Multi-agent cooperative power distribution network load prediction method and system

Abstract

The application relates to the technical field of power distribution network load prediction, in particular to a multi-agent cooperative power distribution network load prediction method and system. The method obtains a communication frequency index, a communication data volume index, a communication time delay index and a node calculation energy consumption index, constructs a joint objective function containing a communication total cost, a prediction accuracy penalty term and a calculation energy consumption penalty term, and introduces the communication time delay as a hard constraint condition; the objective function is distributedly solved by using an alternating direction multiplier method; under the premise of meeting the time delay constraint, local variables are alternately updated to obtain an optimal communication frequency, an optimal data compression strategy and an optimal calculation resource configuration. Through three-dimensional joint optimization of "communication-computation-accuracy", the communication bandwidth occupation and node energy consumption of the power distribution network edge side are significantly reduced while ensuring the prediction accuracy, and the real-time performance and stability of the multi-agent cooperative system are improved.

Description

Technical Field

[0001] This application relates to the field of distribution network load forecasting technology, specifically to a multi-agent collaborative distribution network load forecasting method and system. Background Technology

[0002] Current power distribution network load forecasting technologies are gradually evolving from traditional centralized architectures to distributed collaborative architectures. With the advancement of smart grid construction, distributed deep learning and multi-agent systems are widely used at the edge. Each node agent trains a local model using local measurement data and achieves global knowledge sharing and collaborative updates by interacting with neighboring nodes to exchange model parameters or gradient information. This approach effectively avoids the direct uploading of raw sensitive data, protects user privacy to a certain extent, and reduces the computational burden on the central server, becoming the mainstream technical approach for improving the accuracy of low-voltage power distribution network load forecasting.

[0003] However, despite the advantages of distributed architecture in data privacy protection, it still faces significant communication bottlenecks in practical applications. In low-voltage distribution networks with a large number of nodes, to ensure the convergence and accuracy of the prediction model, agents often need to exchange high-dimensional gradient or parameter information frequently. This frequent communication interaction consumes a large amount of communication bandwidth resources, leading to network congestion, increased transmission latency, and even packet loss. Especially in edge environments with limited communication conditions, the high communication overhead not only reduces the system's real-time response capability but may also cause collaborative prediction failure due to communication blockage, severely restricting the large-scale deployment and application of distributed load forecasting technology. Summary of the Invention

[0004] In view of the above, it is necessary to provide a multi-agent collaborative method and system for power distribution network load forecasting to solve the above problems.

[0005] The first aspect of this application provides a multi-agent collaborative method for distribution network load forecasting, the method comprising: The communication frequency index is obtained based on the change in the gradient of the local model and the event triggering threshold; the communication data volume index is obtained based on the gradient sparsity ratio and the quantization bit width; the communication latency index is obtained based on the communication data volume index and the pre-acquired network state parameters; and the node computing energy consumption index is obtained based on the algorithm complexity of the data compression step. A joint objective function is established, which includes the communication frequency index, communication data volume index, node computing energy consumption index, and prediction accuracy index. The communication delay index is introduced into the joint objective function as a hard constraint. The joint objective function is solved in a distributed manner using the alternating direction multiplier method to obtain the optimal communication frequency, optimal data compression strategy, and optimal computing resource allocation for each node agent. Each node agent performs load prediction using its local deep learning model based on the optimal communication frequency and optimal data compression strategy, updates its local model parameters based on the optimal computing resource configuration, and outputs collaborative prediction results.

[0006] Preferably, the step of obtaining the communication frequency index based on the change in the local model gradient and the event triggering threshold specifically includes: Obtain the change range of the model parameters of the local deep learning model at the current time and the last time the communication was triggered; compare the change range with the event trigger threshold; if the change range is greater than or equal to the event trigger threshold, then record the trigger state at the current time as active, otherwise record it as silent; The number of times the communication is active within a preset time period is counted to obtain the communication frequency index.

[0007] Preferably, the step of obtaining the communication data volume index based on the gradient sparsity ratio and quantization bit width specifically includes: Obtain the total dimension of the model parameters to be transmitted; calculate the proportion of gradients that are not discarded based on the gradient sparsity ratio; determine the size of the compressed data bits for a single communication data packet based on the quantization bit width; multiply the total dimension, the proportion of gradients that are not discarded, and the quantization bit width to obtain the communication data volume index.

