A method for monitoring and optimizing intelligent electrical equipment

Abstract

This invention discloses a method for monitoring and optimizing intelligent electrical equipment, relating to the field of power technology. The method includes: collecting high-dimensional voltage-current time-series data and transient event markers, calculating topological invariant features, and inputting these features into a lightweight neural network model to identify the characteristics of each electrical device; based on the identified features, using causal inference and dynamic regularized regression to perform anomaly attribution and counterfactual energy efficiency prediction, and constructing a multi-objective optimization function through anomaly attribution and counterfactual energy efficiency prediction; based on the multi-objective optimization function, generating equipment operation scheduling strategies through a deep reinforcement learning agent; based on the operation scheduling strategies, sending control commands to the electrical devices, and collecting real-time feedback of operation results for optimization updates. This invention improves the accuracy, robustness, and security of intelligent electrical equipment monitoring and optimization by integrating topological features, causal inference, and secure reinforcement learning.

Description

Technical Field

[0001] This invention relates to the field of power technology, and in particular to a method for monitoring and optimizing intelligent electrical equipment. Background Technology

[0002] With the development of smart grid technology, efficient monitoring and optimization of electrical equipment has become an important way to improve energy efficiency and achieve energy conservation and emission reduction goals. In the past, people used simple methods to monitor and optimize electrical equipment, such as checking whether electricity consumption exceeded a certain fixed value or comparing it with past average levels. However, this method is relatively rigid and cannot react to complex and changing realities—such as sudden weather changes or changes in user habits. Later, with the development of artificial intelligence technology, researchers began to use smarter algorithms to learn the operating patterns of electrical equipment and attempt to automatically adjust the equipment's working status to save energy.

[0003] Despite some progress, existing technologies still face numerous challenges in the monitoring and optimization of intelligent electrical equipment. First, traditional energy efficiency modeling methods typically assume independence between state variables, oversimplifying the factors influencing electricity consumption and ignoring their interrelationships, leading to insufficient model accuracy. Second, existing monitoring systems lack the ability to effectively identify abnormal operating conditions, especially atypical operating conditions, easily resulting in false alarms or missed alarms. Furthermore, the use of reinforcement learning for equipment scheduling optimization ignores equipment safety constraints, potentially leading to unreasonable control decisions and increasing the risk of equipment failure. Therefore, existing technologies suffer from insufficient model accuracy, poor anomaly identification, and unsafe scheduling. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for monitoring and optimizing intelligent electrical equipment, which solves problems such as insufficient model accuracy, poor anomaly identification, and unsafe scheduling.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for monitoring and optimizing intelligent electrical equipment, comprising,

[0008] Collect high-dimensional voltage-current time series data and transient event markers, calculate topological invariant features, and input them into a lightweight neural network model to identify the characteristics of each electrical device;

[0009] Based on the identified characteristics of each electrical device, causal reasoning and dynamic regular regression are used to perform anomaly attribution and counterfactual energy efficiency prediction. A multi-objective optimization function is constructed through anomaly attribution and counterfactual energy efficiency prediction.

[0010] Based on a multi-objective optimization function, a device operation scheduling strategy is generated through a deep reinforcement learning agent. Based on the operation scheduling strategy, control commands are sent to the electrical equipment, and the operation results are collected in real time for optimization and updates.

[0011] As a preferred embodiment of the intelligent electrical equipment monitoring and optimization method of the present invention, the step of using causal reasoning and dynamic regularized regression to perform anomaly attribution and counterfactual energy efficiency prediction based on the identified characteristics of each electrical equipment refers to loading a pre-configured structured causal graph for each electrical equipment, collecting historical operating data of the electrical equipment, constructing a dynamic energy efficiency benchmark model, and calculating the expected power benchmark value, including:

[0012] The initial regression coefficient vector is fitted using the least squares method based on historical operating data. The iteration is performed based on the initial regression coefficient vector. When the updated regression coefficient vector meets the convergence condition, the iteration stops and the final robust coefficient vector is output. Otherwise, the next round of iteration continues. The final robust regression coefficient vector is subjected to mixed regularization to generate the final regression coefficient vector, and a dynamic correction factor is constructed to calculate the expected power benchmark value.

[0013] The hybrid regularization includes preset ridge regularization parameters and Liu-type compression parameters;

[0014] The real-time active power is compared with the expected power baseline. If the relative deviation condition is met, the energy efficiency of the electrical equipment is determined to be abnormal, and an abnormality flag is set. During the diagnostic phase, the posterior probability of being in an abnormal state is calculated based on structured cause-effect graphs and Bayesian inference, and a posterior probability threshold is set. When the posterior probability is greater than the posterior probability threshold If the actual energy consumption is the most likely root cause, then counterfactual energy efficiency prediction is performed. A regression model is used to calculate the daily average energy consumption for both the actual and counterfactual scenarios, and the difference is subtracted to obtain the counterfactual energy savings. .

