Energy efficient scheduling method based on edge computing

Abstract

The invention discloses a high-efficiency energy scheduling method based on edge computing. The method comprises the following steps: step 1, acquiring operation state information at an edge node; step 2, operating the improved LightGBM model, and uploading parameters of the local improved LightGBM model to a federal aggregation server; 3, selecting an optimal scheduling action in the discrete scheduling action set based on the deep Q network; 4, issuing the local scheduling instruction to the energy equipment; 5, the federal aggregation server executes parameter aggregation operation; and step 6, compressing the global energy scheduling model based on a knowledge distillation mechanism. According to the method, low-delay and high-precision energy load prediction and scheduling control can be realized in a multi-edge node cooperation scene, and the method is suitable for intelligent energy consumption management and dynamic load optimization of industrial parks, building clusters and distributed energy systems.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of energy management and edge intelligence, and particularly relates to an energy efficient scheduling method based on edge computing. BACKGROUND

[0002] How to realize efficient coordination and intelligent scheduling of multiple types of energy equipment has become a key problem in the research of smart energy systems. Traditional energy scheduling systems mostly rely on centralized computing architecture, and the scheduling strategy is often based on fixed rules or empirical models, which is difficult to cope with the real-time and dynamic challenges in the multi-source heterogeneous data environment. In addition, with the development of edge computing and Internet of Things technology, energy equipment has been widely deployed in various edge nodes, and the collection and processing capacity of its operation data has been significantly enhanced, providing the possibility of intelligent decision-making locally. However, how to fully tap the data value of each edge node while ensuring data privacy and controllable network resources, and realizing joint training and continuous optimization of the model, is still a major difficulty in current technology.

[0003] In the prior art, the load prediction model usually adopts time series algorithm or simple machine learning model, such as ARIMA, SVR or traditional GBDT, which lacks the ability to model the fusion of short-term fluctuations and long-term trends, resulting in insufficient prediction accuracy. As for scheduling decisions, common methods such as heuristic algorithms or reinforcement learning networks often ignore the coupling relationship between load prediction results and environmental status, making it difficult to achieve collaborative optimization. In addition, some research attempts to introduce the LightGBM model for prediction tasks, but does not combine multi-time scale features and does not consider the adaptive ability of the model on the edge node. More importantly, most current systems still use centralized model updating methods, which cannot meet the requirements of data security and local computing in the edge environment. Although existing research has introduced a federated learning framework to realize distributed model training, it often uses simple parameter averaging or weighting mechanism in the model aggregation process, without designing a special aggregation strategy for tree structure models and deep network models, which affects the generalization performance of the global model and the reliability of scheduling execution.

[0004] Therefore, how to provide an energy efficient scheduling method based on edge computing is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide an energy efficient scheduling method based on edge computing. The present application combines an improved LightGBM load prediction model and a deep Q network scheduling optimization model, and combines federated learning and knowledge distillation mechanism. The whole process of local model training, global parameter aggregation and compression optimization in a multi-edge node environment is described in detail, which has the advantages of high prediction accuracy, fast scheduling response and low communication overhead.

[0006] According to an embodiment of the present application, an energy-efficient scheduling method based on edge computing includes the following steps:

[0007] Step 1: Collecting running state information at the edge node, performing preprocessing operations, and generating standardized feature data;

[0008] Step 2: Running an improved LightGBM model based on standardized feature data to generate local load prediction results, and performing distributed training and parameter updating through a federated learning mechanism;

[0009] Step 3: Based on the local load prediction results and real-time state parameters, selecting the optimal scheduling action in the discrete scheduling action set based on the deep Q network, and generating the optimal scheduling action as a local scheduling instruction;

[0010] Step 4: Issuing the local scheduling instruction to the energy device;

[0011] Step 5: In a predetermined period, each edge node uploads the improved LightGBM model parameters and deep Q network parameters to the federated aggregation server, the federated aggregation server performs parameter aggregation operations to generate a global energy scheduling model;

[0012] Step 6: Based on the knowledge distillation mechanism and model compression mechanism, the global energy scheduling model is compressed and optimized, and the compressed and optimized global energy scheduling model is issued to each edge node.

[0013] Optionally, the edge node collects running state information and performs preprocessing operations, specifically:

[0014] At each edge node, the running state information collected by the energy consumption monitoring device, the environmental perception device and the energy device deployed on site is obtained, and the running state information includes energy consumption monitoring data, environmental parameters and device running state data;

[0015] The preprocessing operation includes performing missing value filling, outlier removal, data format unification and other data cleaning operations on the original running state information to obtain cleaned running state data;

[0016] The cleaned running state data is divided into feature vectors according to the preset feature dimension and time window division method, and the feature vectors include short-term state sequences and long-term trend sequences constructed based on sliding time windows;

[0017] Perform normalization processing and dimension alignment operations on the feature vectors to generate standardized feature data.

[0018] Optionally, the improved LightGBM model is run based on the standardized feature data to generate a local load prediction result, and distributed training and parameter updating are performed through a federated learning mechanism, specifically as follows:

[0019] The initial parameters of the improved LightGBM model are loaded on each edge node, and the standardized feature data is received;

[0020] The short-term state sequence and the long-term trend sequence are combined into a multi-time scale input matrix according to a preset splicing order;

[0021] The multi-time scale input matrix is divided into a training set and a validation set according to time sequence, the training hyperparameters of the improved LightGBM model are set, a local training round is performed, and local improved LightGBM model parameters are obtained;

[0022] The local improved LightGBM model parameters include tree structure level number, split feature index, split threshold interval, and leaf node weight parameters;

[0023] The training hyperparameters include an objective function, a learning rate, a maximum number of leaves, a maximum depth, a subsampling ratio, and a feature subsampling ratio;

[0024] The multi-time scale input matrix is input into the improved LightGBM model, and the node split path traversal of the multi-time scale input matrix is performed in each decision tree to determine the corresponding leaf node output value;

[0025] The decision tree leaf node output value corresponding to the short-term state sequence and the decision tree leaf node output value corresponding to the long-term trend sequence are calculated respectively, and the results are combined according to a preset weighting ratio to obtain a prediction value corresponding to a time slice;

[0026] The continuous prediction values are arranged in time sequence to generate a local load prediction result;

[0027] In a preset period, each edge node uploads the local improved LightGBM model parameters to a federated aggregation server;

[0028] The federated aggregation server performs parameter aggregation operation on the local improved LightGBM model parameters uploaded by each edge node to generate global improved LightGBM model parameters;

[0029] The federated aggregation server distributes the global improved LightGBM model parameters to each edge node.