[0008] Preferably, the step of obtaining the communication latency index based on the communication data volume index and the pre-acquired network status parameters specifically includes: Obtain transmission delay, propagation delay, processing delay, and queuing delay; The transmission delay is obtained by dividing the communication data volume index by the network bandwidth, and the network bandwidth and queuing delay are prior constants measured and issued by the distribution network aggregation node in the previous communication cycle. The communication latency index is obtained by adding the transmission latency, propagation latency, processing latency, and queuing latency together.

[0009] Preferably, the algorithm complexity based on the data compression step is used to obtain the node computing energy consumption index, specifically as follows: The computational complexity of the data compression step is determined, and the computational complexity includes a preset basic computing power cost item, a first dynamic computing power cost item, and a second dynamic computing power cost item; Wherein, the first dynamic computing power cost term is inversely proportional to the proportion of gradients not discarded, and the second dynamic computing power cost term is inversely proportional to the quantization bit width; Based on the hardware environment parameters of the intelligent agent and the computational complexity of each computation step, the dynamic energy consumption of the node executing the computation task is calculated; the dynamic energy consumption is added to the basic idle energy consumption of the node to obtain the node's computational energy consumption index.

[0010] Preferably, the specific formula for the joint objective function is as follows: Describes the minimum value function; This represents the overall optimization objective value of the entire multi-agent cooperative prediction system; Indicates communication frequency specifications; Indicates the amount of communication data; Indicates the communication resource weighting coefficient; This serves as an approximate substitute for the global prediction error. For precision weighting coefficients; Calculate energy consumption indicators for nodes. Energy consumption weighting coefficient; where All are normalized weight coefficients, satisfying .

[0011] Preferably, the constraint condition of the objective function is that the communication delay index is less than or equal to the maximum allowable delay threshold for distribution network control services.

[0012] Preferably, the distributed solution of the joint objective function using the alternating direction multiplier method specifically includes: Construct an augmented Lagrange function by introducing Lagrange multipliers and penalty parameters; Each node agent alternately updates local variables, and the updating of local variables includes solving the communication and compression subproblems and the computation and prediction subproblems; By exchanging local variables and updating the Lagrange multipliers until convergence, the optimal event triggering threshold and the optimal data compression strategy can be obtained.

[0013] Preferably, the solution to the communication and compression subproblem specifically includes: The communication frequency index within the statistical period is transformed into a continuously decreasing fitting function about the event trigger threshold; Under the conditions of fixed computational energy consumption index and Lagrange multipliers, the desired communication frequency index that minimizes the total communication overhead is obtained by optimizing the event triggering threshold.

[0014] Secondly, embodiments of this application also provide a multi-agent collaborative power distribution network load forecasting system, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of any of the methods described above.

[0015] This application has at least the following beneficial effects: This application introduces a communication frequency metric (based on an event-triggered mechanism) and a communication data volume metric (based on adaptive gradient compression). Compared to traditional fixed-period communication, this application triggers communication only when local model parameters change significantly, and eliminates redundant data by dynamically adjusting the sparsity ratio and quantization bit width. This reduces the number and size of invalid data packets sent from the source, effectively alleviating bandwidth pressure on the edge side of the low-voltage distribution network and reducing network packet loss rate caused by data packet accumulation.

[0016] The communication latency metric (including transmission, propagation, processing, and queuing latency) is treated as a hard constraint on the objective function, rather than a simple penalty term. Utilizing the Lagrange multiplier feedback mechanism in the ADMM algorithm, the system can perceive network congestion in real time: when queuing latency increases exponentially due to high-frequency communication, the algorithm forcibly reduces the communication frequency. This ensures that critical load forecasting commands and control data are always transmitted within a safe latency range, avoiding system oscillations caused by communication delays and greatly improving the stability of the distribution network operation.

[0017] A multi-dimensional joint objective function was constructed, encompassing communication overhead, prediction accuracy, and computational energy consumption. By introducing weighting coefficients, the system can find the optimal balance between "reducing communication overhead" and "increasing computational complexity," thus avoiding the drawbacks of simply squeezing the computing power of edge devices to save communication or blindly transmitting the entire data for high accuracy, achieving optimal allocation of global resources.

[0018] A node computing energy consumption index based on a semiconductor dynamic power consumption model reconstruction is introduced. By quantifying the physical energy consumption cost caused by the complexity of data compression algorithms, the system can guide edge nodes to select the compression strategy with the lowest computational cost while meeting communication requirements; effectively reducing the CPU load and heat generation of edge devices, and extending the lifespan of battery-powered devices, making it particularly suitable for large-scale deployment of distribution network IoT scenarios.

