into the trained and mature energy flow optimization model to generate multiple sets of energy allocation strategies. Infeasible strategies are eliminated by checking the equipment capacity constraints. The comprehensive benefit score of each strategy is calculated, and the set of energy allocation strategies sorted by score is output.

[0065] During actual operation, the real-time data acquisition system continuously acquires the latest operating data of each energy device. After feature extraction and subspace decomposition, it generates the current optimized subspace vector set. These vectors are input in real time to the deployed and well-trained energy flow optimization model (i.e., the Actor network of DDPG). The internal processing flow of the model is as follows: 1) The network receives the state vector St (dimension consistent with that during training); 2) Forward propagation is performed through each layer of the network (such as fully connected layers and the ReLU activation function); 3) A continuous and deterministic action vector At is generated at the output layer. This At vector directly corresponds to the scheduling operation combination defined in the state-action space definition table (e.g., [energy storage charging power increment = +120.5kW, photovoltaic power curtailment = 0kW, gas turbine setpoint = 850kW, electricity purchase = -50kW (representing electricity sales)]). This is a basic energy allocation strategy generated by the model.

[0066] To provide decision-making flexibility and robustness, the system does not rely solely on a single model output. It generates multiple energy allocation strategies by introducing policy perturbations or parallel inference. Common methods include: 1) Noisy exploration: A small, controllable random noise (such as Gaussian noise) is superimposed on the output action At of the Actor network to generate several (e.g., 5-10) alternative strategies that fluctuate around the optimal strategy. 2) Post-processing of multi-objective optimization: Using the value assessment Q(St, A) of the Critic network, multiple candidate action points are sampled near the optimal action At, their Q values ​​and various sub-rewards (efficiency, cost, carbon emissions) are calculated, and several representative strategies distributed on the Pareto front are selected. Each strategy specifically describes the setpoint or operational instructions that each controllable device should execute within a future scheduling cycle (e.g., the next 15 minutes). However, these strategies may violate actual physical constraints. Therefore, the system performs rigorous Equipment Capacity Constraint Checking: each strategy is sent to a Constraint Validator. The validator checks the feasibility of each action instruction in the strategy item by item, based on the constraint rules in the state-action space definition table, the real-time status of the equipment (such as current SOC and temperature), and the equipment nameplate parameters (such as maximum power, maximum current, and ramp rate). For example, it checks whether "energy storage charging power +120.5kW" exceeds the current battery's maximum allowable charging power (affected by SOC and temperature) or the converter's rated power. Any action or combination of actions that violates the constraints will cause the entire strategy to be marked as an infeasible strategy and discarded.

[0067] For a feasible strategy set that passes the constraint check, the system needs to evaluate its comprehensive benefits to assist decision-makers in making a choice. This is achieved by calculating the Comprehensive Benefit Score (CBS) for each strategy. The scoring model considers not only the expected immediate reward Rt after the strategy is implemented (predicted by the reward function model), but also longer-term benefit assessments (approximated using the Q-value of the Critic network) and factors such as the strategy's robustness and execution complexity. Typically, the CBS is a weighted combination: CBS = W1 * Q(St, A) + W2 * Strategy Execution Complexity Score + W3 * Strategy Robustness Score. Here, Q(St, A) is the Critic network's long-term value assessment of taking the strategy (action A); a lower strategy execution complexity score (e.g., frequency of instruction changes, number of equipment start-ups and shutdowns) is better; and a higher strategy robustness score (e.g., sensitivity to prediction errors) is better. Weights W1, W2, and W3 are configured according to management needs (e.g., W1=0.8, W2=0.1, W3=0.1). All feasible strategies are sorted from highest to lowest based on their calculated CBS values. The final output Energy Allocation Strategy Collection is this list of feasible strategies sorted in descending order of their comprehensive benefit score. Each strategy record contains a detailed sequence of scheduling instructions, expected changes in key performance indicators (efficiency, cost, carbon emissions), the calculated CBS value, and constraint check passing flags. This collection provides decision-makers with multiple optimized and feasible scheduling options, with the highest-scoring strategy typically being the preferred implementation plan.

[0068] A dynamic optimization model based on a reinforcement learning framework transforms the energy scheduling problem into a sequential decision-making process. The model learns optimal strategies through continuous interaction with the environment, prioritizing the energy allocation scheme with the highest energy efficiency and lowest cost while ensuring supply and demand balance. Markov decision processes ensure the temporal consistency of the strategies, achieving adaptive optimization of the energy system and significantly improving scheduling efficiency. Compared to traditional rule engines, reinforcement learning can automatically adapt to system changes and discover optimal strategies that are difficult to design manually.