[0030] Optionally, based on the local load prediction result and real-time state parameters, the optimal scheduling action is selected from a discrete scheduling action set based on a deep Q network, and the optimal scheduling action is used to generate a local scheduling instruction, specifically as follows:

[0031] The current time slice prediction value from the local load forecast results, the current value, voltage value and power value from the energy consumption monitoring data, the temperature, humidity and light intensity from the environmental parameters, and the equipment start-up and shutdown status, load rate and energy storage device charge status from the equipment operation status data are spliced ​​together in a preset order to form a state vector.

[0032] During the forward propagation of the deep Q-network, the Q-values ​​of candidate scheduling actions in the discrete scheduling action set are calculated based on the state vector;

[0033] The discrete scheduling action set includes energy equipment start-stop actions corresponding to scheduling action identifiers, energy equipment power level switching actions, energy storage device charging actions, energy storage device discharging actions, energy storage device state maintenance actions, and control actions based on energy equipment operating mode switching.

[0034] Compare the Q values ​​of each candidate scheduling action and determine the candidate scheduling action with the largest Q value as the optimal scheduling action;

[0035] The optimal scheduling action is converted into a local scheduling instruction according to a preset format. The local scheduling instruction includes a scheduling action identifier and corresponding execution parameters.

[0036] Optionally, the step of sending local scheduling instructions to energy equipment specifically includes:

[0037] Local scheduling instructions are encapsulated at each edge node. The encapsulation content includes scheduling action identifier, execution parameters, target energy device identifier, execution time parameter, and duration parameter.

[0038] Select a communication link with the target energy equipment and establish a communication connection;

[0039] Local dispatch commands are sent to the target energy equipment via the communication connection.

[0040] The target energy equipment returns a confirmation message indicating receipt of the instruction.

[0041] The target energy equipment parses the scheduling action identifier and execution parameters according to the local scheduling instructions, and outputs the operating status information and execution result information to each edge node;

[0042] Each edge node records its running status information and execution result information, and generates a scheduling execution record.

[0043] Optionally, within a preset period, each edge node uploads the improved LightGBM model parameters and deep Q-network parameters to the federated aggregation server. The federated aggregation server performs parameter aggregation operations to generate a global energy scheduling model, specifically as follows:

[0044] Within a preset period, each edge node uploads its local improved LightGBM model parameters and local deep Q network parameters to the federated aggregation server;

[0045] The local deep Q-network parameters include the weight parameters and bias parameters of each layer;

[0046] The federated aggregation server receives the local improved LightGBM model parameters and local deep Q network parameters uploaded by each edge node and establishes a parameter index table;

[0047] The federated aggregation server performs hierarchical tree-structured aggregation operations on the improved LightGBM model, including:

[0048] The nodes at the same level are aligned according to the tree structure hierarchy number. At higher-level nodes, a majority vote is performed based on the split feature index to determine the global split feature index. At higher-level nodes, the global split threshold is determined based on the split threshold range according to the preset threshold selection rule to generate the global tree structure. The federated aggregation server performs aggregation at the leaf nodes based on the leaf node weight parameters uploaded by each edge node to generate the global improved LightGBM model.

[0049] The federated aggregation server performs parameter aggregation operations on the deep Q network, including: aligning the weight parameters and bias parameters of each layer according to the network hierarchy, and performing aggregation to obtain the global deep Q network;

[0050] The global energy scheduling model generated by the federated aggregation server includes a global improved LightGBM model and a global deep Q network model.

[0051] Optionally, the step of compressing and optimizing the global energy scheduling model based on the knowledge distillation mechanism and the model compression mechanism, and then distributing the compressed and optimized global energy scheduling model to each edge node, specifically involves:

[0052] Using the global energy scheduling model as the teacher model, a distillation training dataset is constructed, and the distillation training dataset is input into the teacher model to generate soft-labeled output results.

[0053] A student model is constructed in a federated aggregation server, the student model comprising an improved LightGBM model and a deep Q-network model;

[0054] Input the distillation training dataset into the student model to obtain the prediction results of the student model;

[0055] In the federated aggregation server, the difference between the soft-label output of the teacher model and the prediction of the student model is used as the distillation loss to perform parameter optimization and update the parameters of the improved LightGBM model and the deep Q network.

[0056] After completing the knowledge distillation operation, the federated aggregation server performs model compression operation on the improved LightGBM model and the deep Q network model. The model compression operation includes pruning, quantization and redundant parameter removal to obtain a compressed and optimized global energy scheduling model.

[0057] The federated aggregation server distributes the compressed and optimized global energy scheduling model to each edge node to replace and update the local improved LightGBM model and the local deep Q network model in the edge nodes.

[0058] Optionally, the output result of generating soft tags is specifically as follows:

[0059] The running status data in the distillation training dataset is input into the global improved LightGBM model to obtain the load prediction value sequence for the corresponding time slice. The load prediction value sequence is then normalized by temperature coefficient to generate soft labels in the form of prediction probability distribution.

[0060] The load forecast sequence, current, voltage and power values ​​from energy consumption monitoring data, temperature, humidity and light intensity from environmental parameters, and equipment start-up and shutdown status, load rate and energy storage device state of charge from equipment operation status data are spliced ​​together in a preset order to form a distillation state vector.

[0061] The distillation state vector is input into the global deep Q-network model to obtain the Q-value set corresponding to each candidate scheduling action in the discrete scheduling action set. Softmax normalization is then performed on the Q-value set to generate soft labels in the form of the probability distribution of candidate scheduling actions.