[0015] As a preferred embodiment of the intelligent electrical equipment monitoring and optimization method of the present invention, the step of constructing a multi-objective optimization function through anomaly attribution and counterfactual energy efficiency prediction refers to acquiring external information and combining real-time active power, anomaly indicators and counterfactual energy saving to construct a multi-objective optimization objective function with causal constraints.

[0016] As a preferred embodiment of the intelligent electrical equipment monitoring and optimization method of the present invention, the method is as follows: based on a multi-objective optimization function, a deep reinforcement learning agent generates an equipment operation scheduling strategy and constructs a corresponding reinforcement learning reward function. The negative value of the multi-objective optimization objective function is decomposed into an instant reward by time step. The instant rewards at all times are summarized to obtain the cumulative reward of the scheduling cycle. The deep reinforcement learning agent is trained using a multi-agent near-end strategy optimization algorithm.

[0017] The multi-agent proximal policy optimization algorithm employs a hybrid architecture of a centralized Critic network and a distributed Actor network. The centralized Critic network outputs an estimate of the long-term cumulative reward, which guides the policy updates of all Actor networks. Each device corresponds to a distributed Actor network that generates control commands. For each device, abnormal periods are selected from historical operation records, and the historical optimal action sequence is extracted. Based on the historical optimal action sequence, a kernel density estimation method is used to construct the prior policy probability distribution of the device. When updating the policy network parameters, an inverse KL divergence penalty term is added to the standard PPO pruning objective function to obtain the overall optimization objective, and the penalty coefficient of the inverse KL divergence penalty term is updated. The overall optimization objective is recalculated using the adjusted penalty coefficient, and the optimal policy is output after training. .

[0018] As a preferred embodiment of the intelligent electrical equipment monitoring and optimization method of the present invention, wherein: the step of sending control commands to electrical equipment based on the operation scheduling strategy refers to issuing the optimal strategy to electrical equipment with intelligent control interfaces through the home LAN. The execution content includes start / stop commands, delayed start, power adjustment, and maintenance suggestion push.

[0019] As a preferred embodiment of the intelligent electrical equipment monitoring and optimization method of the present invention, the real-time acquisition and feedback of operating results for optimization and updating refers to executing the optimal strategy. Then, feedback data is collected and analyzed to adjust the weights of the corresponding causal edges or update the prior probabilities. At the same time, a sliding window mechanism is used to update the local historical dataset, and the energy efficiency regression model is refitted using the least squares method and uploaded to the cloud server in accordance with the federated learning protocol.

[0020] As a preferred embodiment of the intelligent electrical equipment monitoring and optimization method of the present invention, the acquisition of high-dimensional voltage-current time-series data and transient event markers refers to acquiring the voltage and current signals of the main circuit to form an original time-series data sequence, and calculating the current change in real time during the acquisition process. And set the event trigger threshold. and steady-state confirmation threshold It determines whether the electrical equipment has entered a steady state or is turned off, records the end time, and finally generates a list of transient event timestamps.

[0021] As a preferred embodiment of the intelligent electrical equipment monitoring and optimization method of the present invention, the following steps are taken: The topological invariant features are calculated and input into a lightweight neural network model to identify the features of each electrical equipment. Based on the transient event timestamp list, the voltage-current subsequence corresponding to the m-th start-stop event is extracted from the original time-series data sequence, and the voltage-current subsequence is plotted as a closed curve on a two-dimensional plane. The number of turns is calculated based on the closed curve. Simultaneously, continuous homology analysis is performed on the closed curve to obtain feature vectors. The number of turns and feature vectors constitute the topological invariant features. The voltage-current subsequence is input into a pre-trained lightweight neural network model, and the extracted time-series embedded features are fused together with the topological invariant features and input into the classification and regression module to output the start-stop status and real-time active power of each electrical equipment.

[0022] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the intelligent electrical equipment monitoring and optimization method as described in the first aspect of the present invention.

[0023] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the intelligent electrical equipment monitoring and optimization method as described in the first aspect of the present invention.

[0024] The beneficial effects of this invention are as follows: This invention, through an intelligent monitoring and optimization method based on structured causal graphs (SCM) and Bayesian inference, not only overcomes the problems of multicollinearity and sensitivity to anomalous samples in traditional energy efficiency modeling, but also enhances the model's adaptability to operating condition fluctuations by combining a dynamic calibration mechanism for the coefficient of variation. At the same time, by combining counterfactual energy-saving assessment technology, it achieves accurate location and quantitative analysis of energy efficiency anomalies caused by potential root causes. The use of the multi-agent proximal policy optimization (MAPPO) algorithm with prior policy constraints ensures that energy efficiency optimization is achieved while also taking into account the safe operating boundaries of the equipment. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart of a smart electrical equipment monitoring and optimization method in Example 1.