[0019] This application establishes a dynamic closed-loop feedback mechanism of "prediction-evaluation-re-triggering". When the distribution network load fluctuates drastically (such as power surges caused by renewable energy grid connection), the system can automatically detect prediction errors and trigger a new round of ADMM collaborative optimization. This mechanism enables the system to adapt to the complex time-varying environment of the distribution network in real time, maintain high-precision prediction capabilities, and solve the shortcomings of traditional static configuration methods that cannot adapt to dynamic network environments. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the steps of a multi-agent collaborative distribution network load forecasting method provided in one embodiment of this application. Detailed Implementation

[0021] The following description, in conjunction with the accompanying drawings, details the specific scheme of the multi-agent collaborative distribution network load forecasting method and system provided in this application.

[0022] Please see Figure 1 The diagram illustrates a flowchart of a multi-agent cooperative distribution network load forecasting method according to an embodiment of this application. The method includes the following steps: The first step: Obtain communication frequency indicators based on the change in local model gradients and event triggering thresholds.

[0023] Existing multi-agent cooperative prediction methods mostly employ a time-triggered communication mechanism, meaning that regardless of whether the local model changes significantly, the agent will forcibly send data at fixed time steps. During periods when the distribution network load is relatively stable, this communication is entirely redundant.

[0024] This application aims to quantify and control the execution effect of an "Event-Triggered Mechanism" at the algorithmic level. By monitoring the magnitude of changes in the local model's gradient in real time, communication is triggered only when the change exceeds a threshold, thereby controlling the communication frequency index. It is incorporated into the control system as a dynamic variable, rather than a fixed system parameter.

[0025] Communication frequency indicators In the statistical period The specific expression inside is: ; This metric is used to quantify the "necessity of communication." Through this indicator, the system can eliminate redundant and low-information interactions, reducing the number of communications over time.

[0026] In the formula for communication frequency indicators, the trigger function The logical judgment is as follows: Current moment intelligent agent Local model parameters; Model parameters at the last time communication was triggered; The preset event trigger threshold, σ, is set by the system based on the statistical mean of the historical model update gradient, with a reference range of 0.01 to 0.05. : Represents the norm of a vector (such as the L2 norm), used to measure the magnitude of parameter changes.

[0027] In this application, the system does not communicate according to a fixed time step, but instead monitors the update magnitude of the local model parameters in real time. Communication only occurs when the local model has learned significantly new knowledge. Only when set to 1 will a communication be triggered; otherwise, silence will be maintained. This application utilizes communication frequency indicators. The originally uncontrollable and mechanical communication behavior is transformed into an intelligent variable that can be described by mathematical models and dynamically adjusted by optimization algorithms. This can directly cut off a large amount of invalid data interaction from the time dimension, significantly reduce the bandwidth occupation of the power distribution network communication channel, and alleviate network congestion.

[0028] The second step: Obtain communication data volume indicators based on gradient sparsity ratio and quantization bit width.

[0029] Modern power distribution network load forecasting often employs deep learning models (such as LSTM and CNN), which contain massive amounts of weight parameters and high-dimensional gradients. In a low-voltage power distribution network with a large number of nodes, if all of this raw data is transmitted in full for each coordination, it will instantly fill the already limited communication bandwidth, leading to severe network congestion.

[0030] Existing distributed algorithms often employ fixed data precision and transmission strategies, which cannot adapt to the real-time fluctuating communication conditions of power distribution networks. Therefore, this application obtains communication data volume indicators. This is to quantify and control the execution effect of "adaptive gradient compression (such as sparsification and quantization)" at the algorithm level. The size of the data packet in a single communication is incorporated as a dynamic variable into the control system, allowing it to elastically scale according to network congestion levels.

[0031] Single communication data volume index The calculation formula is: ; : Total dimension (total number) of the parameters to be transmitted in the model; Gradient sparsity ratio ( ), indicating the proportion of minor gradients discarded; Quantization bit width (e.g., compressing a 32-bit floating-point number to an 8-bit integer). (corresponding to the number of bits after compression), used to determine the size of the compressed data bits in a single communication data packet.