[0062] The soft label in the form of the predicted probability distribution and the soft label in the form of the candidate scheduling action probability distribution together constitute the soft label output result.

[0063] The beneficial effects of this invention are:

[0064] This invention achieves efficient perception and intelligent processing of local energy consumption monitoring data, environmental parameters, and equipment operating status by deploying an improved LightGBM model and a deep Q-network model at edge nodes. The improved LightGBM model, used for training and predicting feature vectors across multiple time scales, effectively improves the timeliness and accuracy of load forecasting, accurately reflecting short-term state changes and long-term trend fluctuations, thus providing reliable support for subsequent scheduling strategies. The deep Q-network, through joint modeling of local load forecasting results and real-time state parameters, dynamically calculates the Q-value of each action in the discrete scheduling action set, ensuring that the selected scheduling instructions have optimal energy efficiency and execution effect under the current state.

[0065] This invention constructs a global model aggregation mechanism based on federated learning. Each edge node synchronizes only its local model parameters without uploading the original data, effectively protecting data privacy. Simultaneously, hierarchical aggregation and alignment strategies ensure the structural consistency and performance stability of the aggregated model. By introducing knowledge distillation and model compression mechanisms, the deployment efficiency of the global model on edge devices is further improved. While maintaining prediction and scheduling accuracy, the model size and computational resource consumption are significantly reduced, ensuring the method possesses good practicality and scalability. Attached Figure Description

[0066] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0067] Fig. 1 This is an overall flowchart of an energy-efficient scheduling method based on edge computing proposed in this invention;

[0068] Fig. 2 This is a schematic diagram of the improved LightGBM model structure of an energy-efficient scheduling method based on edge computing proposed in this invention.

[0069] Fig. 3 This is a flowchart of federated aggregation and knowledge distillation graphs for an energy-efficient scheduling method based on edge computing proposed in this invention. Detailed Implementation

[0070] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0071] refer to Figs. 1-3 An energy-efficient scheduling method based on edge computing includes the following steps:

[0072] Step 1: Collect operational status information at edge nodes, perform preprocessing operations, and generate standardized feature data;

[0073] Step 2: Run the improved LightGBM model based on standardized feature data to generate local load prediction results, and perform distributed training and parameter updates through a federated learning mechanism;

[0074] Step 3: Based on the local load forecast results and real-time status parameters, select the optimal scheduling action from the discrete scheduling action set based on the deep Q network, and generate local scheduling instructions from the optimal scheduling action;

[0075] Step 4: Send local dispatch instructions to energy equipment;

[0076] Step 5: Within a preset period, each edge node uploads the improved LightGBM model parameters and the deep Q network parameters to the federated aggregation server. The federated aggregation server performs parameter aggregation operations to generate a global energy scheduling model.

[0077] Step 6: Based on the knowledge distillation mechanism and model compression mechanism, compress and optimize the global energy scheduling model, and distribute the compressed and optimized global energy scheduling model to each edge node.

[0078] This invention proposes an efficient energy scheduling method based on edge computing. Combining an improved LightGBM model and a deep Q-network model, it constructs a closed-loop scheduling process from operational status information collection, standardization processing, load forecasting, scheduling action generation, command issuance, to federated aggregation and model optimization. This method collects multi-source status data in real time from edge nodes, uses the improved LightGBM model for local load forecasting, and dynamically selects the optimal scheduling action based on the deep Q-network model to achieve fine-grained control of energy equipment. Simultaneously, it aggregates model parameters from each node through a federated learning mechanism to construct a global energy scheduling model. Furthermore, by incorporating knowledge distillation and model compression strategies, it achieves lightweight deployment and continuous optimization of the model, improving the intelligence level of energy scheduling and its adaptability to edge deployment.

[0079] In this embodiment, the step of collecting operating status information at edge nodes and performing preprocessing operations specifically includes:

[0080] At each edge node, the operating status information collected by energy consumption monitoring equipment, environmental sensing equipment and energy equipment deployed on site is acquired. The operating status information includes energy consumption monitoring data, environmental parameters and equipment operating status data.

[0081] The preprocessing operation includes: performing data cleaning operations such as missing value imputation, outlier removal, and data format unification on the original running status information to obtain cleaned running status data;

[0082] The cleaned operational status data is used to construct feature vectors according to preset feature dimensions and time window division methods. The feature vectors include short-term status sequences and long-term trend sequences constructed based on sliding time windows.

[0083] Short-term state sequences and long-term trend sequences are generated according to multi-timescale division rules, specifically including:

[0084] The continuously collected operational status data are arranged in chronological order to form an original time series; a first time window parameter and a second time window parameter are set on the original time series, corresponding to the short-term window length and the long-term window length, respectively;

[0085] The first time window is slid with a fixed step size to extract multiple short-term segments from the original time series. The length of each segment is equal to the length of the short-term window, and short-term state sequences are constructed sequentially. The second time window is slid with the same or different step sizes to extract multiple long-term segments from the original time series. The length of each segment is equal to the length of the long-term window, and long-term trend sequences are constructed sequentially. Feature statistics, normalization, and missing value imputation operations are performed on the short-term state sequences and long-term trend sequences, respectively, to generate short-term feature vectors and long-term feature vectors. The short-term feature vectors and long-term feature vectors are concatenated or combined according to preset concatenation rules to form a complete multi-timescale feature vector.

[0086] Normalization and dimension alignment operations are performed on the feature vectors to generate standardized feature data.

[0087] This invention constructs a preprocessing and feature extraction mechanism for multi-source heterogeneous operating status information at the edge node side. By collecting energy consumption monitoring data, environmental parameters, and equipment operating status data, it performs cleaning operations such as missing value imputation, outlier removal, and format unification. Combined with a sliding time window strategy, it constructs short-term state sequences and long-term trend sequences, further extracts and splices short-term feature vectors and long-term feature vectors to form multi-time-scale feature vectors, and through normalization and dimension alignment processing, finally generates standardized feature data that can be used for model input, providing a high-quality time-series feature foundation for subsequent load forecasting and intelligent scheduling.