[0027] Figure 2 This is a flowchart of the energy efficiency monitoring and anomaly diagnosis method for electrical equipment in Example 1. Detailed Implementation

[0028] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0029] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0030] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0031] Example 1, referring to Figure 1 and Figure 2 This is the first embodiment of the present invention, which provides a method for monitoring and optimizing intelligent electrical equipment, including the following steps:

[0032] S1. Collect high-dimensional voltage-current time series data and transient event markers, calculate topological invariant features, input them into a lightweight neural network model, and identify the features of each electrical device.

[0033] Specifically, acquiring high-dimensional voltage-current time-series data and transient event markers involves installing an intelligent sensing module integrating voltage transformers (PTs) and current transformers (CTs) at the power inlet of the user's main distribution box. This module synchronously acquires the voltage and current signals of the main circuit at a fixed sampling frequency, ensuring time alignment between the two to form the original time-series data sequence.

[0034]

[0035] in, It is the original time-series data sequence. It is the voltage signal at the nth sampling time. It is the current signal at the nth sampling time;

[0036] During the data acquisition process, the change in current is calculated in real time by subtracting the current signals between adjacent sampling points. Set event trigger thresholds through statistical analysis and steady-state confirmation threshold When the change in current First time greater than the event trigger threshold When this occurs, an electrical equipment start / stop event is determined to have occurred, and the start time of the event is recorded. And continuously monitor it; when the current change is less than the steady-state confirmation threshold for m consecutive cycles. When the time is reached, it is determined that the electrical equipment has entered a steady state or is turned off, and the end time is recorded. Finally, a list of transient event timestamps is generated:

[0037]

[0038] in, It is a list of timestamps for transient events. It is the m-th start / stop event, that is, the start / stop event corresponding to the m-th cycle.

[0039] By employing a high-precision synchronous sampling and dual-threshold dynamic event detection mechanism, the system achieves accurate capture and complete time boundary definition of power equipment start-up and shutdown events. This not only effectively suppresses false triggering caused by grid noise, but also provides a high-quality, structured transient event data foundation for subsequent non-intrusive load identification, equipment-level energy efficiency modeling, and anomaly diagnosis, significantly improving the reliability, timing accuracy, and analytical granularity of the intelligent power monitoring system.

[0040] Furthermore, topological invariant features are calculated and input into a lightweight neural network model to identify the characteristics of each electrical device based on a list of transient event timestamps. From the original time series data sequence Extract the voltage-current subsequence corresponding to the m-th start / stop event. The voltage-current subsequence is plotted as a closed curve on a two-dimensional plane. The horizontal axis represents voltage, and the vertical axis represents current.

[0041] Based on closed curves The number of turns, representing the net number of revolutions the curve makes around the origin (a standard invariant in topology), is calculated using the following formula:

[0042]

[0043] in, Here, di(t) is the number of windings, di(t) is the differential of the current, and du(t) is the differential of the voltage. It is pi. dt is the derivative of current and voltage with respect to time, obtained through the difference approximation; di and du are obtained through the difference approximation.

[0044] At the same time, for closed curves Persistent homology analysis was performed to extract persistent barcodes with 0-dimensional homology (i.e., connected components) and 1-dimensional homology (i.e., ring structure). Four statistical features (maximum lifetime value, mean, number of entries, and variance) were extracted for each feature, generating a total of eight scalar values ​​to form an 8-dimensional feature vector. , will wrap around 8-dimensional feature vectors The voltage-current subsequences are input into a pre-trained lightweight neural network model to form topological invariant features. The extracted temporal embedding features are then fused with the topological invariant features and input into the classification and regression module to output the start / stop status of each electrical device. and real-time active power ;

[0045] The pre-trained lightweight neural network model employs a hybrid architecture of a one-dimensional convolutional neural network (1D-CNN) followed by a bidirectional long short-term memory network (Bi-LSTM). The 1D-CNN contains two convolutional layers, each with a kernel size of 5 and output channels of 32 and 64 respectively, using ReLU activation and batch normalization to extract local temporal features. The Bi-LSTM contains a single-layer hidden unit with a hidden state dimension of 128, used to model the temporal dependencies of the start-up and shutdown processes. The model ultimately outputs device type classification probability, start-up and shutdown status, and real-time active power through a fully connected layer. The model is deployed at the user-side edge. Computing nodes (such as smart gateways or embedded AI chips) are trained using an edge-cloud collaborative federated transfer learning framework. Each user terminal independently trains its model based on locally collected voltage-current subsequences and uploads only the temporal embedding feature tensor output by the Bi-LSTM layer (which is then averaged to obtain a 128-dimensional vector) to the cloud server. Before uploading, noise following a Laplace distribution is added to the feature tensor to achieve differential privacy protection. The cloud server performs a weighted average on the received feature tensors (the weights are the proportion of data volume of each terminal), updates the general feature encoder, and distributes the updated model parameters to each terminal for local fine-tuning.