[0032] The computing resources of the edge nodes of the distribution network are limited, and this application acquires and controls... This means that the system must complete data filtering and compression locally before performing communication. This not only reduces the transmission power consumption of the RF module, but also avoids the high CPU load caused by packaging and unpacking all data, allowing edge devices to concentrate more computing power on load prediction itself.

[0033] The third step: Obtain the communication latency index based on the communication data volume index and the pre-acquired network status parameters.

[0034] In the first two steps, this application reduces the communication frequency ( ) and data volume ( To save bandwidth, some power distribution network communication channels (such as 4G / 5G or carrier communication) are time-varying. If only the data volume is considered while ignoring the channel queuing time, critical control commands (such as emergency load shedding) may be blocked, causing system oscillations.

[0035] Obtain communication latency metrics This is to introduce "timeliness constraints" into the ADMM distributed optimization algorithm and quantify the end-to-end time for data to travel from the edge node to the cloud / aggregation node for reception and processing. This enables the system to distinguish between "latency-sensitive control commands" and "latency-insensitive statistical data," thereby allowing for tiered transmission.

[0036] In the context of multi-agent collaboration in power distribution networks, communication latency metrics Queuing latency cannot be simply regarded as a fixed value, but rather a dynamic cumulative value composed of four parts: transmission, propagation, processing, and queuing. Especially in high-concurrency or bandwidth-constrained power distribution network communication environments (such as HPLC, 4G / 5G slicing), queuing latency is often the main cause of system oscillation.

[0037] In this application, the formula is used. Obtain communication latency metrics, including: , Indicates transmission delay, used to describe the time required to push data packets onto the communication channel; Indicates the amount of communication data; Indicates the link bandwidth.

[0038] , L represents the propagation delay, used to quantify the time required for electromagnetic waves to propagate in a physical medium (optical fiber, twisted pair, air); L represents the physical distance of the communication link; v represents the signal propagation speed (speed of light or speed of radio waves).

[0039] This represents processing latency, used to quantify the time consumed by an agent or router in receiving, verifying, parsing data packet headers, and performing encryption / decryption operations.

[0040] Queuing delay, used to describe the time a data packet waits to be forwarded in the output queue of a network node (such as a switch, base station, or aggregator). (This is related to communication frequency metrics.) Highly correlated. When When excessive levels cause network congestion, It will increase exponentially, which is the main factor leading to instability in distribution network control.

[0041] It is important to note that and The measurement or detection is performed by the distribution network edge router or aggregation node in the previous communication cycle and sent to each local agent through in-band signaling or broadcast messages as a priori constant for the current calculation cycle.

[0042] Based on the above formula, it can be understood that the communication latency index... Communication data volume indicators There is a significant positive linear coupling relationship between them. According to the physical definition of transmission delay, the amount of communication data... This constitutes a direct determinant of transmission delay. (In link bandwidth) Under constant conditions, reduce the amount of data in a single communication by using an adaptive compression algorithm. It can linearly reduce the channel dwell time of data packets (i.e., transmission delay). This will enable the achievement of end-to-end communication latency metrics. This mechanism provides a deterministic physical guarantee for alleviating bandwidth bottlenecks at the edge of the distribution network.

[0043] Communication latency metrics With communication frequency indicators The relationship between them exhibits a highly complex nonlinear dynamic game, and its core mechanism lies in the sensitive response of network queuing delay to communication frequency.

[0044] In low-load range: when the communication frequency When the network throughput is below the critical threshold, the communication channel is in a light-load state, and the queuing delay is... Approaching zero, at this point Primarily dominated by deterministic factors such as transmission delay and propagation delay, the system exhibits good delay predictability.

[0045] In high-load range: when communication frequency When the number of cases continues to rise and exceeds the network's critical threshold, network congestion occurs, resulting in queuing delays. Follow The increase exhibits an exponential, non-linear surge. This deterioration in non-linearity leads to a decline in communication latency metrics. It exhibits a severe long-tail effect and high jitter.

[0046] System-level impact: The long-tail effect not only significantly reduces the effective throughput of communication channels, but also disrupts the timing synchronization in the multi-agent collaborative prediction process, leading to a lag in global state information and consequently causing phase delays in control commands or even system oscillations. Therefore, it is essential to... Strict constraints are imposed below the network critical threshold to avoid the nonlinear deterioration range of queuing delay.

[0047] The fourth step: Obtain the node computing energy consumption index based on the algorithm complexity of the data compression step.