[0088] In this embodiment, the step of running the improved LightGBM model based on standardized feature data to generate local load prediction results, and performing distributed training and parameter updates through a federated learning mechanism, specifically involves:

[0089] Load the initial parameters of the improved LightGBM model at each edge node and receive standardized feature data;

[0090] The short-term state sequence and the long-term trend sequence are combined into a multi-time-scale input matrix according to a preset splicing order;

[0091] The multi-timescale input matrix is ​​divided into training and validation sets according to time order. The training hyperparameters of the improved LightGBM model are set, and local training rounds are executed to obtain the parameters of the local improved LightGBM model.

[0092] The parameters of the locally improved LightGBM model include tree structure level number, split feature index, split threshold range and leaf node weight parameters.

[0093] The tree structure level number is determined based on the level of the node during the decision tree generation process. The split feature index is determined based on the optimal feature selected by gain calculation when the node splits. The split threshold interval is obtained when the optimal split point of the feature is calculated. The leaf node weight parameter is obtained when the output value of the leaf node is calculated based on the gradient and the second derivative.

[0094] Training hyperparameters include the objective function, learning rate, maximum number of leaves, maximum depth, subsampling ratio, and feature subsampling ratio;

[0095] The multi-timescale input matrix is ​​input into the improved LightGBM model, and node splitting path traversal is performed on the multi-timescale input matrix in each decision tree to determine the corresponding leaf node output value.

[0096] The output values ​​of the decision tree leaf nodes corresponding to the short-term state sequence and the output values ​​of the decision tree leaf nodes corresponding to the long-term trend sequence are calculated separately and weighted, and then combined according to the preset weighting ratio to obtain the predicted value of the corresponding time slice.

[0097] Arrange consecutive forecast values ​​in chronological order to generate local load forecast results;

[0098] Within a preset period, each edge node uploads the parameters of its local improved LightGBM model to the federated aggregation server;

[0099] The federated aggregation server performs parameter aggregation operations on the local improved LightGBM model parameters uploaded by each edge node to generate global improved LightGBM model parameters;

[0100] The federated aggregation server distributes the global improved LightGBM model parameters to each edge node.

[0101] This invention is based on an edge computing architecture. An improved LightGBM model is loaded into each edge node. By receiving standardized feature data and constructing a multi-time-scale input matrix, local load prediction modeling is performed by combining short-term state sequences and long-term trend sequences. Local model parameters such as tree structure level numbers, split feature indices, split threshold intervals, and leaf node weight parameters are obtained using gain calculation and gradient optimization mechanisms. These parameters are periodically uploaded to a federated aggregation server through a distributed federated learning mechanism, achieving secure aggregation and global updates of multi-node model parameters. This effectively improves the model's generalization ability and prediction accuracy, while taking into account both data privacy protection and energy load modeling accuracy.

[0102] In this embodiment, the step of selecting the optimal scheduling action from the discrete scheduling action set based on the local load forecasting results and real-time status parameters using a deep Q-network, and generating a local scheduling instruction from the optimal scheduling action, specifically involves:

[0103] The current time slice prediction value from the local load forecast results, the current value, voltage value and power value from the energy consumption monitoring data, the temperature, humidity and light intensity from the environmental parameters, and the equipment start-up and shutdown status, load rate and energy storage device charge status from the equipment operation status data are spliced ​​together in a preset order to form a state vector.

[0104] During the forward propagation of the deep Q-network, the Q-values ​​of candidate scheduling actions in the discrete scheduling action set are calculated based on the state vector. The Q-value calculation process of the deep Q-network includes: inputting the state vector into the deep Q-network, performing forward computation through fully connected layers and nonlinear activation functions to obtain the hidden layer representation; performing linear mapping in the output layer, with the dimension of the output vector consistent with the size of the discrete scheduling action set, and each component in the output vector corresponding to the Q-value of each candidate scheduling action in the discrete scheduling action set, forming a set of Q-values ​​for candidate scheduling actions; calculating the set of Q-values ​​for candidate scheduling actions at the next time step on the target network based on the state vector at the next time step and taking the maximum value as the target Q-value; using the difference between the Q-values ​​of candidate scheduling actions and the target Q-value as the loss function, performing backpropagation and parameter update; after the parameter update is completed, outputting the set of Q-values ​​for candidate scheduling actions through the forward computation process to determine the optimal scheduling action.

[0105] The discrete scheduling action set includes energy equipment start-stop actions corresponding to scheduling action identifiers, energy equipment power level switching actions, energy storage device charging actions, energy storage device discharging actions, energy storage device state maintenance actions, and control actions based on energy equipment operating mode switching.

[0106] Compare the Q values ​​of each candidate scheduling action and determine the candidate scheduling action with the largest Q value as the optimal scheduling action;

[0107] The optimal scheduling action is converted into a local scheduling instruction according to a preset format. The local scheduling instruction includes a scheduling action identifier and corresponding execution parameters.

[0108] The execution parameters are control values ​​and control configurations corresponding to the scheduling action identifier, specifically including target power value, power level number, execution time, duration, charging and discharging power, and target state of charge of the energy storage device.

[0109] In the process of efficient energy dispatching, this invention constructs a state vector based on local load forecasting results and multi-source operating status data. A deep Q-network is used to evaluate the Q-values ​​of each candidate dispatching action in the discrete dispatching action set. A fully connected network and a nonlinear activation function are used to achieve deep characterization of state features. The output layer outputs a Q-value vector consistent with the size of the action set. A target network is then used to calculate the target Q-value and perform parameter updates, ultimately determining the optimal dispatching action with the largest Q-value. This optimal dispatching action is then transformed into a local dispatching instruction, which includes a dispatching action identifier and its control parameters. This enables start-stop control, power switching, and operating mode adjustment of energy equipment and energy storage devices, significantly improving the intelligence and response efficiency of the dispatching process.