[0046] By transforming voltage-current trajectories into topological closed curves, combining winding count and continuous coherence features to extract the inherent geometric invariants of the device, and fusing them with deep temporal features, the accuracy and robustness of non-intrusive load identification are significantly improved. At the same time, relying on the edge-cloud collaborative federated transfer learning framework, the model can be generalized across domains while ensuring user electricity privacy. Ultimately, this achieves intelligent power equipment monitoring and optimization with high accuracy, low latency, and strong privacy protection.

[0047] S2. Based on the identified characteristics of each electrical device, causal reasoning and dynamic regular regression are used to perform anomaly attribution and counterfactual energy efficiency prediction. A multi-objective optimization function is constructed through anomaly attribution and counterfactual energy efficiency prediction.

[0048] Specifically, based on the identified characteristics of each electrical device, causal reasoning and dynamic regularized regression are used for anomaly attribution and counterfactual energy efficiency prediction. This involves loading a pre-configured structured causal graph for each electrical device (the causal graph is pre-configured by expert knowledge or manufacturer technical specifications before the deployment of the electrical device). The structured causal graph is a directed acyclic graph (DAG) consisting of a set of observable variable nodes. and the causal relationship constitutes the boundary, in which, These are state variables (external or internal, such as environmental parameters, user settings, component health status, etc.) that affect the energy efficiency of electrical equipment. It is the measured active power of electrical equipment o, and the historical operating data of electrical equipment o are collected. ,in, It is the j-th state variable of the y-th sample. Let M be the measured active power of the y-th sample, and M be the total number of samples. A dynamic energy efficiency benchmark model is constructed to calculate the expected power benchmark value under the current operating conditions, including fitting an initial regression coefficient vector on historical operating data using the least squares (OLS) method.

[0049]

[0050] in, It is the initial regression coefficient vector of the j-th state variable. It is a transpose operation;

[0051] Based on the initial regression coefficient vector, the prediction power for each sample is calculated using the regression coefficients obtained in the previous iteration.

[0052]

[0053] in, It is the predicted power of the regression coefficients from the previous round. These are the intercept term from the previous round and the regression coefficients of each state variable, respectively.

[0054] Predicted power based on current regression coefficients Calculate the relative residuals for each sample:

[0055]

[0056] in, It is the relative residual, which represents the degree of difference between the actual power and the predicted power;

[0057] Based on the relative residuals, calculate the membership weight for each sample:

[0058]

[0059] in, It is the membership weight of the y-th sample;

[0060] Construct a weighted matrix (diagonal matrix) based on the membership weights of each sample:

[0061]

[0062] in, It is the updated regression coefficient vector. It is a state variable design matrix. It is the historical power observation vector. It is a transpose operation;

[0063] If the updated regression coefficient vector satisfies the convergence condition, the iteration stops and the final robust coefficient vector is output; otherwise, the next iteration continues. The convergence condition is:

[0064]

[0065] in, It is the convergence threshold, set through machine learning. It is the final robust coefficient vector. It is the final robustness coefficient vector of the j-th state variable;

[0066] The regression model is further regularized using preset regularization parameters (verified in numerous energy efficiency modeling experiments to be applicable to common electrical equipment such as HVAC, refrigerators, and water pumps; automatically loaded by the system during equipment deployment based on equipment type, requiring no online parameter tuning) to avoid instability or distortion of the physical meaning of regression coefficients due to multicollinearity. The final regression coefficient vector is then calculated.

[0067]

[0068] in, It is the final regression coefficient vector. It is the identity matrix. It is a ridge regularization parameter used to control the diagonal loading intensity. It is a Liu-type compression parameter used to control the compression of the initial estimate. The degree of preservation;

[0069] The final regression coefficient vector used for energy efficiency benchmark calculation is obtained as follows:

[0070]

[0071] in, It is the final regression coefficient vector. It is the final regression coefficient vector of the j-th state variable;

[0072] To improve the adaptability of the energy efficiency benchmark model to operating condition fluctuations, dynamic calibration is further performed by incorporating historical variation information of key state variables. From all state variables, the state variable with the largest absolute value of the regression coefficient is selected as the key variable with the greatest impact on power.

[0073]

[0074] in, The key variable is k, which is the overall state variable.

[0075] Calculate key variables based on historical operating data of electrical equipment o The coefficient of variation is used to measure the degree of volatility.

[0076]

[0077] in, Key variable coefficient of variation, It is the arithmetic mean of the key variables in historical data. It is the sample standard deviation;

[0078] The calculation formula is:

[0079]

[0080] Where M is the total number of samples. It is the key variable for the y-th sample;

[0081] The calculation formula is:

[0082]

[0083] Through key variables coefficient of variation Constructing a dynamic correction factor:

[0084]

[0085] in, It is a dynamic correction factor. It is the adjustment coefficient;

[0086] Substitute the current state variable observations into the regression model and multiply by the dynamic correction factor. Calculate the expected power baseline value:

[0087]

[0088] in, This is the expected power reference value. It is the actual observed value of the j-th state variable at the current time;