[0069] S204, Apply the digital twin verification algorithm to the energy allocation strategy set to perform physical-digital space consistency verification and generate a reliable energy scheduling scheme. The digital twin verification algorithm ensures the feasibility of the strategy through a two-way mapping mechanism. Specifically, a full-element digital twin model can be constructed based on the physical parameters of the energy system, and each strategy in the energy allocation strategy set can be converted into a control command that the model can recognize, outputting the input command set of the twin model; First, the construction of a full-element digital twin model requires the integration of multi-dimensional physical parameters of the energy system. For example, for thermal power generating units, real-time data such as boiler combustion efficiency (unit: percentage %), turbine speed (unit: revolutions per minute RPM), and generator output power (unit: megawatts MW) need to be collected; for photovoltaic power plants, irradiance (unit: watts per square meter W / m²) needs to be integrated. 2 Parameters such as component temperature (unit: degrees Celsius °C) and inverter conversion efficiency (unit: percentage %) are used. These parameters are mathematically correlated through physical equations (such as the first law of thermodynamics and Ohm's law) and equipment characteristic curves (such as the flow-head curve of a pump and the charge-discharge rate-C-rate curve of a battery), forming a dynamically simulated virtual system. The model accuracy must meet engineering-grade standards; for example, the simulation error of the steam turbine thermal efficiency should not exceed ±0.5%.

[0070] Secondly, the strategy conversion process requires parsing the structured instructions in the energy allocation strategy set. Suppose a strategy requires "increasing the discharge power of the energy storage system to 80% of its rated value during the period from 14:00 to 15:00," the conversion module will break it down as follows: Timestamp instructions: Start timestamp Unix Time 1715673600 (corresponding to May 15, 2024, 14:00), end timestamp 1715677200 (15:00); Device action instructions: Energy storage system output power target value = rated power × 80% (if the rated power is 100MW, then the target value is 80MW); Constraints: Battery state of charge (SOC) must not be lower than 20%.

[0071] The converted instructions are in JSON or XML format and include fields such as {"device_id": "ESS_Unit01", "action_type": "discharge", "target_value": 80, "unit": "MW", "constraints":["SOC≥20%"]} to ensure that the twin model can be recognized.

[0072] Finally, instruction set generation needs to be compatible with multiple brands of equipment through a protocol adaptation layer. For example, Siemens gas turbines use the Modbus TCP protocol, while Tesla energy storage systems use the CAN bus protocol. The adaptation layer converts the unified instructions into device-specific control codes (such as Modbus function code 06 for writing registers) and adds check bits (such as CRC-16 cyclic redundancy check). The output instruction set is sorted chronologically to form a queue with priority tags, ensuring that high-priority instructions (such as emergency peak shaving) are executed first.

[0073] The simulation engine of the digital twin model is started, and each instruction in the input instruction set is dynamically simulated. The changes of key parameters during the simulation process are recorded, and digital space operation curves are generated. At the same time, real-time physical operation data of the energy system is collected through IoT sensors to generate physical space operation curves and output a set of digital-physical operation curves. The dynamic simulation engine operates based on a physical mechanism model. Taking power grid load dispatch as an example: Digital space simulation: When the command "increase wind power output by 30MW" is input, the engine calls the power characteristic model of the wind turbine (including the wind speed-power conversion function), and combines it with meteorological forecast data to simulate changes in intermediate parameters such as rotor speed (unit: revolutions per minute, RPM) and gearbox torque (unit: Newton-meters, Nm), and finally outputs the predicted grid frequency (unit: Hertz, Hz) and node voltage (unit: kilovolts, kV). The simulation step size is set to milliseconds (e.g., 0.1 seconds) to ensure the capture of transient processes (such as short-circuit current impacts).

[0074] Curve generation: Key parameters (such as frequency deviation Δf) are stored in time series with a sampling interval of 0.5 seconds to form a digital space running curve. The data format is {timestamp: 1715673600.5, parameter: "grid_frequency", value: 49.98}.

[0075] Physical space data acquisition relies on Internet of Things (IoT) sensor networks: Sensor types include: Electrical quantities: smart meters (accuracy class 0.2S) to collect voltage / current harmonic distortion rate (THD) (unit: percentage %); Mechanical quantities: vibration sensors (range 0-10g) to monitor bearing abnormalities; Environmental quantities: temperature and humidity sensors (accuracy ±0.5℃) to record the status of the computer room.

[0076] Data transmission adopts industrial protocols (such as IEC 61850), and is preprocessed through an edge computing gateway (such as 5-minute moving average filtering) to generate physical space operation curves. The timestamps are strictly aligned with the digital space (synchronized using the NTP network time protocol with an error of <1 millisecond).

[0077] The curve set output adopts a dual-channel storage structure: digital curves are stored in the time-series database InfluxDB, with tags including simulation_id=SIM_2024051501; physical curves are stored in the relational database MySQL, with tags including device_id=WTG_Unit07. The two are linked through a unified transaction ID to form comparable data pairs.