[0048] In order to achieve the aforementioned (Event triggering) and (Gradient compression) Edge devices (such as smart meters and DTUs) need to perform complex model inference and gradient calculations locally. For devices that rely on battery power or have difficulty obtaining power, the energy consumption from computation may far exceed the energy consumption from communication.

[0049] This application obtains node computing energy consumption indicators. This is used to quantify the physical cost of "algorithm complexity," making it easier to know how many joules of computing power the device consumes to save 1kb of communication, thus determining whether this compression strategy is worthwhile.

[0050] In a preferred embodiment of this application, the node calculates energy consumption indicators. The calculation follows the semiconductor dynamic power consumption model, whose basic physical expression is: Where K represents all computational steps included in a collaborative cycle, in this application For data preprocessing, For LSTM forward propagation, For gradient backpropagation, Compress the data; Indicates the first The dynamic power consumption of a task is the instantaneous power generated by the charging and discharging of the internal capacitors of transistors and the switching of logic gate states when the processor is executing a specific computing task. Indicates the first The static power consumption of a task is the basic power generated by semiconductor leakage current, basic clock maintenance and peripheral circuit standby when the device is powered on, even if no effective computing task is being performed. Indicates the first The execution time of each task.

[0051] Furthermore, in order to establish a direct mapping relationship between energy consumption and the computational complexity of edge-side AI algorithms to facilitate subsequent joint optimization, this embodiment expands and reconstructs the above physical model as follows: ;in, V represents the effective capacitance coefficient, a constant used to characterize the physical architecture characteristics of the internal processor (such as CPU, NPU, or DSP) chip of an intelligent agent, and is related to the chip's manufacturing process (such as 7nm, 14nm) and internal circuit structure; V represents the operating voltage, that is, the power supply voltage of the processor when it is performing its own calculation tasks. For the first The computational complexity required for each task; To execute the first Processor operating frequency for each task; This represents the basic idle energy consumption of a node, which is the basic energy consumption of the agent within a coordination cycle due to semiconductor leakage current, basic clock maintenance, and peripheral circuit standby. This energy consumption is independent of the specific AI computing task and only related to the total execution time of the task. The energy consumption formula... and When considered as hardware environment parameters, the formula clearly shows: The size is directly calculated by the algorithm for the load. Therefore, this application aims to reduce the total energy consumption of the system by optimizing the complexity of the AI ​​algorithm. This represents the dynamic energy consumption of a node performing computational tasks; Among them, the computational complexity of the data compression task Positively correlated with compression depth, specifically mapped as follows + ,in , , These are constant coefficients pre-calibrated based on the factory performance of the hardware equipment. The basic computing power cost item can usually be an integer constant ranging from 100 to 500, depending on the instruction set architecture of the intelligent agent's MCU. In this embodiment, the value is 300. The first dynamic computational cost to characterize sparsity; This represents the proportion of gradients discarded. The closer it is to 1, the more drastic the compression. However, to accurately select so many gradients to discard, more complex sorting and threshold search algorithms are needed, causing a non-linear increase in computational cost. When the discard ratio approaches 1, the denominator approaches 0, and the computational cost of this term increases rapidly. Characterizing the second dynamic computing cost of low-bit quantization; This refers to the number of bits retained. A smaller bit width results in a higher compression ratio, but to prevent model distortion, extremely low bit width quantization typically requires real-time calculation of dynamic scaling factors or asymmetric compensation. The algorithm complexity increases inversely with decreasing bit width. The system subsequently executes the ADMM algorithm to solve the cost function, directly outputting the numerical results of the percentage of gradients discarded and the quantization bit width within the current period.

[0052] This reconstructed formula transforms the previously abstract concept of "computational energy consumption" into an analytical expression that can be directly processed by mathematical optimization algorithms such as ADMM. The system can then accurately assess the impact of data compression. And increased "and due to reduced communication frequency" The quantitative game relationship between "saving communication energy consumption" and "accuracy" provides precise mathematical support for the three-dimensional joint optimization of "communication-computation-accuracy".

[0053] The fifth step is to establish a joint objective function that includes the communication frequency index, communication data volume index, node computing energy consumption index, and prediction accuracy index, and to introduce the communication delay index as a hard constraint into the joint objective function; to use the alternating direction multiplier method to solve the joint objective function in a distributed manner, and to obtain the optimal communication frequency, optimal data compression strategy, and optimal computing resource allocation for each node agent.

