Load identification and regulation method and system based on edge AI and online self-learning

Abstract

This invention belongs to the field of intelligent industrial load management technology, specifically relating to a load identification and control method and system based on edge AI and online self-learning. It includes the following steps: collecting multi-source heterogeneous raw load data through an edge intelligent terminal deployed on the user side; inputting the load feature data into a lightweight deep learning model deployed on the edge intelligent terminal; initiating an online self-learning process based on the confidence level; generating and executing an optimal control strategy for specific electrical equipment based on a preset finite state machine and multi-objective optimization model, according to the identification results and received power grid control instructions; and uploading the model parameters and control effect data locally on the edge intelligent terminal to a cloud master station for global model aggregation and updating. This invention achieves millisecond-level real-time accurate load identification by deploying a lightweight CNN-LSTM hybrid model on the edge.

Description

Technical Field

[0001] This invention belongs to the field of intelligent industrial load management technology, specifically relating to a load identification and control method and system based on edge AI and online self-learning. Background Technology

[0002] With the accelerated construction of new power systems, power load management is shifting from a traditional, extensive, and orderly electricity consumption model to a refined and interactive demand response model. The industrial sector, as the main body of electricity load, comprises complex and diverse load systems including production equipment, air conditioning, energy storage devices, and charging piles, representing a significant source of adjustable resources. However, achieving accurate, rapid, and adaptive identification and control of these loads remains a key technological challenge.

[0003] Currently, load identification technologies are mainly divided into two categories: one is based on invasive load monitoring, which requires the installation of sensors on each electrical device, resulting in high costs and difficult deployment and maintenance, hindering large-scale promotion; the other is based on non-invasive load monitoring, which collects total load data at the user's main power line and uses algorithms to decompose the operating status of each device. Existing non-invasive methods largely rely on centralized computing in the cloud, uploading massive amounts of load data to a central platform for model training and identification. This model has significant drawbacks: firstly, the huge data transmission volume leads to high communication latency, making it difficult to meet the millisecond-level response requirements for real-time load control; secondly, the centralized processing mode is highly dependent on the main station's computing power, posing a single point of failure risk; and thirdly, uploading detailed electricity consumption data raises user privacy concerns, reducing user willingness to participate.

[0004] At the control level, existing strategies are mostly based on preset, fixed rules or simple threshold judgments, lacking precise perception of load dynamic characteristics and real-time operating status. For example, they cannot effectively distinguish the characteristic differences caused by different models and different degrees of aging within the same type of equipment, resulting in overly coarse control commands that may affect normal user production or reduce control effectiveness. In addition, when new electrical equipment is added to the user side or the operating characteristics of existing equipment drift, traditional static models cannot automatically update and adapt, leading to decreased identification accuracy, control strategy failure, and limiting the potential release of load resources to participate in grid interaction.

[0005] Although the concept of edge computing has been introduced into the power sector for local data processing to alleviate cloud pressure, most existing edge solutions only perform simple data collection and forwarding, or run fixed, lightweight models, lacking intelligence with continuous learning capabilities. They cannot solve the core problem of dynamic load changes in industrial sites, and struggle to automatically identify unknown new equipment without human intervention, and maintain high-precision identification and control capabilities over long periods. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for load identification and regulation based on edge AI and online self-learning, so as to solve the problems existing in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a load identification and control method based on edge AI and online self-learning, the method comprising the following steps: Step S101: Collect multi-source heterogeneous raw load data through edge intelligent terminals deployed on the user side, and preprocess and extract features from the raw data to obtain load feature data; Step S102: Input the load feature data into a lightweight deep learning model deployed on the edge intelligent terminal to identify the load type and output the identification result and the corresponding confidence level; Step S103: Based on the confidence level, start the online self-learning process, and use samples with confidence levels lower than a preset threshold and generated pseudo-labels to incrementally update the lightweight deep learning model; Step S104: Based on the identification results and the received power grid control instructions, generate and execute the optimal control strategy for the specific power-consuming equipment based on the preset finite state machine and multi-objective optimization model; Step S105: Upload the model parameters and control effect data of the edge intelligent terminal to the cloud master station, whereby the cloud master station performs global model aggregation and update, and then distributes the updated model to the edge intelligent terminal to achieve edge-cloud collaborative iteration.

[0008] Preferably, in step S101, the collected raw data of multi-source heterogeneous loads includes electrical data and non-electrical data; the feature extraction includes: Calculate the instantaneous active power P(t) and reactive power Q(t); Detect load switching events and extract the feature vector of the events. The feature vector This includes changes in active power ΔP, changes in reactive power ΔQ, and peak starting inrush current. Steady-state power factor after switching and the duration of the transient process At least one of them.