[0078] The dynamic time warping algorithm is used to calculate the deviation value of the corresponding parameter in the digital-physical operation curve set. When the deviation value exceeds the deviation threshold, the deviation source is located through a two-way mapping mechanism, the corresponding parameter in the twin model is corrected, and the simulation is repeated until the deviation value is lower than the deviation threshold. A deviation correction report is then output. Deviation calculation uses the Dynamic Time Warping (DTW) algorithm: Taking the power grid frequency curve as an example, we take the digital spatial sequence S_{sim} = [f_{sim1}, f_{sim2}, ...,f_{simn}] and the physical spatial sequence S_{phy} = [f_{phy1}, f_{phy2}, ..., f_{phym}] (n and m are the number of time points of different lengths, f_{simn} represents the nth frequency value of the digital spatial sequence, and f_{phym} represents the mth frequency value of the physical spatial sequence).

[0079] Construct a cumulative distance matrix D, where the matrix element d(i,j) represents the Euclidean distance (in Hertz Hz) between S_{sim}[i] and S_{phy}[j].

[0080] The minimum warping path is found using dynamic programming, and the minimum distance d_{min} (in millihertz, mHz) between the two normalized sequences is calculated as the deviation value. The deviation threshold δ_{thres} is set to 50 mHz (according to the IEEE 1547 standard).

[0081] Two-way mapping mechanism to locate the source of deviation: Forward mapping (physical to digital): When d_{min} > δ_{thres} is detected, source tracing analysis is triggered. For example, the physical sensor shows that the transformer oil temperature is 85℃, while the simulation value is 78℃. The forward mapping engine retrieves the heat loss equation of the transformer in the twin model and finds that the cooling coefficient K is set to 0.75, while the actual equipment needs to be corrected to 0.68 due to aging.

[0082] Reverse mapping (digital → physical): If the simulation predicts that the current of a certain line exceeds the limit (>1200A), but the physical sensor does not alarm, the reverse mapping checks the sensor calibration record and finds that it has a +5% range error, triggering on-site calibration.

[0083] Model correction and resimulation: Correct parameters: Update the transformer cooling coefficient K from 0.75 to 0.68; Rerun simulation: Generate new numerical curves using the corrected model; Iteration termination condition: When d_{min} ≤ δ_{thres} or the maximum number of iterations (e.g., 3 times) is reached.

[0084] For example, the deviation correction report includes: deviation parameter name (e.g., "transformer T1 oil temperature"); original model value / physical value (78℃ / 85℃); corrected model value (83.2℃); final deviation value (d_{min}=12 mHz).

[0085] Preliminary energy allocation strategies are selected based on the deviation correction report. The strategies are then ranked comprehensively based on their execution costs and energy utilization efficiency. The top 20% of the optimal strategies are selected, and implementation steps and emergency backup plans are added to generate a reliable energy dispatch plan.

[0086] Strategy filtering is based on dual filtering: Feasibility filtering: Strategies that trigger constraint alarms in the simulation are eliminated. For example, if a strategy requires energy storage to discharge 80MW, but the simulation shows that the battery SOC will drop to 18% within 15 minutes (below the safety threshold of 20%), this strategy is marked as "infeasible". Deviation tolerance filtering: Only strategies with all critical parameter DTW deviations < δ_{thres} are retained. The list of critical parameters includes: grid frequency, critical node voltage, spinning reserve capacity (unit: megawatts, MW), etc.

[0087] The overall ranking uses a weighted scoring model: Evaluation dimensions: Energy efficiency η (weight 0.6): Calculated as actual power generation / theoretical maximum power generation (unit: percentage %); Execution cost C (weight 0.3): Includes fuel cost (RMB / MWh) and equipment wear and tear cost (such as discount for start-up and shutdown cycles); Carbon emission intensity E (weight 0.1): Converted to carbon dioxide equivalent (unit: tons / MWh tCO2e / MWh).

[0088] Scoring formula: Score = 0.6 × η_max + 0.3 × C_min + 0.1 × E_min. Where η_max is the highest historical efficiency (e.g., 92%), C_min is the lowest cost (e.g., 200 RMB / MWh), and E_min is the lowest carbon intensity (e.g., 0.4 tCO2e / MWh). Scheme enhancement: Implementation steps breakdown: Decompose the top 20% of strategies into actionable instruction sequences. For example: Time window [14:00-14:05]: 1. Command #001: Start gas turbine GT101, target output 50MW (ramp rate 4MW / min); 2. Command #002: Reduce the power limit of photovoltaic inverter PV_Inverter05 to 85%. Emergency backup plan: Rollback logic is designed for major failure scenarios: Scenario 1: A generator unit fails and exits – activate the backup diesel generator unit (response time < 2 minutes); Scenario 2: Load surges by 10% – call upon demand response resources (reduce interruptible load by 200MW). The final generated trusted energy dispatch plan document includes: a Gantt chart of strategy execution sequence; a table of equipment operation command codes; expected energy efficiency / cost / carbon emission quantitative indicators; and an emergency scenario handling flowchart.