[0089] Real-time active power Compared with the expected power reference value The comparison is performed. If the relative deviation condition is met, it is determined that the electrical equipment o has an energy efficiency abnormality, triggering the diagnostic process. The relative deviation condition is:

[0090]

[0091] in, This is the threshold for determining anomalies, referring to the general criteria for determining equipment performance anomalies in IEC 60068-2 and ISO 13374 series standards;

[0092] When the abnormality determination conditions are met, the abnormality flag of the electrical equipment o is set. ,otherwise During the diagnostic phase, posterior probabilities are calculated based on structured causal graphs and Bayesian inference:

[0093]

[0094] in, Given the observation of a "high power" event, the posterior probability that the j-th state variable is in an abnormal state, where "high power" represents the real-time active power. Significantly higher than the expected power baseline , This indicates that the j-th state variable is in an abnormal state. It is a priori probability, set based on equipment runtime or maintenance records. It is the likelihood function, calculated from the residual distribution of the energy efficiency regression model (assuming normality);

[0095] Setting the posterior probability threshold through statistical analysis When the posterior probability of any state variable is greater than the posterior probability threshold If so, it is identified as the most likely root cause. And perform counterfactual (assuming that an abnormal variable is "fixed" to a normal state while other conditions remain unchanged, thereby constructing a "what if there is no problem" virtual operating scenario) energy efficiency prediction, and calculate the average daily energy consumption of the actual scenario and the counterfactual scenario respectively through regression model;

[0096] The formula for calculating the average daily energy consumption in a real-world scenario is:

[0097]

[0098] in, This represents the average daily energy consumption under real-world conditions, and Q represents the future step size. It is a time interval;

[0099] The formula for calculating the average daily energy consumption in a counterfactual scenario is:

[0100]

[0101] in, This is the average daily energy consumption under a counterfactual scenario. It is the value of the l-th state variable at time t under the counterfactual scenario;

[0102] The possible values ​​are:

[0103]

[0104] in, It is the reference value of the j-th state variable under normal conditions. These are the actual state variable observations at time t;

[0105] Daily energy consumption in real-world scenarios Daily energy consumption in counterfactual scenarios Subtraction yields counterfactual energy savings .

[0106] By integrating structured causal graphs and robust weighted regression, a dynamic energy efficiency benchmark model with physical interpretability was constructed. By utilizing adaptive weighting of relative residuals and hybrid regularization strategies, the impact of data noise and multicollinearity on model stability was effectively overcome. Furthermore, based on the dynamic correction mechanism of the coefficient of variation of key variables, the adaptability of the benchmark value to operating condition fluctuations was significantly improved, avoiding misjudgment. After anomaly detection, by combining Bayesian inference and counterfactual simulation, not only can the root cause be accurately located, but the energy-saving potential after repair can also be quantified, realizing a closed loop from "perceiving anomalies" to "attribution diagnosis" and then to "benefit assessment".

[0107] Furthermore, by constructing a multi-objective optimization function through anomaly attribution and counterfactual energy efficiency prediction, it obtains external information such as the current time-of-use electricity price (from the real-time electricity price table published by the power grid company), user comfort constraints (the physical boundaries of environmental parameters acceptable to users, obtained through machine learning, such as temperature), and power grid demand response instructions (electricity prices or control signals published by the power grid, obtained through smart meters, communication protocols, or platform interfaces), combined with real-time active power. Abnormal signs and counterfactual energy saving Construct a multi-objective optimization objective function with causal constraints:

[0108]

[0109] in, It is a multi-objective optimization objective function. It is a scheduling strategy, that is, the sequence of control actions the system takes for all electrical devices over a future period of time. It is the total energy consumption in the future. This refers to electricity expenses during peak hours. It is a normalized user comfort score. It is an indicator function. These are non-negative weighting coefficients used to balance the importance of each objective. They are set using a normalized weighting method. 'o' represents the electrical equipment, i.e., the electrical equipment index variable.

[0110] Future total energy consumption The calculation formula is:

[0111]

[0112] in, It is used to convert watt-seconds to kilowatt-hours. It is a time interval;

[0113] Peak electricity expenses The calculation formula is:

[0114]

[0115] in, This is the current time-of-use electricity price. It is the set of times during peak periods.

[0116] By constructing a multi-objective optimization function that integrates electricity price, comfort, demand response, and equipment energy efficiency diagnostic results, and using counterfactual energy saving as the scheduling incentive term for abnormal equipment, a unified optimization of economy, comfort, grid coordination, and equipment safe operation is achieved. By using a normalized weighting mechanism to flexibly balance multi-objective conflicts, the scheduling strategy can not only respond to external signals but also actively avoid high-risk equipment, significantly improving the overall benefits and robustness of the smart power system.

[0117] S3. Based on a multi-objective optimization function, a device operation scheduling strategy is generated through a deep reinforcement learning agent. Based on the operation scheduling strategy, control commands are sent to the electrical equipment, and the operation results are collected in real time for optimization and updates.