[0054] In edge computing for power distribution networks, reducing communication overhead, minimizing equipment energy consumption, and ensuring prediction accuracy are often conflicting engineering considerations (e.g., excessive data compression). It will reduce energy consumption, but it will severely compromise prediction accuracy. If adjustment is made solely based on human experience or simple heuristic rules, it is easy to fall into the trap of local optima.

[0055] Based on this, this application constructs the objective function. This transforms the complex engineering trade-offs described above into a rigorous mathematical optimization problem. This is achieved by introducing weighting coefficients (…). The system can automatically evaluate the combined benefits of different strategies within a unified mathematical framework, thereby finding the global optimal solution (or Pareto optimal solution) in multidimensional conflicts.

[0056] The communication frequency index was previously defined in this application. Communication data volume indicators Communication delay and calculate energy consumption Multiple independent metrics. Without overall coordination, individual agents may operate independently, leading to an imbalance in the overall resource allocation of the system. Objective function All core metrics (such as total communication overhead) Prediction error Calculate energy consumption ) and constraints (such as delay limits) This allows distributed optimization algorithms such as ADMM (Alternating Direction Multiplier Method) to have a clear convergence direction, ensuring that each edge node is approaching the global optimum when making local decisions.

[0057] The objective function established in this application is specifically as follows: Describes the minimum value function; This represents the overall optimization objective value of the entire multi-agent collaborative prediction system.

[0058] Indicates communication resource penalty items, Indicates the total communication cost. This represents the communication resource weighting coefficient.

[0059] This represents the prediction accuracy penalty term. The prediction error, calculated on the local validation set based on the current local model parameters, serves as an approximate substitute for the global prediction error. This is the accuracy weighting coefficient; this item aims to ensure the quality of core business operations in the coordinated control of the distribution network and prevent the distortion of prediction results due to excessive compression of communication.

[0060] This indicates an energy consumption penalty item. Calculate energy consumption indicators for nodes. This is the energy consumption weighting coefficient. This parameter aims to minimize the computing power consumption of edge devices from a physical domain perspective, thereby extending device battery life.

[0061] All are normalized weight coefficients, satisfying The data is dynamically distributed by the distribution network master station based on the current business scenario's preferences for real-time performance, accuracy, and energy consumption.

[0062] Constraints: .

[0063] For communication latency metrics, This is the maximum allowable delay threshold for distribution network control services; this item serves as an inequality constraint to ensure that adjustments to the communication strategy do not exceed the physical limits of system stability.

[0064] when Approaching When this happens, the corresponding Lagrange multipliers increase sharply, and this penalty term will force the optimizer to reduce the communication frequency in the next iteration. This avoids the surge in latency caused by network congestion.

[0065] In edge computing environments, resources are highly coupled and mutually constrained. For example, to reduce communication overhead ( The system typically employs more complex gradient compression algorithms, which inevitably leads to higher computational energy consumption. The error rate increases; however, excessive compression can lead to information loss, increasing the prediction error. The objective function increases through weighting coefficients ( By unifying the costs of these three types of conflicts within a single mathematical framework, the ADMM algorithm can find the global optimal solution (Pareto optimality) in multidimensional conflicts, thus avoiding system imbalance caused by single-dimensional optimization.

[0066] Communication delay ( It has nonlinear characteristics, when the communication frequency ( When excessive latency causes network congestion, the latency deteriorates exponentially, directly leading to distribution network control instability. Therefore, this model does not include... Instead of simply adding it to the objective function for linear weighting, it is treated as a hard constraint. This logically ensures that no matter how the system optimizes the allocation of communication and computing resources, it will never come at the expense of the real-time performance and security of the power distribution network control.

[0067] Because the objective function contains inequality constraints ( This application transforms the problem into a standard form of equality constraint problem by introducing slack variables or penalty terms. Specifically, this application uses a first-order Taylor expansion to transform the non-convex problem... The term is approximated as a convex function to facilitate ADMM iterative solution. Subsequently, the augmented Lagrangian function is constructed: in, For Lagrange multipliers, Here, x is the agent's local decision variable matrix, containing locally compressed gradients and model parameters; z is the global consensus reference variable matrix issued by the aggregator; A and B are preset constant diagonal matrices used to match the dimensions of local and global variables; and c is the maximum allowed consensus deviation vector, usually taken as zero.