[0089] Various scheduling strategies are simulated and executed in a digital twin environment, and their actual effectiveness is verified through high-fidelity simulation. A bidirectional mapping mechanism compares the virtual simulation results with the feedback from the physical system in real time, identifies and corrects model deviations, ensures the executability of the strategies in real-world scenarios, significantly reduces trial-and-error costs, and avoids system risks caused by inappropriate strategies. Digital twin verification provides a safe buffer for strategy implementation and enhances the reliability of decision-making.

[0090] S205, implement blockchain notarization and smart contract execution for the trusted energy dispatch scheme to generate tamper-proof energy transaction records, which serve as the final execution basis for the digital management of the energy system.

[0091] Specifically, the trusted energy dispatch scheme can be structured and parsed to extract key information including scheme number, execution time interval, list of energy equipment involved, and energy allocation of each stage. The SHA-256 hash algorithm is used to calculate the unique hash value of the scheme. The key information is bound to the hash value to generate scheme summary information, ensuring that the summary information can uniquely identify the scheme content. The scheme summary and hash value are then output. Structured parsing and key information extraction Trusted energy dispatch schemes are stored in JSON or XML format, containing multi-level nested data structures. The parsing process first identifies the top-level tags: Scheme ID: A unique identifier, such as "ESS-2025-0382", where "ESS" represents the energy dispatch system, "2025" is the year, and "0382" is the serial number.

[0092] Execution time window: Start and end timestamps accurate to milliseconds (e.g., Start_Time: 2025-04-11T09:00:00.000Z, End_Time: 2025-04-11T17:30:00.000Z).

[0093] Energy Equipment List (Device_List): Lists the equipment types and instance codes involved, such as "Wind Turbine #WTG-07", "Photovoltaic Array #PV-Farm-12", and "Energy Storage Unit #BESS-23".

[0094] Energy Allocation: Records device-level allocation strategies as key-value pairs, such as "WTG-07:15.2MW", "PV-Farm-12: 8.7MW", "BESS-23: Charging power 3.5MW / Discharging power 2.1MW".

[0095] The parsing engine iterates through all fields, removes comments and redundant metadata, and retains only the core fields mentioned above to generate a standardized data dictionary.

[0096] Hash calculation and digest binding The extracted key information is concatenated into a string in a fixed order and input into the SHA-256 hash algorithm (SecureHash Algorithm 256-bit). This algorithm performs 64 rounds of non-linear transformations (such as bit rotation and modular addition) on the input to generate a 256-bit (32-byte) hexadecimal hash value (e.g., "a1b2c3d4e5f6..."). This hash value is collision-resistant: even if one character in the scheme changes (e.g., "15.2MW" changes to "15.3MW"), the new hash value will be completely different (e.g., "f9e8d7c6b5a4..."). The system encapsulates the original key information and hash value into a scheme digest (Scheme_Digest), with an example structure as follows: { "Scheme_ID": "ESS-2025-0382", "Hash_Value": "a1b2c3d4e5f6...", "Critical_Data": {raw key-value pairs} }

[0097] The digest ensures the uniqueness of the content through a digital fingerprint mechanism: any tampering will invalidate the hash value.

[0098] Integrity verification and output Perform double validation before output: Reverse verification: Recalculate the SHA-256 using the original data in the digest and compare it to the stored hash value; Field completeness check: Confirm that no critical fields are missing (such as omitting a device). After successful verification, output the scheme digest and hash value to the next stage, and simultaneously back them up locally to a secure storage area.

[0099] The solution summary and hash value are uploaded to the Energy Consortium Chain, and a consensus request is initiated. Each node participates in the consensus by verifying the completeness of the solution summary and the correctness of the hash value. When more than 2 / 3 of the nodes confirm that it is valid, the notarization is completed. The on-chain record includes the notarization time, node signature and hash value, and outputs the blockchain notarization certificate. Consortium blockchain node architecture and upload mechanism The Energy Consortium Chain consists of multiple participating nodes, including: Generator Nodes: such as operators of thermal power plants and wind farms; Grid Operator Nodes: responsible for balancing the regional power grid; and Regulatory Nodes: such as energy administration bureaus.

[0100] Each node deploys a blockchain client and establishes secure communication through two-way authentication using CA certificates. During upload, the scheduling system packages the scheme digest and hash value into a notarization request transaction (Notarization_Tx) and broadcasts it to all nodes in the network. The transaction structure includes the request ID, the initiator's digital signature, the digest data, and the hash value.