[0110] In this embodiment, the step of sending the local scheduling command to the energy equipment specifically involves:

[0111] Local scheduling instructions are encapsulated at each edge node. The encapsulation content includes scheduling action identifier, execution parameters, target energy device identifier, execution time parameter, and duration parameter.

[0112] Select a communication link with the target energy equipment and establish a communication connection; the communication link includes two types: wired communication link and wireless communication link. The wired communication link includes point-to-point or bus-type connection methods established through physical wires or optical fibers, specifically including Ethernet links, RS485 buses, CAN buses, optical fiber communication links, and other wired interfaces that conform to field industrial communication standards.

[0113] The wireless communication links include communication methods established through radio signals, specifically including Wi-Fi communication links, Bluetooth or Bluetooth Low Energy communication links, ZigBee communication links, LoRa communication links, cellular mobile communication links, and other short-range or wide-area wireless communication protocols.

[0114] Local dispatch commands are sent to the target energy equipment via the communication connection.

[0115] The target energy equipment returns a confirmation message indicating receipt of the instruction.

[0116] The target energy device parses the scheduling action identifier and execution parameters according to the local scheduling instructions, and outputs operating status information and execution result information to each edge node. The execution result information is the feedback content after the scheduling action is completed, specifically including the device's execution status identifier, the actual completion status of the execution action, the real-time measurement value after the target power value or power level switch, the change in state of charge of the energy storage device after the charging and discharging action is completed, and the deviation information between the parameter and the preset execution parameters. Each edge node receives and records the execution result information to form a scheduling execution record, and provides a basis for subsequent model parameter updates and scheduling strategy optimization.

[0117] Each edge node records its running status information and execution result information, and generates a scheduling execution record.

[0118] In this embodiment, within a preset period, each edge node uploads the improved LightGBM model parameters and the deep Q-network parameters to the federated aggregation server. The federated aggregation server performs parameter aggregation operations to generate a global energy scheduling model, specifically as follows:

[0119] Within a preset period, each edge node uploads its local improved LightGBM model parameters and local deep Q network parameters to the federated aggregation server;

[0120] The local deep Q-network parameters include the weight parameters and bias parameters of each layer;

[0121] The federated aggregation server receives the local improved LightGBM model parameters and local deep Q network parameters uploaded by each edge node and establishes a parameter index table;

[0122] The federated aggregation server performs hierarchical tree-structured aggregation operations on the improved LightGBM model, including:

[0123] The nodes at the same level are aligned according to the tree structure hierarchy number. At higher-level nodes, a majority vote is performed based on the split feature index to determine the global split feature index. At higher-level nodes, the global split threshold is determined based on the split threshold range according to the preset threshold selection rule to generate the global tree structure. The federated aggregation server performs aggregation at the leaf nodes based on the leaf node weight parameters uploaded by each edge node to generate the global improved LightGBM model.

[0124] The federated aggregation server performs parameter aggregation operations on the deep Q network, including: aligning the weight parameters and bias parameters of each layer according to the network hierarchy, and performing aggregation to obtain the global deep Q network;

[0125] The global energy scheduling model generated by the federated aggregation server includes a global improved LightGBM model and a global deep Q network model.

[0126] This invention encapsulates generated local scheduling commands into a standard format at edge nodes and reliably distributes them to target energy devices via wired or wireless communication links, enabling refined execution of operations such as device start-up and shutdown, power switching, and energy storage control. Simultaneously, edge nodes receive operational status information and execution feedback from the devices, forming a complete scheduling execution record, thus improving the traceability and system stability of the scheduling loop. Furthermore, within a preset period, this invention uses a federated learning mechanism to structurally aggregate the improved LightGBM model parameters and deep Q-network parameters uploaded by each edge node. Employing hierarchical decision tree fusion and neural network weight alignment methods, it constructs a global load prediction model and a global policy decision model, ultimately generating an integrated global energy scheduling model, achieving a unified approach of multi-node collaborative optimization and global policy consistency.

[0127] In this embodiment, the step of compressing and optimizing the global energy scheduling model based on the knowledge distillation mechanism and the model compression mechanism, and then distributing the compressed and optimized global energy scheduling model to each edge node, specifically involves:

[0128] Using the global energy scheduling model as the teacher model, a distillation training dataset is constructed, and the distillation training dataset is input into the teacher model to generate soft-labeled output results.

[0129] A student model is constructed in a federated aggregation server, the student model comprising an improved LightGBM model and a deep Q-network model;

[0130] Input the distillation training dataset into the student model to obtain the prediction results of the student model;

[0131] In the federated aggregation server, the difference between the soft-label output of the teacher model and the prediction of the student model is used as the distillation loss to perform parameter optimization and update the parameters of the improved LightGBM model and the deep Q network.

[0132] After completing the knowledge distillation operation, the federated aggregation server performs model compression operation on the improved LightGBM model and the deep Q network model. The model compression operation includes pruning, quantization and redundant parameter removal to obtain a compressed and optimized global energy scheduling model.

[0133] The federated aggregation server distributes the compressed and optimized global energy scheduling model to each edge node to replace and update the local improved LightGBM model and the local deep Q network model in the edge nodes.

[0134] This invention introduces knowledge distillation and model compression mechanisms to efficiently optimize the global energy scheduling model generated by federated aggregation, further improving its adaptability and operational efficiency for edge deployment. Specifically, firstly, the global energy scheduling model is used as the teacher model to construct a distillation training dataset and generate soft-label output results for candidate scheduling actions. Then, based on a student model containing an improved LightGBM model and a deep Q-network model, the distillation loss between the prediction results and the soft labels is calculated, completing parameter optimization and model transfer. Subsequently, the federated aggregation server performs compression operations such as pruning, quantization, and redundant parameter removal on the optimized student model to obtain a simplified global energy scheduling model, which is then distributed to each edge node to replace the original local model. This achieves a balance between model inference efficiency and resource consumption, ensuring intelligent scheduling capabilities in edge computing environments.