[0009] Preferably, the lightweight deep learning model in step S102 is a hybrid model of convolutional neural network (CNN) and long short-term memory network (LSTM); the load type identification includes: The power sequence is standardized. The standardized power sequence and event features are input together into the CNN module to extract local spatial features. The spatial local features are input into the LSTM module to capture temporal dependencies and obtain the final feature representation of the load. The probability distribution of the output load type is used as the identification result through a fully connected layer and a Softmax function.

[0010] Preferably, the online self-learning process in step S103 includes: When the confidence level is lower than the preset confidence level threshold, the current sample is marked as a sample to be learned; For the samples to be learned, a fast clustering method based on k-nearest neighbors k-NN is used to assign pseudo-labels to them; An incremental dataset is constructed using new samples with the pseudo-labels, and a loss function is calculated based on the incremental dataset; The parameters of the lightweight deep learning model are updated using gradient descent.

[0011] Preferably, in step S104, a finite state machine is defined for each controllable electrical device, whose states include multiple states such as shutdown, startup, normal operation, reduced power operation, and standby; the objective function of the multi-objective optimization model is to minimize the overall cost. The comprehensive cost Costs are influenced by user comfort , power grid regulation deviation cost and the benefits of participating in demand response Weighted composition.

[0012] Preferably, the user comfort affects the cost. Load adjustment of each controlled equipment and their importance weights Related; the power grid regulation deviation cost The square of the difference between the actual total adjustment and the target adjustment; the benefit of participating in demand response. It is determined by the product of the total regulation amount and the unit compensation electricity price.

[0013] Preferably, the global model aggregation and update in step S105 adopts a federated learning approach, whereby the cloud master station aggregates local model parameters from multiple edge intelligent terminals and generates an updated global model by averaging or weighted averaging.

[0014] This invention also discloses a load identification and control system based on edge AI and online self-learning, the system comprising: An edge smart terminal deployed on the user side is used to execute steps S101 to S104 of the method; The cloud master station is communicatively connected to multiple edge intelligent terminals and is used to execute step S105 of the method.

[0015] The beneficial effects of this invention are as follows: By deploying a lightweight CNN-LSTM hybrid model on the edge, this invention achieves millisecond-level real-time accurate load identification; by introducing an online self-learning mechanism based on confidence assessment and pseudo-label generation, the system can automatically identify new devices and dynamically update the model, significantly improving adaptive capabilities and long-term operation and maintenance efficiency; based on a finite state machine and multi-objective optimization intelligent control strategy, it achieves device-level fine-grained control, maximizing the value of load regulation while ensuring user production; the constructed edge-cloud collaborative architecture effectively reduces cloud dependence and network load, fundamentally protecting user data privacy and system robustness; ultimately, it creates considerable economic benefits for power users, provides flexible regulation capabilities for the power grid, significantly improves the level of new energy consumption and system operating efficiency, and has significant economic and social benefits. Attached Figure Description

[0016] Figure 1 This is a flowchart of the process of the present invention. Detailed Implementation

[0017] In the description of this disclosure, it should be understood that the terms “center,” “upper,” “lower,” “front,” “rear,” “left,” “right,” “vertical,” “horizontal,” “top,” “bottom,” “inner,” and “outer,” etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this disclosure and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this disclosure.

[0018] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this disclosure, unless otherwise stated, "a plurality of" means two or more.

[0019] In the description of this disclosure, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," "linking," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection, an electrical connection, or a communication connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this disclosure according to the specific circumstances.

[0020] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings and preferred embodiments.

[0021] This invention belongs to the field of power system load management and artificial intelligence application technology, specifically relating to a load identification and control method based on edge AI and online self-learning. This method addresses the needs for refined management and rapid response to diverse loads on the industrial user side, ultimately improving the accuracy, reliability, and response speed of user-side load participation in grid interaction services such as demand response and ancillary services.

[0022] like Figure 1 As shown, to achieve the above-mentioned objectives, this invention provides a method and system for load identification and control based on edge AI and online self-learning. This method deploys an edge intelligent terminal integrated with a lightweight AI model on the user side to achieve real-time accurate identification, dynamic model updating, and refined collaborative control of load devices. Its core lies in constructing a closed-loop system of "perception-identification-learning-decision-execution." Specifically, it includes the following steps: Step 101: Data Acquisition and Feature Extraction of Multi-Source Heterogeneous Loads on the Edge Side Edge intelligent terminals collect raw data through smart meters, current / voltage sensors, and non-electrical quantity sensors (such as temperature and vibration sensors), and perform preprocessing and feature extraction.