[0101] Consensus Verification Process Upon receiving a request, the node performs the following verifications: Digest integrity verification: checks whether fields are complete (e.g., whether Scheme_ID and Device_List exist); Hash correctness verification: recalculates the digest data using local SHA-256 and compares the result with the hash value in the transaction.

[0102] After successful verification, the node signs the transaction and broadcasts a "confirmation vote (Ack_Vote)". The system counts the number of votes, and consensus is triggered when the number of confirming nodes exceeds 2 / 3 of the total number of nodes (e.g., ≥7 out of 10 nodes confirm).

[0103] Block generation and evidence output After consensus is reached, the leader node (selected by a round-robin algorithm) packages the transaction into a new block and executes it. Timestamp annotation: The time of the record is written in Coordinated Universal Time (UTC) format (e.g., 2025-04-11T09:05:23.456Z). Multi-signature: All confirming nodes are appended with an Elliptic Curve Digital Signature (ECDSA Signature). On-chain records: Block data includes {transaction content, evidence storage time, node signature list, and block hash}.

[0104] The final generated blockchain notary certificate (Blockchain_Notary_Certificate) has the following format example: { "Block_Height": 14208, / / Block height "Tx_ID": "TX-9a8b7c6d", / / Transaction ID "Notary_Time": "2025-04-11T09:05:23.456Z", "Signing_Nodes": ["Node_A_Sig", "Node_B_Sig", ...] / / Signature list }

[0105] The certificate is returned to the scheduling system as evidence stored on the blockchain.

[0106] Smart contracts are written based on trusted energy dispatch schemes. The contract terms include trigger conditions, execution actions, and settlement rules. The contract is compiled into executable code through an on-chain virtual machine, deployed to the consortium blockchain and associated with the evidence storage certificate, and the deployed smart contract is output. Contract terms design Write smart contract terms based on the business logic of the scheduling scheme: Trigger Condition: Logical judgment based on real-time data, such as: If the real-time electricity price > the threshold Price_Threshold (0.12 $ / kWh) AND the photovoltaic output > the demand value Demand_Level (5MW); then the energy storage discharge is triggered.

[0107] Execution_Action: Equipment control commands, such as: SET BESS-23.mode = "Discharge"; SET BESS-23.power = 2.1MW.

[0108] Settlement Rules (Settlement_Rule): Measuring energy transactions, such as: Energy storage discharge revenue = Discharge volume × Real-time electricity price × Revenue sharing ratio (0.85). Terms are described using Solidity-like syntax and support time-driven (e.g., executed daily at 14:00) and event-driven (e.g., electricity price fluctuations).

[0109] Compilation and On-Chain Deployment Contract code is compiled via an on-chain virtual machine: the front-end compiler converts high-level language code into bytecode; the bytecode is input into the virtual machine to generate a contract account address, such as "0x7a3b...c9d1".

[0110] Deployment requires associating with a notarized certificate: the transaction ID (Tx_ID) from the certificate is written into the contract metadata to establish a two-way index. The deployment process includes: Gas Fee payment: the fee is calculated based on the contract complexity (e.g., 0.05 ETH); Multi-node verification: consortium blockchain nodes verify the legality of the association between the contract bytecode and the notarized certificate. A successfully deployed contract obtains a unique contract deployment certificate (Deployment_Receipt), containing the contract address and the notarized transaction ID.

[0111] Security Audit and Output Before deployment, three audits are performed: logical conflict detection: checking whether actions and conditions contradict each other (such as setting charge / discharge simultaneously); resource permission verification: confirming that the contract can only access authorized device interfaces; and evidence association verification: ensuring that the Tx_ID referenced by the contract actually exists on the chain. After the audits pass, the deployed smart contract is output to the consortium blockchain, and its status is marked as "Active".

[0112] Using smart contracts, trigger conditions are automatically determined based on real-time collected energy system operation data. When the conditions are met, the corresponding scheduling operation is executed. Each completed operation generates a transaction record containing the operation time, execution result, and equipment status. The hash value of the transaction record is written into the blockchain and linked with the evidence to form an immutable energy transaction record.

[0113] Condition triggering and operation execution Smart contracts listen to external data sources (called "oracles"): Energy Oracle: Pushes real-time data such as grid frequency, electricity price, and equipment power.

[0114] Contractual data is continuously compared with preset conditions: when the real-time electricity price is > 0.12 $ / kWh and the photovoltaic output is > 5MW, a discharge command is automatically triggered.

[0115] The execution process calls the device control API: POST / bess / 23 / control { "mode": "discharge", "power": 2100} to the energy storage system via the RESTful interface.