[0135] In this embodiment, the generation of soft tag output results specifically refers to:

[0136] The running status data in the distillation training dataset is input into the global improved LightGBM model to obtain the load prediction value sequence for the corresponding time slice. The load prediction value sequence is then normalized by temperature coefficient to generate soft labels in the form of prediction probability distribution.

[0137] The load forecast sequence, current, voltage and power values ​​from energy consumption monitoring data, temperature, humidity and light intensity from environmental parameters, and equipment start-up and shutdown status, load rate and energy storage device state of charge from equipment operation status data are spliced ​​together in a preset order to form a distillation state vector.

[0138] The distillation state vector is input into the global deep Q-network model to obtain the Q-value set corresponding to each candidate scheduling action in the discrete scheduling action set. Softmax normalization is then performed on the Q-value set to generate soft labels in the form of the probability distribution of candidate scheduling actions.

[0139] The soft label in the form of the predicted probability distribution and the soft label in the form of the candidate scheduling action probability distribution together constitute the soft label output result.

[0140] The prediction results of the student model are as follows:

[0141] The standardized feature data from the distillation training dataset is input into the improved LightGBM model. Node splitting path traversal is performed in each decision tree to determine the corresponding leaf node output value. Combination operation is then performed on the leaf node output values ​​to generate the predicted value sequence for the corresponding time slice.

[0142] The distillation state vectors from the distillation training dataset are input into a lightweight deep Q-network model. Forward computation is performed sequentially through fully connected layers and nonlinear activation functions to obtain the hidden layer representation. Then, a linear mapping is performed on the output layer to generate the Q-value set corresponding to each candidate scheduling action in the discrete scheduling action set.

[0143] The student model prediction result is composed of the predicted value sequence and the Q-value set.

[0144] This invention achieves knowledge transfer and lightweight reconstruction of a global energy scheduling model by constructing a refined soft-label generation and student model prediction mechanism. Specifically, the operating state data from the distillation training dataset is first input into a global improved LightGBM model and a global deep Q-network model, respectively, generating load prediction sequences and Q-value sets of candidate scheduling actions. These are then processed using temperature coefficient normalization and Softmax normalization to construct soft labels in the form of prediction probability distributions and scheduling action probability distributions, serving as the output benchmark for the teacher model. Subsequently, a student model, composed of the improved LightGBM model and the lightweight deep Q-network model, is used to infer the standardized feature data and distillation state vectors from the distillation training dataset, generating corresponding prediction sequences and candidate action Q-value sets as the student model's prediction results. The soft-label output results and the student model prediction results are used together for distillation loss calculation and parameter updates, ensuring that the compressed model maintains performance while possessing higher adaptability for edge deployment.

[0145] In this embodiment, the improved LightGBM model performs a convention processing on multi-timescale features during the input construction stage: after generating standardized feature data, the short-term state sequence and the long-term trend sequence are combined into a multi-timescale input matrix according to a preset splicing order. It is agreed that the feature columns corresponding to the short-term state sequence are located in a continuous column interval, and the feature columns corresponding to the long-term trend sequence are located in another continuous column interval. The column interval and column index range are fixed in the model configuration for subsequent source determination and output classification of split feature indexes.

[0146] During training, splitting information is recorded: when splitting is performed at each decision tree node, the splitting feature index is used to determine whether the splitting feature belongs to the short-term state sequence interval or the long-term trend sequence interval. The splitting feature index and splitting threshold interval are written into the tree structure, and the splitting feature source information is retained when traversing the node splitting path. This allows the samples falling into the leaf node to be classified into the short-term source group or the long-term source group according to the source of the splitting feature in the splitting path. At the same time, the tree structure level number, splitting feature index, splitting threshold interval and leaf node weight parameters are generated to form the parameters of the locally improved LightGBM model.

[0147] In the prediction output stage, the output values ​​of the leaf nodes are grouped and combined: after inputting the multi-timescale input matrix into the model, the node splitting path is traversed in each decision tree and the output values ​​of the leaf nodes are determined. Based on the splitting feature source information recorded during the training phase, the output values ​​of the leaf nodes are assigned to the short-term source group and the long-term source group respectively. The output values ​​of the leaf nodes of the short-term source group and the output values ​​of the leaf nodes of the long-term source group are combined according to the preset weighting ratio to obtain the predicted values ​​of the corresponding time slices. The predicted values ​​of the continuous time slices are arranged in chronological order to generate the local load prediction results.

[0148] In the dataset partitioning and hyperparameter setting stages, time order constraints are maintained: the multi-timescale input matrix is ​​partitioned into training and validation sets according to time order, and the objective function, learning rate, maximum number of leaves, maximum depth, subsampling ratio, and feature subsampling ratio are set. Local training rounds are performed in combination with fixed short-term state sequence intervals and long-term trend sequence intervals. The output includes locally improved LightGBM model parameters containing tree structure level numbers, split feature indices, split threshold intervals, and leaf node weight parameters, which are used for subsequent federated aggregation and updates.

[0149] This invention proposes a structured improvement scheme for handling multi-timescale features in the improved LightGBM model. By fixing the column intervals of short-term state sequences and long-term trend sequences during the input stage, recording the source of split features and generating tree structure level numbers, split feature indices, split threshold intervals, and leaf node weight parameters during the training stage, and grouping and combining the output values ​​of leaf nodes based on the split source information and fusing them according to a preset weighting ratio to generate prediction results during the prediction stage, while maintaining time order constraints during dataset partitioning and hyperparameter setting, the output parameters of the local improved LightGBM model are traceable and structurally consistent, which can be used for subsequent federated aggregation and updates.

[0150] Example 1:

[0151] To verify the feasibility of this invention in practice, it was applied to a comprehensive energy management scenario in a large industrial park. This industrial park is equipped with various energy devices, including gas-fired generator sets, distributed photovoltaic systems, energy storage battery packs, and several key energy load points. Due to the significant peak-valley electricity consumption differences in the industrial park's production model, coupled with issues such as fluctuating environmental conditions, frequent equipment start-ups and shutdowns, and unreasonable charging and discharging strategies for energy storage devices, energy dispatching often suffers from large load forecasting deviations, untimely dispatching actions, low equipment operating efficiency, and insufficient energy utilization. Traditional centralized energy dispatching methods require uploading large amounts of real-time operational data to a central server, which not only increases communication overhead but also leads to delays in dispatching instructions, making it difficult to meet the real-time dispatching needs of complex scenarios.