[0068] Each agent solves the local subproblem independently and alternately, based on local data and information from the previous round of interactions with neighboring nodes: (1) Communication and Compression Subproblem (Optimization) and After fixing the energy consumption variables and multipliers, the agent calculates based on the current channel state (such as bandwidth, queuing delay). And local gradient changes, to solve for the total communication overhead Minimum communication frequency and data compression ratio. During the solution process, the communication frequency of the statistical period is considered. The process is transformed into a continuously decreasing fitting function with respect to the event triggering threshold σ. By optimizing the value of the triggering threshold σ, the agent indirectly obtains the desired communication frequency that minimizes the total communication overhead. .

[0069] (2) Calculation and prediction sub-problems (optimization) and After fixing the communication variables, the agent dynamically adjusts the model complexity (such as the number of LSTM hidden layer nodes or quantization bits) based on the current remaining battery power and local load data, minimizing computational energy consumption while meeting prediction accuracy requirements. .

[0070] (3) After each agent completes local optimization, it exchanges updated local variables (such as compressed gradients or model parameters) only with neighboring nodes or edge aggregators through the communication network. Subsequently, the Lagrange multipliers are updated according to the residuals of the global consistency constraint. : This step is equivalent to imposing a "penalty" or "reward" on agents that fail to meet global constraints, guiding each node to adjust its strategy in the next iteration.

[0071] (4) The system continues to perform the above alternating updates and multiplier iterations until both the original residual and the dual residual are less than the preset convergence threshold. At this point, the algorithm converges to the global Pareto optimal solution, and each agent obtains the optimal communication frequency under the current operating conditions. Optimal data volume And the optimal computing configuration, and based on this, execute load forecasting and collaborative interaction for the next cycle.

[0072] The sixth step: Each node agent uses its local deep learning model to predict the load based on the optimal communication frequency and optimal data compression strategy, updates the local model parameters based on the optimal computing resource configuration, and outputs the collaborative prediction results.

[0073] After the ADMM algorithm converges iteratively, each agent obtains the optimal cooperative strategy (i.e., the optimal communication frequency) under the current operating conditions. Optimal data volume (and optimal computing configuration), and based on this, perform the following specific steps to complete high-precision load forecasting: (1) Global Consistency Parameter Aggregation: Each agent receives globally consistent variables (such as the global average gradient or aggregated model weights) from neighboring nodes or edge aggregators, coordinated by the ADMM algorithm. Since this parameter is subject to hard time constraints during the ADMM iteration process ( With the protection of the intelligent agent, the agent can confirm that the parameter meets the requirements of power distribution network control in terms of timeliness and accuracy.

[0074] (2) Local Model Adaptive Update: The agent fuses the aggregated global consistency parameters with the locally computed local gradients to update the network weights of the local deep learning model (such as LSTM). During this process, the agent strictly follows the optimal computational configuration (such as specific quantization bits or model pruning ratios) obtained from ADMM, minimizing local computational energy consumption while ensuring prediction accuracy. (Control it within the optimal range.)

[0075] (3) High-precision load forecast output: Each agent uses the updated local model, inputs the latest local historical load and meteorological characteristic data, and independently generates load forecast results for the current and future periods. Since the model integrates the collaborative information of multiple nodes in the entire network and has undergone joint optimization of communication and computing resources, its output forecast results have both high precision with a global perspective and ensure real-time response capability at the edge of the distribution network.

[0076] (4) Closed-loop feedback and dynamic re-triggering: After the prediction result is output, the agent calculates the local prediction error in real time. If the error or local gradient change exceeds the preset event trigger threshold again, the system will automatically enter the next ADMM co-optimization cycle to re-evaluate and adjust. and This forms a dynamic closed loop of "prediction-evaluation-adaptive optimization".

[0077] Based on the same inventive concept as the above methods, this application also provides a multi-agent collaborative distribution network load forecasting system, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of any of the above methods.

Claims

1. A multi-agent collaborative method for distribution network load forecasting, characterized in that, The method includes the following steps: The communication frequency index is obtained based on the change in the gradient of the local model and the event triggering threshold; the communication data volume index is obtained based on the gradient sparsity ratio and the quantization bit width; the communication latency index is obtained based on the communication data volume index and the pre-acquired network state parameters; and the node computing energy consumption index is obtained based on the algorithm complexity of the data compression step. A joint objective function is established, which includes the communication frequency index, communication data volume index, node computing energy consumption index, and prediction accuracy index. The communication delay index is introduced into the joint objective function as a hard constraint. The joint objective function is solved in a distributed manner using the alternating direction multiplier method to obtain the optimal communication frequency, optimal data compression strategy, and optimal computing resource allocation for each node agent. Each node agent performs load prediction using its local deep learning model based on the optimal communication frequency and optimal data compression strategy, updates its local model parameters based on the optimal computing resource configuration, and outputs collaborative prediction results.