[0116] The execution result (such as "Success: BESS-23 discharging at 2100kW") is returned to the contract by the equipment.

[0117] Transaction record generation and on-chain evidence storage Each operation generates an Energy Transaction Record (Energy_Transaction_Record), which includes: Operation Time (Op_Time): UTC timestamp of instruction completion; Execution Result (Op_Result): Success / failure code and details (e.g., "Success_Code: 200"); Device Status (Device_Status): Device snapshot after the action (e.g., "SOC: 65%"). Record format example: { "Tx_Type": "BESS_Discharge", "Op_Time": "2025-04-11T14:30:45.789Z", "Result": "Success: 2100kW for 15min", "Device_State": { "BESS-23": { "SOC": 65%, "Temp": 32°C}} }

[0118] The SHA-256 hash value of the record is calculated, written to the blockchain as a new transaction, and associated with the initial evidence certificate (traceable through the Tx_ID in the contract metadata).

[0119] Chain association and immutability verification Blockchain uses Merkle trees to build data associations: the hash value of a transaction record serves as a leaf node; together with the root hash of the block containing the initial proof of ownership, a new Merkle root is generated. Any modification (such as tampering with the discharge amount "2100kW" → "2000kW") will result in: a change in the record hash value → invalidation of the Merkle root → disruption of block continuity.

[0120] The final output is an immutable energy transaction record, with the following characteristics: traceability: traceable to the original scheduling scheme through the contract address; audit transparency: regulatory nodes can verify the consistency of data throughout the entire chain; legal validity: multi-signature and timestamp constitute judicial evidence.

[0121] Verified scheduling schemes are written to the blockchain and automatically executed using smart contracts. Each energy transaction and scheduling operation generates a timestamped encrypted record, forming a complete audit trail. Distributed ledger technology ensures the transparency and tamper-proof nature of the records, establishing a trustworthy energy trading ecosystem and resolving trust issues in multi-party collaboration. Blockchain-based evidence storage provides reliable evidence for regulation, while smart contracts enable automated execution, improving management efficiency.

[0122] Another embodiment of the present invention provides an energy digital management system, see [link to relevant documentation]. Figure 3 The system may include: The receiving module 301 is used to receive multi-source heterogeneous energy data acquisition signals and generate an initial energy feature map based on a spatiotemporal correlation feature extraction algorithm, wherein each feature node is associated with the operating parameter dimension of the energy system. The encoding module 302 is used to encode the initial energy feature map using a graph neural network, construct an energy knowledge graph vector set, and decompose the high-dimensional vector in the knowledge graph vector set into several subspace vectors to capture the dynamic characteristics of energy flow. Each subspace vector represents the multi-dimensional operating state of the corresponding energy device. The scheduling module 303 is used to make real-time scheduling decisions on the subspace vector using a reinforcement learning-driven energy flow optimization model to generate a set of energy allocation strategies, wherein the optimization model maximizes energy utilization efficiency through a Markov decision process. Application module 304 is used to apply a digital twin verification algorithm to the energy allocation strategy set to perform physical-digital space consistency verification and generate a reliable energy scheduling scheme. The digital twin verification algorithm ensures the feasibility of the strategy through a bidirectional mapping mechanism. The execution module 305 is used to implement blockchain notarization and smart contract execution for the trusted energy dispatch scheme, and generate tamper-proof energy transaction records as the final execution basis for the digital management of the energy system.

[0123] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0124] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0125] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.

[0126] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A digital energy management method, characterized in that, The method includes: It receives multi-source heterogeneous energy data acquisition signals and generates an initial energy feature map based on a spatiotemporal correlation feature extraction algorithm, wherein each feature node is associated with the operating parameter dimension of the energy system; The initial energy feature map is encoded by a graph neural network to construct an energy knowledge graph vector set. The high-dimensional vectors in the knowledge graph vector set are decomposed into several subspace vectors to capture the dynamic characteristics of energy flow. Each subspace vector represents the multi-dimensional operating state of the corresponding energy device. A reinforcement learning-driven energy flow optimization model is used to make real-time scheduling decisions on the subspace vectors to generate a set of energy allocation strategies. The optimization model maximizes energy utilization efficiency through a Markov decision process. The digital twin verification algorithm is applied to the energy allocation strategy set to perform physical-digital space consistency verification and generate a reliable energy scheduling scheme. The digital twin verification algorithm ensures the feasibility of the strategy through a two-way mapping mechanism. The trusted energy dispatch scheme is implemented with blockchain notarization and smart contract execution to generate tamper-proof energy transaction records, which serve as the final execution basis for the digital management of the energy system.