[0118] Specifically, based on a multi-objective optimization function, a device operation scheduling strategy is generated using deep reinforcement learning agents to solve for the multi-objective optimization objective function. The optimal policy is minimized, and the corresponding reinforcement learning reward function is constructed. The multi-objective optimization objective function is then applied. Negative values ​​are decomposed into immediate rewards over time steps:

[0119]

[0120] in, It is the instant reward at time t. It is the total energy consumption increment of electrical equipment o at time t. It refers to the electricity cost at time t. It is the normalized comfort score at time t, calculated from the measured indoor temperature and the user-defined comfort range. The deviation is obtained by linear mapping;

[0121] The calculation formula is:

[0122]

[0123] Where O represents the main electrical equipment;

[0124] The calculation formula is:

[0125] in, This is the current time-of-use electricity price;

[0126] The cumulative reward R for the scheduling cycle is obtained by summing the instantaneous rewards at all times. This reward is used to construct the policy gradient signal, which serves as the regression target of the value function of the Critic network. The deep reinforcement learning agent is trained using the Multi-Agent Proximal Policy Optimization (MAPPO) algorithm. The core architecture consists of a centralized Critic network and a distributed Actor network.

[0127] The centralized Critic network is responsible for evaluating the long-term value of the current overall scheduling strategy from a global perspective. The input to the Critic network is the global state vector.

[0128]

[0129] It is the global state vector. It is the overall comfort target value set for the user;

[0130] Critic network output scalar value , representing the estimated cumulative reward that can be obtained in the global state, used to guide the policy updates of all Actor networks;

[0131] The distributed Actor network consists of an independent Actor network for each electrical device, generating control commands based solely on its local information. This ensures that real-time communication between devices is not required during execution. For each type of electrical device, the input to the distributed Actor network is the local observation vector of the electrical device.

[0132]

[0133] in, It is the local observation vector of the electrical device o. These are the operating status characteristics of the electrical equipment (such as the current gear, running time, water tank temperature, etc., collected by the equipment's sensors or smart sockets).

[0134] The Actor network outputs a probability distribution of a discrete action. That is, the electrical equipment o in the local observation vector The probability distribution of actions under the circumstances, the action space includes a combination of start and stop commands (0 (off) or 1 (on)), delayed start (for interruptible devices (such as electric water heaters and washing machines, postpone the start time to the off-peak period of time-of-use electricity price), power adjustment (for continuously running devices (such as air conditioners, adjust the operating parameters within the user comfort constraint range), and maintenance suggestion push (for devices diagnosed with abnormalities, push specific operation and maintenance prompts to the user terminal).

[0135] To ensure that faulty equipment is not assigned high-risk actions during scheduling, periods without faults (i.e., fault markers) for each electrical device are selected from historical operation records. (Time period), and extract the historical best action sequence executed by the agent that ultimately brings high cumulative rewards. Where N is the number of valid samples, This is the xth historically optimal action. Based on the historically optimal action sequence, the kernel density estimation (KDE) method is used to construct the prior policy probability distribution of the electrical equipment, which is used to characterize the safe operation behavior pattern.

[0136]

[0137] in, This is the prior policy probability distribution of electrical equipment o. This is the action currently to be evaluated. is the Gaussian kernel bandwidth (standard KDE parameter, valued according to Silverman rule), and exp(·) is the natural exponential function;

[0138] When updating the policy network parameters, during the MAPPO algorithm training process, to incorporate the prior policy probability distribution of electrical equipment into the policy update, an inverse KL divergence penalty term is added to the standard PPO pruning objective function, forming the overall optimization objective:

[0139]

[0140] in, It is the overall optimization goal. It is the pruning objective function of the standard PPO. This is the output of the policy network that is currently awaiting updates. It is the prior policy distribution. It is the reverse KL divergence. It is the KL penalty coefficient. It is the set of trainable parameters for the policy network;

[0141] The calculation formula is:

[0142]

[0143] To improve the stability of the training process, the daily scheduling cycle is divided into peak periods. Hegu period Two fixed time periods are used as analysis segments, and the instantaneous rewards of all time steps within each segment are collected. And the empirical distribution is fitted using kernel density estimation (KDE):

[0144]

[0145] in, It is a fit to an empirical distribution. It is a fragment Number of time steps within, It is the Gaussian kernel bandwidth;

[0146] Based on the fitted empirical distribution, calculate the information entropy:

[0147]

[0148] in, It is information entropy;

[0149] The entropy difference is calculated by subtracting the information entropy of two adjacent segments. Entropy difference threshold is set through statistical analysis. When entropy difference Greater than the entropy difference threshold This will trigger the KL penalty coefficient. Dynamic adjustment:

[0150]

[0151] in, This is the adjusted KL penalty coefficient. It is the information entropy of adjacent segments;

[0152] The adjusted KL penalty coefficient is used in the overall optimization objective to achieve adaptive control of the policy update step size. After training, the agent's output makes... The optimal strategy that is minimized is:

[0153]

[0154] in, This is the optimal strategy.