2. The multi-agent collaborative distribution network load forecasting method as described in claim 1, characterized in that, The method of obtaining communication frequency indicators based on the change in local model gradients and event triggering thresholds specifically includes: Obtain the change range of the model parameters of the local deep learning model at the current time and the last time the communication was triggered; compare the change range with the event trigger threshold; if the change range is greater than or equal to the event trigger threshold, then record the trigger state at the current time as active, otherwise record it as silent; The number of times the communication is active within a preset time period is counted to obtain the communication frequency index.

3. The multi-agent collaborative distribution network load forecasting method as described in claim 1, characterized in that, The method for obtaining communication data volume metrics based on gradient sparsity ratio and quantization bit width specifically includes: Obtain the total dimension of the model parameters to be transmitted; calculate the proportion of gradients that are not discarded based on the gradient sparsity ratio; determine the size of the compressed data bits for a single communication data packet based on the quantization bit width; multiply the total dimension, the proportion of gradients that are not discarded, and the quantization bit width to obtain the communication data volume index.

4. The multi-agent collaborative distribution network load forecasting method as described in claim 1, characterized in that, The process of obtaining the communication latency index based on the communication data volume index and pre-acquired network status parameters specifically includes: Obtain transmission delay, propagation delay, processing delay, and queuing delay; The transmission delay is obtained by dividing the communication data volume index by the network bandwidth, and the network bandwidth and queuing delay are prior constants measured and issued by the distribution network aggregation node in the previous communication cycle. The communication latency index is obtained by adding the transmission latency, propagation latency, processing latency, and queuing latency together.

5. The multi-agent collaborative distribution network load forecasting method as described in claim 3, characterized in that, The algorithm complexity based on the data compression step is used to obtain the node's computational energy consumption index, specifically as follows: The computational complexity of the data compression step is determined, and the computational complexity includes a preset basic computing power cost item, a first dynamic computing power cost item, and a second dynamic computing power cost item; Wherein, the first dynamic computing power cost term is inversely proportional to the proportion of gradients not discarded, and the second dynamic computing power cost term is inversely proportional to the quantization bit width; Based on the hardware environment parameters of the intelligent agent and the computational complexity of each computation step, the dynamic energy consumption of the node executing the computation task is calculated; the dynamic energy consumption is added to the basic idle energy consumption of the node to obtain the node's computational energy consumption index.

6. The multi-agent collaborative distribution network load forecasting method as described in claim 1, characterized in that, The specific formula for the joint objective function is as follows: Describes the minimum value function; This represents the overall optimization objective value of the entire multi-agent cooperative prediction system; Indicates communication frequency specifications; Indicates the amount of communication data; Indicates the communication resource weighting coefficient; This serves as an approximate substitute for the global prediction error. For precision weighting coefficients; Calculate energy consumption indicators for nodes. Energy consumption weighting coefficient; where All are normalized weight coefficients, satisfying .

7. The multi-agent collaborative distribution network load forecasting method as described in claim 6, characterized in that, The objective function is constrained by the communication latency index being less than or equal to the maximum allowable latency threshold for distribution network control services.

8. The multi-agent collaborative distribution network load forecasting method as described in claim 1, characterized in that, The distributed solution of the joint objective function using the alternating direction multiplier method specifically includes: Construct an augmented Lagrange function by introducing Lagrange multipliers and penalty parameters; Each node agent alternately updates local variables, and the updating of local variables includes solving the communication and compression subproblems and the computation and prediction subproblems; By exchanging local variables and updating the Lagrange multipliers until convergence, the optimal event triggering threshold and the optimal data compression strategy can be obtained.

9. The multi-agent collaborative distribution network load forecasting method as described in claim 8, characterized in that, The solution to the communication and compression subproblem specifically includes: The communication frequency index within the statistical period is transformed into a continuously decreasing fitting function about the event trigger threshold; Under the conditions of fixed computational energy consumption index and Lagrange multipliers, the desired communication frequency index that minimizes the total communication overhead is obtained by optimizing the event triggering threshold.

10. A multi-agent collaborative distribution network load forecasting system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-9.

Images