2. The method according to claim 1, characterized in that, The system receives multi-source heterogeneous energy data acquisition signals and generates an initial energy feature map based on a spatiotemporal correlation feature extraction algorithm. Each feature node is associated with a dimension of the energy system's operating parameters, including: It receives multi-source heterogeneous energy data acquisition signals, adds a unified timestamp to each signal through high-precision clock synchronization technology, generates a spatiotemporal index label by combining the GPS coordinates installed on the device, filters out invalid acquisition points with signal strength below the strength threshold, and outputs a set of raw energy signals with spatiotemporal index. The original energy signal set with spatiotemporal index is preprocessed, and the signal components of different frequency bands are separated by wavelet packet decomposition algorithm. The noise components generated by electromagnetic interference are eliminated by adaptive threshold denoising. At the same time, the missing data is supplemented by interpolation method based on the similarity of adjacent time moments, and the denoised and completed energy data sequence is output. Based on the denoised and completed energy data sequence, the operating parameter dimensions of each energy device are extracted, the CFS feature selection algorithm is used to calculate the mutual information value between parameters, the core parameters that are strongly correlated with the energy system state are retained, and a parameter-device association table is generated. Based on the parameter-equipment association table, a spatiotemporal association rule mining algorithm is used to calculate the co-occurrence frequency and association strength of different equipment parameters. The equipment parameters are used as feature nodes and the association strength is used as edge weights to construct an initial energy feature map containing node attributes and edge attributes.

3. The method according to claim 2, characterized in that, The initial energy feature map is encoded using a graph neural network to construct an energy knowledge graph vector set. The high-dimensional vectors in the knowledge graph vector set are then decomposed into several subspace vectors to capture the dynamic characteristics of energy flow. Each subspace vector represents the multidimensional operating state of the corresponding energy device, including: The importance of nodes in the initial energy feature map is evaluated. The PageRank algorithm is used to calculate the importance score of each feature node. The top 90% of nodes are retained to simplify the map structure. Weakly related edges with weights below the weight threshold are deleted, and a simplified energy feature map is output. The simplified energy feature map is input into the graph attention neural network. The dependencies between nodes are learned through the multi-head attention mechanism. The attribute features and associated features of each node are fused and encoded into a 128-dimensional high-dimensional vector. Each dimension of the vector corresponds to the abstract expression of the feature, and the energy knowledge graph vector set is output. Based on the dynamic characteristics of energy flow, the kernel principal component analysis algorithm is used to decompose the high-dimensional vectors in the energy knowledge graph vector set into subspaces. The decomposition dimension is set according to the energy transfer efficiency of each link, and a preliminary subspace vector set is output. The initial subspace vector set is processed by the Gram-Schmidt orthogonalization process to eliminate the coupling correlation between different vectors, so that each subspace vector independently represents the operating state of a single energy link, and the optimized subspace vector set is output.

4. The method according to claim 3, characterized in that, The energy flow optimization model driven by reinforcement learning performs real-time scheduling decisions on the subspace vectors to generate a set of energy allocation strategies. The optimization model maximizes energy utilization efficiency through a Markov decision process, including: The optimized subspace vector set is used as the state input for reinforcement learning. The state space is defined as the combination of real-time operating parameters of each energy link, and the action space is the executable scheduling operation. Each action corresponds to a specific parameter adjustment range, and the state-action space definition table is output. A multi-objective reward function is designed, with maximizing energy utilization efficiency as the primary objective and minimizing equipment operating costs and carbon emission intensity as secondary objectives. The weight allocation is determined by the analytic hierarchy process (AHP). Real-time state parameters are substituted into the function to calculate the instantaneous reward value, and the reward function calculation model is output. The DDPG algorithm is used to train the energy flow optimization model. The state-action space definition table and the reward function calculation model are used as inputs. The Actor-Critic network is updated iteratively through Markov decision process. An experience replay mechanism is introduced to store and sample effective decision samples, and a mature energy flow optimization model is output. The subspace vectors collected in real time are input into the trained and mature energy flow optimization model to generate multiple sets of energy allocation strategies. Infeasible strategies are eliminated by checking the equipment capacity constraints. The comprehensive benefit score of each strategy is calculated, and the set of energy allocation strategies sorted by score is output.

5. The method according to claim 4, characterized in that, The energy allocation strategy set is applied to a digital twin verification algorithm to perform physical-digital space consistency verification and generate a reliable energy scheduling scheme. The digital twin verification algorithm ensures strategy feasibility through a bidirectional mapping mechanism, including: A full-element digital twin model is constructed based on the physical parameters of the energy system. Each strategy in the energy allocation strategy set is converted into a control command that the model can recognize, and the input command set of the twin model is output. The simulation engine of the digital twin model is started, and each instruction in the input instruction set is dynamically simulated. The changes of key parameters during the simulation process are recorded, and digital space operation curves are generated. At the same time, real-time physical operation data of the energy system is collected through IoT sensors to generate physical space operation curves and output a set of digital-physical operation curves. The dynamic time warping algorithm is used to calculate the deviation value of the corresponding parameter in the digital-physical operation curve set. When the deviation value exceeds the deviation threshold, the deviation source is located through a two-way mapping mechanism, the corresponding parameter in the twin model is corrected, and the simulation is repeated until the deviation value is lower than the deviation threshold. A deviation correction report is then output. Preliminary energy allocation strategies are selected based on the deviation correction report. The strategies are then ranked comprehensively based on their execution costs and energy utilization efficiency. The top 20% of the optimal strategies are selected, and implementation steps and emergency backup plans are added to generate a reliable energy dispatch plan.