[0023] (1) Acquire one-dimensional time series electrical data, including voltage and current: Voltage value at time t (unit: volts V) Current value at time t (unit: amperes A) (2) Calculate instantaneous active power and reactive power : (1) (2) in, Let t be the phase difference between voltage and current at time t (in radians).

[0024] (3) Extract transient characteristics of load switching events. When a power change exceeding a threshold is detected... And record the event feature vector. : (3) in, This represents the change in active power (unit: watts W). This represents the change in reactive power (unit: Var). Peak value of the inrush current (unit: Ampere A). The steady-state power factor after switching is given. The duration of the transient process (in seconds).

[0025] Step 102: Load Identification Based on Lightweight Convolutional Neural Network (CNN) and Long Short-Term Memory Network (LSTM) Build and run a lightweight hybrid deep learning model to extract load features and perform device type identification.

[0026] (1) Standardize the power sequence: (4) in, This represents the average active power within the time window. This represents the standard deviation of active power within the time window.

[0027] (2) The standardized power sequence Event characteristics A lightweight CNN module is used as a common input to extract local spatial feature maps. (5) in, The convolution kernel weight matrix is... For convolution operations, For the input feature matrix, For bias vectors, It is a linear rectification activation function.

[0028] (3) Feature map Input a lightweight LSTM module to capture time dependencies and obtain the final load feature representation. : (6) (7) (8) (9) (10) (11) in, , , Let represent the activation vectors of the forget gate, input gate, and output gate, respectively. It is the Sigmoid activation function. , , , This is the corresponding weight matrix. , , , For the corresponding bias vector, In cellular state, Candidate cell state, This is element-wise multiplication (Hadamard product).

[0029] (4) Output the probability distribution of load identification through a fully connected layer and a Softmax function. : (12) (13) in, This represents the output vector (logits) of the fully connected layer. For the weights and biases of the fully connected layer, , K Given the total number of known load categories, This is the probability of predicting a load of type i.

[0030] 103: Dynamic Model Update Based on Online Self-Learning Mechanism When the identification confidence is low or an unknown device is detected, an online learning process is initiated to dynamically update the edge-side model.

[0031] (1) Calculate the confidence level of the prediction results If it is below the threshold If it is a sample, then it is marked as a sample to be learned. (14) in, The confidence threshold is typically set to 0.8 to 0.9.

[0032] (2) For samples with low confidence, a fast clustering method based on k-nearest neighbors (k-NN) is used for initial labeling to form pseudo-labels. : (15) in, The labels of the k nearest known samples in the feature space. This is the mode function.

[0033] (3) Construct an incremental dataset using the newly collected samples and their pseudo-labels, and calculate its loss function. : (16) in, This represents the number of samples in the incremental dataset. The pseudo-label (one-hot encoded) for the i-th sample. For all weight parameters of the model, is the L2 regularization coefficient.

[0034] (4) Update model parameters using mini-batch gradient descent. The learning rate is : (17) in, For the loss function with respect to the parameters The gradient of.

[0035] Step 104: Generation of intelligent control strategy based on finite state machine (FSM) and multi-objective optimization Based on the identification results and power grid instructions, a refined control strategy is generated.

[0036] (1) Define a finite state machine for each controllable load, with the following state set: For example: {Power off, Starting up, Normal operation, Reduced power operation, Standby}.

[0037] (2) Define a state transition function T, which is triggered by an event E (such as a control command or equipment status): (18) (3) Establish an optimization model with the objectives of minimizing the impact on user comfort, achieving the best adjustment effect, and maximizing user benefits. The objective function is to minimize the overall cost. : (19) in, For comfort / production impact cost, For the cost of power grid regulation deviation, To participate in the benefits of demand response, For the weighting coefficients, satisfying .

[0038] (4) Comfort cost With load control amount And importance weight Related: (20) in, The number of devices being regulated. The importance weight of the k-th device (e.g., core production equipment has a high weight). This represents the load regulation of the kth device (unit: kW).

[0039] (5) Cost of power grid regulation deviation Measuring the actual total adjustment With target adjustment The gap: (twenty one) (twenty two) (6) Demand response benefits By regulation quantity and compensation electricity price Decide: (twenty three) in, Compensation price per unit of regulation (unit: yuan / kWh).