[0155] By accurately mapping multi-objective optimization problems to reinforcement learning reward functions, a business-goal-oriented intelligent scheduling mechanism was constructed. Employing a centralized Critic and distributed Actor architecture, it ensures global collaboration while meeting the low-communication, high-real-time deployment requirements of edge devices. Combined with a KDE prior strategy based on historical security behavior and by using reverse KL divergence constraints to constrain the action selection of abnormal devices, high-risk operations are effectively avoided. Furthermore, a dynamic adjustment mechanism for KL coefficients based on reward distribution entropy is designed to achieve adaptive and stable control during the training process. This not only achieves multi-objective collaboration in electricity cost saving, comfort assurance, and energy efficiency improvement, but also significantly enhances the system's security robustness and policy convergence under device anomalies and environmental disturbances.

[0156] Furthermore, based on the operation scheduling strategy, control commands are sent to electrical devices via the home LAN (such as Wi-Fi or Zigbee) to distribute the optimal strategy to electrical devices with smart control interfaces. The execution content includes start / stop commands (0 (off) or 1 (on)), delayed start (for interruptible devices (such as electric water heaters and washing machines), the start time is postponed to the off-peak period of time-of-use electricity price), power adjustment (for continuously running devices (such as air conditioners), the operating parameters are adjusted within the user comfort constraint range), and maintenance suggestion push (for devices diagnosed with abnormalities, specific operation and maintenance prompts are pushed to the user terminal).

[0157] By precisely distributing optimization strategies to various smart electrical devices through the home LAN, the system can automatically execute differentiated commands such as start / stop, delay, and power adjustment, and proactively push maintenance suggestions to abnormal devices. This not only ensures the physical implementation of scheduling strategies, but also integrates automated control with user-coordinated operation and maintenance, significantly improving the intelligence level, operating economy, and equipment reliability of the home energy system.

[0158] Furthermore, real-time collection of operational results feedback allows for optimization and updates, instructing the execution of future control strategies for electrical equipment. Then, in the next scheduling cycle, actual energy consumption, user behavior feedback (recording whether users adopt maintenance suggestions (yes / no) and whether they manually override scheduling instructions) and comfort scores (obtained through user active scoring or passive monitoring (such as whether the indoor temperature is continuously within the set range)) feedback data are collected.

[0159] The actual energy consumption calculation formula is:

[0160]

[0161] in, This is the actual energy consumption. It is the measured power of the electrical equipment at time t. It is used to convert watt-seconds to kilowatt-hours;

[0162] Based on the collected feedback data analysis, when users adopt maintenance suggestions, the actual energy consumption is then considered. The actual energy savings are calculated by comparing the baseline energy consumption before scheduling with the predicted counterfactual energy savings. If the actual energy-saving effect deviates from the prediction, the causal graph parameter correction is triggered, the weight of the corresponding causal edge is adjusted or the prior probability is updated. At the same time, the local historical dataset is updated using a sliding window mechanism, and the energy efficiency regression model is refitted using the least squares method to ensure that it reflects the latest equipment status. Finally, the data is uploaded to the cloud server in accordance with the federated learning protocol for global model aggregation only, while the original data is kept locally.

[0163] By collecting feedback from actual operation to build a closed-loop verification mechanism, and combining it with online fine-tuning of the energy efficiency model via a sliding window, the system effectively addresses equipment aging and environmental changes. By comparing actual and counterfactual energy savings, the system dynamically corrects the causal graph parameters, enabling continuous evolution of diagnostic knowledge. Furthermore, by relying on federated learning for uploading, the system achieves collaborative optimization through collective intelligence while protecting user privacy, significantly improving the accuracy, adaptability, and reliability of the system's long-term operation.

[0164] This embodiment also provides a computer device applicable to a method for monitoring and optimizing intelligent electrical equipment, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the method for monitoring and optimizing intelligent electrical equipment as proposed in the above embodiment.

[0165] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0166] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the intelligent electrical equipment monitoring and optimization method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