6. The method according to claim 5, characterized in that, The implementation of blockchain-based notarization and smart contract execution for the trusted energy dispatch scheme, generating immutable energy transaction records as the final basis for the digital management of the energy system, includes: The trusted energy dispatch scheme is structured and parsed to extract key information including scheme number, execution time interval, list of energy equipment involved, and energy allocation of each stage. The SHA-256 hash algorithm is used to calculate the unique hash value of the scheme. The key information is bound to the hash value to generate scheme summary information, ensuring that the summary information can uniquely identify the scheme content. The scheme summary and hash value are output. The solution summary and hash value are uploaded to the Energy Consortium Chain, and a consensus request is initiated. Each node participates in the consensus by verifying the completeness of the solution summary and the correctness of the hash value. When more than 2 / 3 of the nodes confirm that it is valid, the notarization is completed. The on-chain record includes the notarization time, node signature and hash value, and outputs the blockchain notarization certificate. Smart contracts are written based on trusted energy dispatch schemes. The contract terms include trigger conditions, execution actions, and settlement rules. The contract is compiled into executable code through an on-chain virtual machine, deployed to the consortium blockchain and associated with the evidence storage certificate, and the deployed smart contract is output. Using smart contracts, trigger conditions are automatically determined based on real-time collected energy system operation data. When the conditions are met, the corresponding scheduling operation is executed. Each completed operation generates a transaction record containing the operation time, execution result, and equipment status. The hash value of the transaction record is written into the blockchain and linked with the evidence to form an immutable energy transaction record.

7. An energy digital management system, characterized in that, The system includes: The receiving module is used to receive multi-source heterogeneous energy data acquisition signals and generate an initial energy feature map based on a spatiotemporal correlation feature extraction algorithm, wherein each feature node is associated with the operating parameter dimension of the energy system. The encoding module is used to encode the initial energy feature map using a graph neural network, construct an energy knowledge graph vector set, and decompose the high-dimensional vectors in the knowledge graph vector set into several subspace vectors to capture the dynamic characteristics of energy flow. Each subspace vector represents the multi-dimensional operating state of the corresponding energy device. The scheduling module is used to make real-time scheduling decisions on the subspace vector using a reinforcement learning-driven energy flow optimization model, and generate a set of energy allocation strategies, wherein the optimization model maximizes energy utilization efficiency through a Markov decision process. The application module is used to apply a digital twin verification algorithm to the energy allocation strategy set to perform physical-digital space consistency verification and generate a reliable energy scheduling scheme. The digital twin verification algorithm ensures the feasibility of the strategy through a two-way mapping mechanism. The execution module is used to implement blockchain notarization and smart contract execution for the trusted energy dispatch scheme, generating tamper-proof energy transaction records as the final execution basis for the digital management of the energy system.

8. The system according to claim 7, characterized in that, The receiving module is specifically used for: It receives multi-source heterogeneous energy data acquisition signals, adds a unified timestamp to each signal through high-precision clock synchronization technology, generates a spatiotemporal index label by combining the GPS coordinates installed on the device, filters out invalid acquisition points with signal strength below the strength threshold, and outputs a set of raw energy signals with spatiotemporal index. The original energy signal set with spatiotemporal index is preprocessed, and the signal components of different frequency bands are separated by wavelet packet decomposition algorithm. The noise components generated by electromagnetic interference are eliminated by adaptive threshold denoising. At the same time, the missing data is supplemented by interpolation method based on the similarity of adjacent time moments, and the denoised and completed energy data sequence is output. Based on the denoised and completed energy data sequence, the operating parameter dimensions of each energy device are extracted, the CFS feature selection algorithm is used to calculate the mutual information value between parameters, the core parameters that are strongly correlated with the energy system state are retained, and a parameter-device association table is generated. Based on the parameter-equipment association table, a spatiotemporal association rule mining algorithm is used to calculate the co-occurrence frequency and association strength of different equipment parameters. The equipment parameters are used as feature nodes and the association strength is used as edge weights to construct an initial energy feature map containing node attributes and edge attributes.

9. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-6 when it is run.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-6.

Images