[0040] (7) Solve the above multi-objective optimization problem to obtain the optimal control instruction set. This triggers the state transition of the corresponding load state machine.

[0041] Step 105: Edge-Cloud Collaboration and Model Iteration Edge terminals perform control and periodically synchronize with the cloud master station.

[0042] (1) The edge terminal will transfer local model parameters Important event logs and aggregated control effect data are uploaded to the cloud main site.

[0043] (2) The cloud main station aggregates data from multiple edge terminals to perform a global model. The training and updating of the model are carried out, and the optimized model is regularly distributed to various edge terminals to achieve knowledge sharing and iteration. (twenty four) in, The number of edge terminals participating in the aggregation. These are the model parameters for the i-th edge terminal.

[0044] By cyclically executing the above five core steps, this invention achieves real-time and accurate identification, adaptive learning, and collaborative optimization control of diverse loads on the industrial user side, effectively improving the flexibility and value of load-side resources.

[0045] It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention.

Claims

1. A load identification and control method based on edge AI and online self-learning, characterized in that, The method includes the following steps: Step S101: Collect multi-source heterogeneous raw load data through edge intelligent terminals deployed on the user side, and preprocess and extract features from the raw data to obtain load feature data; Step S102: Input the load feature data into a lightweight deep learning model deployed on the edge intelligent terminal to identify the load type and output the identification result and the corresponding confidence level; Step S103: Based on the confidence level, start the online self-learning process, and use samples with confidence levels lower than a preset threshold and generated pseudo-labels to incrementally update the lightweight deep learning model; Step S104: Based on the identification results and the received power grid control instructions, generate and execute the optimal control strategy for the specific power-consuming equipment based on the preset finite state machine and multi-objective optimization model; Step S105: Upload the model parameters and control effect data of the edge intelligent terminal to the cloud master station, whereby the cloud master station performs global model aggregation and update, and then distributes the updated model to the edge intelligent terminal to achieve edge-cloud collaborative iteration.

2. The method according to claim 1, characterized in that, In step S101, the collected raw data of multi-source heterogeneous loads includes electrical data and non-electrical data; the feature extraction includes: Calculate the instantaneous active power P(t) and reactive power Q(t); Detect load switching events and extract the feature vector of the events. The feature vector Including changes in active power ΔP, changes in reactive power ΔQ, and peak starting inrush current. Steady-state power factor after switching and the duration of the transient process At least one of them.

3. The method according to claim 1, characterized in that, The lightweight deep learning model in step S102 is a hybrid model of Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM); the load type identification includes: Standardize the power sequence; The standardized power sequence and event features are input together into the CNN module to extract local spatial features. The spatial local features are input into the LSTM module to capture temporal dependencies and obtain the final feature representation of the load. The probability distribution of the output load type is used as the identification result through a fully connected layer and a Softmax function.

4. The method according to claim 1, characterized in that, The online self-learning process in step S103 includes: When the confidence level is lower than the preset confidence level threshold, the current sample is marked as a sample to be learned; For the samples to be learned, a fast clustering method based on k-nearest neighbors k-NN is used to assign pseudo-labels to them; An incremental dataset is constructed using new samples with the pseudo-labels, and a loss function is calculated based on the incremental dataset; The parameters of the lightweight deep learning model are updated using gradient descent.

5. The method according to claim 1, characterized in that, In step S104, a finite state machine is defined for each controllable electrical device, whose states include shutdown, startup, normal operation, reduced power operation, and standby; the objective function of the multi-objective optimization model is to minimize the overall cost. The comprehensive cost Costs are influenced by user comfort , power grid regulation deviation cost and the benefits of participating in demand response Weighted composition.

6. The method according to claim 5, characterized in that, User comfort affects costs Load adjustment of each controlled equipment and their importance weights Related; the power grid regulation deviation cost The square of the difference between the actual total adjustment and the target adjustment; the benefit of participating in demand response. It is determined by the product of the total regulation amount and the unit compensation electricity price.

7. The method according to claim 1, characterized in that, The global model aggregation and update in step S105 adopts a federated learning approach. The cloud master station aggregates local model parameters from multiple edge intelligent terminals and generates an updated global model by calculating the average or weighted average.

8. A load identification and control system based on edge AI and online self-learning, characterized in that, The system includes: An edge smart terminal deployed on the user side is used to perform steps S101 to S104 of the method as described in any one of claims 1 to 7; The cloud master station is communicatively connected to multiple edge intelligent terminals and is used to perform step S105 of the method as described in claim 1 or 7.

Images