[0152] In this scenario, the method of this invention is deployed on multiple edge nodes in the park. These edge nodes collect energy consumption monitoring data, environmental parameters, and equipment operating status data, such as current, voltage, power, temperature, humidity, light intensity, equipment start-up / shutdown status, load rate, and energy storage state of charge. Through data preprocessing, missing value imputation, outlier removal, and normalization are performed to generate standardized feature data. Based on this, the edge nodes run an improved LightGBM model, utilizing a multi-timescale input matrix to simultaneously consider short-term state sequences and long-term trend sequences, predicting the load for the next 15, 30, and 60 minutes. The average error rate of the prediction results is controlled within 5%, which is approximately 3 percentage points lower than the traditional LightGBM model. Subsequently, a deep Q-network constructs a state vector based on the predicted values ​​and real-time status, and combines this with a discrete scheduling action set to optimize equipment start-up / shutdown, power level switching, and energy storage charging / discharging. Through forward propagation and Q-value comparison, the system can quickly output the optimal scheduling action and convert it into local scheduling commands to be sent to the energy equipment, achieving a second-level response. Specific experimental data are shown in Table 1.

[0153] Table 1 Comparative Experimental Data on Energy Dispatch in Industrial Parks

[0154] Time interval Average load prediction error (%) Scheduling response delay (s) Energy storage state of charge reasonable interval maintenance rate (%) Equipment invalid start-stop times (times / day) Energy utilization rate improvement (%) Traditional centralized method (control group) 8.2 2.5 68 15 — Improved LightGBM+local scheduling 5.1 1.7 81 10 4.0 Improved LightGBM+deep Q network scheduling 4.9 1.2 85 9 5.2 Inventive whole process+federated learning aggregation 4.6 1.0 86 8 6.0 Inventive whole process+distillation and model compression optimization 4.5 0.9 87 7 6.5

[0155] During the experiment, parameters were uploaded and aggregated between the edge nodes and the federated aggregation server every 30 minutes. Table 1 shows that when 10 edge nodes participated in training simultaneously, the prediction accuracy of the globally improved LightGBM model improved by approximately 7% compared to single-point training, and the scheduling success rate of the global deep Q network improved by approximately 12%. In the compression optimization stage, through knowledge distillation and pruning operations, the model size was compressed by approximately 45%, enabling the global energy scheduling model distributed to the edge nodes to run stably on embedded platforms with limited computing resources, and reducing the average prediction and scheduling latency by 0.8 seconds. In a month-long actual operation verification, application data from typical workdays and peak load days showed that energy utilization improved by 6.5%, the proportion of time the energy storage device's state of charge remained within a reasonable range increased from 68% to 87%, and the number of ineffective start-ups and shutdowns of the equipment decreased by approximately 40%, effectively reducing equipment wear and tear.

[0156] The above application results demonstrate that the present invention can achieve high-precision load forecasting and intelligent energy dispatching in complex multi-source environments, avoiding the excessive communication overhead and dispatching lag problems caused by traditional centralized architectures. At the same time, it maintains low computational burden and high dispatching efficiency in edge environments, fully meeting the needs of refined energy management in industrial parks.

[0157] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for efficient energy scheduling based on edge computing, characterized in that, Includes the following steps: Step 1: Collect operational status information at edge nodes, perform preprocessing operations, and generate standardized feature data; Step 2: Run the improved LightGBM model based on the standardized feature data to generate local load forecasting results, and upload the local improved LightGBM model parameters to the federated aggregation server; Step 3: Based on the local load forecast results and real-time status parameters, select the optimal scheduling action from the discrete scheduling action set based on the deep Q network, and generate local scheduling instructions from the optimal scheduling action; Step 4: Send local dispatch instructions to energy equipment; Step 5: Within a preset period, each edge node uploads the deep Q network parameters to the federated aggregation server. The federated aggregation server performs parameter aggregation operations to generate a global energy scheduling model. Step 6: Compress the global energy scheduling model based on the knowledge distillation mechanism and the model compression mechanism, and distribute the compressed and optimized global energy scheduling model to each edge node.

2. The energy-efficient scheduling method based on edge computing according to claim 1, characterized in that, The process of collecting operational status information at edge nodes and performing preprocessing operations specifically includes: At each edge node, the operating status information collected by energy consumption monitoring equipment, environmental sensing equipment and energy equipment deployed on site is acquired. The operating status information includes energy consumption monitoring data, environmental parameters and equipment operating status data. The preprocessing operation includes: performing missing value imputation, outlier removal, and data format unification on the original running status information to obtain cleaned running status data; The cleaned operational status data is used to construct feature vectors according to preset feature dimensions and time window division methods. The feature vectors include short-term status sequences and long-term trend sequences constructed based on sliding time windows. Normalization and dimension alignment operations are performed on the feature vectors to generate standardized feature data.

3. The energy-efficient scheduling method based on edge computing according to claim 1, characterized in that, The process involves running an improved LightGBM model based on standardized feature data to generate local load forecasting results, and then uploading the parameters of the local improved LightGBM model to the federated aggregation server. Specifically: The short-term state sequence and the long-term trend sequence are combined into a multi-time-scale input matrix according to a preset splicing order; The multi-timescale input matrix is ​​divided into training and validation sets according to time order. The training hyperparameters of the improved LightGBM model are set, and local training rounds are executed to obtain the parameters of the local improved LightGBM model. The parameters of the locally improved LightGBM model include tree structure level number, split feature index, split threshold range and leaf node weight parameters. Training hyperparameters include the objective function, learning rate, maximum number of leaves, maximum depth, subsampling ratio, and feature subsampling ratio; The multi-timescale input matrix is ​​input into the improved LightGBM model, and node splitting path traversal is performed on the multi-timescale input matrix in each decision tree to determine the corresponding leaf node output value. The output values ​​of the decision tree leaf nodes corresponding to the short-term state sequence and the output values ​​of the decision tree leaf nodes corresponding to the long-term trend sequence are calculated separately and weighted, and then combined according to the preset weighting ratio to obtain the predicted value of the corresponding time slice. Arrange consecutive forecast values ​​in chronological order to generate local load forecast results; Within a preset period, each edge node uploads the parameters of its local improved LightGBM model to the federated aggregation server.