Claims

1. A method for intelligent monitoring and optimization of power utilization equipment, characterized in that: The application relates to a method for energy saving of power consumption equipment. The method comprises the following steps: collecting high-dimensional voltage-current time series data and transient event labels, calculating topological invariant features, and inputting the topological invariant features into a lightweight neural network model to identify characteristics of each power consumption equipment; Based on the identified characteristics of each power consumption equipment, causal reasoning and dynamic regular regression are used to perform abnormal attribution and counterfactual energy efficiency prediction, and a multi-objective optimization function is constructed based on the abnormal attribution and counterfactual energy efficiency prediction; Based on the multi-objective optimization function, a device operation scheduling strategy is generated by a deep reinforcement learning agent, control instructions are sent to the power consumption equipment based on the operation scheduling strategy, and the operation results are collected in real time for optimization and updating; The method comprises the following steps: based on each power consumption equipment, loading a preconfigured structured causal graph, collecting historical operation data of the power consumption equipment, constructing a dynamic energy efficiency benchmark model to calculate an expected power benchmark value, and performing the following steps: Based on the historical operation data, an initial regression coefficient vector is fitted by using a least square method, and an iteration is performed based on the initial regression coefficient vector; when the iteration updated regression coefficient vector meets a convergence condition, the iteration is stopped, a final robust coefficient vector is output, otherwise the next round of iteration is continued; a mixed regularization processing is applied to the final robust regression coefficient vector to generate a final regression coefficient vector, and a dynamic correction factor is constructed to calculate the expected power benchmark value; The mixed regularization comprises a preset ridge type regularization parameter and a Liu type compression parameter; Comparing the real-time active power with the expected power reference value, if the relative deviation condition is met, it is determined that the energy efficiency of the electrical equipment is abnormal, an abnormal flag is set, in the diagnosis stage, the posterior probability of being in an abnormal state is calculated based on the structured causal diagram and Bayesian inference, and a posterior probability threshold is set When the posterior probability is greater than the posterior probability threshold , the most possible root cause is identified, and counterfactual energy efficiency prediction is performed, the daily average energy consumption of the actual scenario and the counterfactual scenario is calculated respectively using the regression model, and the counterfactual energy saving amount is obtained by subtraction ; The multi-objective optimization function is based on a deep reinforcement learning agent generation device operation scheduling strategy pointer, a corresponding reinforcement learning reward function is constructed, a negative value of a multi-objective optimization objective function is decomposed into an immediate reward according to a time step, cumulative rewards of a scheduling period are obtained by summarizing immediate rewards at all times, and a multi-agent proximal policy optimization algorithm is used to train a deep reinforcement learning agent, and an optimal strategy is output after training .

2. The method of claim 1, wherein: The multi-objective optimization function with causal constraints is constructed by combining real-time active power, abnormal flags and counterfactual energy saving amounts based on external information.

3. The method of claim 2, wherein: The multi-agent proximal policy optimization algorithm adopts a hybrid architecture of centralized Critic network and distributed Actor network; the centralized Critic network outputs an estimate of long-term cumulative rewards, and the estimate of long-term cumulative rewards is used to guide policy updates of all Actor networks; each power consumption equipment corresponds to a distributed Actor network to generate control instructions; for each power consumption equipment, abnormal time periods are selected from historical operation records, and historical optimal action sequences are extracted; based on the historical optimal action sequences, a kernel density estimation method is used to construct a prior strategy probability distribution of the power consumption equipment; when updating strategy network parameters, a reverse KL divergence penalty term is added to a standard PPO clipping objective function to obtain a total optimization objective, and a penalty coefficient of the reverse KL divergence penalty term is updated; and the total optimization objective is recalculated by using the adjusted penalty coefficient.

4. The method of claim 3, wherein: The control instruction is sent to the power utilization equipment through a home local area network to the power utilization equipment with a smart control interface The execution content includes start-stop instruction, delayed start, power adjustment, and maintenance suggestion push.

5. The method of claim 4, wherein: The real-time collection operation result feedback optimization update refers to executing an optimal strategy After that, feedback data is collected, the collected feedback data is analyzed, the weight of the corresponding causal edge is adjusted or the prior probability is updated, at the same time, a sliding window mechanism is used to update a local historical data set, and a least square method is used to refit an energy efficiency regression model, and uploaded to a cloud server according to a federal learning protocol.

6. The method of claim 5, wherein: The collecting high-dimensional voltage-current time series data and transient event label refers to collecting voltage signals and current signals of the main loop to form a raw time series data sequence. In the collecting process, the current change amount is calculated in real time , and the event trigger threshold and the steady state confirmation threshold are set, the power consuming equipment is determined to enter the steady state or be turned off, the end time is recorded, and finally the transient event timestamp list is generated.

7. The method of claim 6, wherein: The calculation topology invariant feature is input into a lightweight neural network model, and each electrical equipment feature is identified; according to a transient event timestamp list, a voltage-current sub-sequence corresponding to the mth start-stop event is intercepted from the original time sequence data sequence, and the voltage-current sub-sequence is drawn as a closed curve on a two-dimensional plane; the number of turns is calculated based on the closed curve; meanwhile, the closed curve is subjected to persistent homology analysis to obtain a feature vector; the number of turns and the feature vector constitute the topology invariant feature; the voltage-current sub-sequence is input into a pre-trained lightweight neural network model to extract a time sequence embedding feature; the topology invariant feature and the time sequence embedding feature are fused; and the fused feature is input into a classification and regression module to output the start-stop state and real-time active power of each electrical equipment.

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that: The processor executes the computer program to realize the steps of the intelligent electrical equipment monitoring and optimization method of any one of claims 1-7.

9. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to realize the steps of the intelligent electrical equipment monitoring and optimization method of any one of claims 1-7.

Images