4. The energy-efficient scheduling method based on edge computing according to claim 1, characterized in that, Based on local load forecasting results and real-time status parameters, the optimal scheduling action is selected from the discrete scheduling action set using a deep Q-network, and the optimal scheduling action is used to generate a local scheduling instruction. Specifically: The current time slice prediction value from the local load forecast results, the current value, voltage value and power value from the energy consumption monitoring data, the temperature, humidity and light intensity from the environmental parameters, and the equipment start-up and shutdown status, load rate and energy storage device charge status from the equipment operation status data are spliced ​​together in a preset order to form a state vector. During the forward propagation of the deep Q-network, the Q-values ​​of candidate scheduling actions in the discrete scheduling action set are calculated based on the state vector; The discrete scheduling action set includes energy equipment start-stop actions corresponding to scheduling action identifiers, energy equipment power level switching actions, energy storage device charging actions, energy storage device discharging actions, energy storage device state maintenance actions, and control actions based on energy equipment operating mode switching. Compare the Q values ​​of each candidate scheduling action and determine the candidate scheduling action with the largest Q value as the optimal scheduling action; The optimal scheduling action is converted into a local scheduling instruction according to a preset format. The local scheduling instruction includes a scheduling action identifier and corresponding execution parameters.

5. The energy-efficient scheduling method based on edge computing according to claim 1, characterized in that, The process of sending local dispatch instructions to energy equipment specifically involves: Local scheduling instructions are encapsulated at each edge node. The encapsulation content includes scheduling action identifier, execution parameters, target energy device identifier, execution time parameter, and duration parameter. Select a communication link with the target energy equipment and establish a communication connection; The local scheduling command is sent to the target energy equipment via the communication connection. The target energy equipment returns a confirmation message indicating receipt of the instruction. The target energy equipment parses the scheduling action identifier and execution parameters according to the local scheduling instructions, and outputs the operating status information and execution result information to each edge node; Each edge node records its running status information and execution result information, and generates a scheduling execution record.

6. The energy-efficient scheduling method based on edge computing according to claim 1, characterized in that, Within a preset period, each edge node uploads its deep Q-network parameters to the federated aggregation server. The federated aggregation server then performs parameter aggregation to generate a global energy scheduling model, specifically: Within a preset period, each edge node uploads its local deep Q network parameters to the federated aggregation server; The local deep Q-network parameters include the weight parameters and bias parameters of each layer; The federated aggregation server receives the local improved LightGBM model parameters and local deep Q network parameters uploaded by each edge node and establishes a parameter index table; The federated aggregation server performs hierarchical tree-structured aggregation operations on the improved LightGBM model, including: The nodes at the same level are aligned according to the tree structure hierarchy number. At higher-level nodes, a majority vote is performed based on the split feature index to determine the global split feature index. At higher-level nodes, the global split threshold is determined based on the split threshold range according to the preset threshold selection rule to generate the global tree structure. The federated aggregation server performs aggregation at the leaf nodes based on the leaf node weight parameters uploaded by each edge node to generate the global improved LightGBM model. The federated aggregation server performs parameter aggregation operations on the deep Q network, including: aligning the weight parameters and bias parameters of each layer according to the network hierarchy, and performing aggregation to obtain the global deep Q network; The global energy scheduling model generated by the federated aggregation server includes a global improved LightGBM model and a global deep Q network model.

7. The energy-efficient scheduling method based on edge computing according to claim 1, characterized in that, The global energy scheduling model is compressed based on knowledge distillation and model compression mechanisms, and the compressed and optimized global energy scheduling model is then distributed to each edge node. Specifically: Using the global energy scheduling model as the teacher model, a distillation training dataset is constructed, and the distillation training dataset is input into the teacher model to generate soft-labeled output results. A student model is constructed in a federated aggregation server, the student model comprising an improved LightGBM model and a deep Q-network model; Input the distillation training dataset into the student model to obtain the prediction results of the student model; In the federated aggregation server, the difference between the soft-label output of the teacher model and the prediction of the student model is used as the distillation loss to perform parameter optimization and update the parameters of the improved LightGBM model and the deep Q network. After completing the knowledge distillation operation, the federated aggregation server performs model compression operation on the improved LightGBM model and the deep Q network model. The model compression operation includes pruning, quantization and redundant parameter removal to obtain the compressed global energy scheduling model. The federated aggregation server distributes the compressed and optimized global energy scheduling model to each edge node to replace and update the local improved LightGBM model and the local deep Q network model in the edge nodes.

8. The energy-efficient scheduling method based on edge computing according to claim 7, characterized in that, The specific output of the generated soft tag is as follows: The standardized feature data from the distillation training dataset is input into the global improved LightGBM model to obtain the load prediction sequence for the corresponding time slice. The load prediction sequence is then normalized by temperature coefficient to generate soft labels in the form of prediction probability distribution. The load forecast sequence, current, voltage and power values ​​from energy consumption monitoring data, temperature, humidity and light intensity from environmental parameters, and equipment start-up and shutdown status, load rate and energy storage device state of charge from equipment operation status data are spliced ​​together in a preset order to form a distillation state vector. The distillation state vector is input into the global deep Q-network model to obtain the Q-value set corresponding to each candidate scheduling action in the discrete scheduling action set. Softmax normalization is then performed on the Q-value set to generate soft labels in the form of the probability distribution of candidate scheduling actions. The soft label in the form of the predicted probability distribution and the soft label in the form of the candidate scheduling action probability distribution together constitute the soft label output result.

Images