Energy consumption management and control system based on smart home

Abstract

The invention belongs to the technical field of smart home, and particularly relates to a smart home-based energy consumption management and control system, which comprises an edge sensing module, an intelligent learning module, an intelligent decision module, a learning reasoning unit and an adaptive module, and is characterized in that environment, equipment and user data are collected in real time, and feature vectors are constructed; training a local user habit model based on the local historical data, and outputting a user intention vector; generating a multi-device cooperative control instruction through a deep reinforcement learning model according to the environment state and the user intention; constructing a household equipment relation graph by means of a graph neural network, and performing global influence deduction and correction on a control instruction; and by monitoring user behavior distribution changes, a learning rapid adaptation mechanism is triggered when habitual structural deviation is detected, and a local model is updated. Aiming at the problems that the control strategy is rigid, the dynamic living habits cannot be adapted and global collaborative optimization is lacked, privacy-safe, personalized and self-adaptive household energy consumption fine management is realized.

Description

Technical Field

[0001] This invention belongs to the field of smart home technology, specifically relating to a smart home energy consumption management system. Background Technology

[0002] Smart home systems have become an important part of modern life. Smart home control systems primarily utilize various sensors, actuators, controllers, and communication technologies to monitor and manage the home environment, providing a more comfortable, convenient, and energy-efficient living experience. Smart home systems can connect various smart devices within the home to meet the family's needs for security, convenience, comfort, and energy conservation. However, traditional smart home systems often rely on preset fixed rules or static thresholds to control and manage smart devices, while users' lifestyles are dynamically changing, such as adjustments to work hours, seasonal differences in sleep patterns, and the addition of new family members. The system cannot perceive these changes, causing it to continuously execute outdated energy consumption strategies, leading to a gradual deterioration in energy efficiency and potentially frequent conflicts with the user's actual needs, thus reducing the user experience. Furthermore, the system fails to adequately consider personalized user needs and environmental changes, resulting in low execution efficiency. It cannot balance the inherent characteristics of household energy consumption—which varies with lifestyle and seasons—with electricity prices dynamically adjusting to grid load conditions. The system lacks the ability to dynamically reconstruct or tiered set energy consumption thresholds based on learned new habits, thus failing to achieve a refined balance between user comfort and energy efficiency that adapts to changing user conditions.

[0003] In summary, existing smart home energy management systems lack a continuous learning mechanism and dynamic adjustment capability for users' lifestyles, making it difficult for their control strategies to evolve in sync with users' ever-changing lifestyles. This results in reduced energy efficiency and a decline in user experience over long-term use. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, this invention provides a smart home energy consumption management system. The system includes an edge sensing module, which is used to collect environmental status data, device status data, user interaction commands, and electricity prices in real time, and construct a multi-dimensional feature vector, wherein the multi-dimensional feature vector includes an environmental status vector. The intelligent learning module is used to periodically train a local user habit model based on historical interaction data stored locally at home, in a privacy-preserving manner. It inputs multi-dimensional feature vectors into the local user habit model and outputs user intent vectors. The intelligent decision-making module, based on a pre-trained reinforcement learning strategy model, takes the environmental state and user intent vector as input, outputs multi-device joint control actions, and controls each home appliance separately with the multi-device joint control instructions; The learning and reasoning module is used to construct and update a heterogeneous family graph with devices and spaces as nodes and physical or functional relationships as edges. It uses a message passing mechanism to perform global impact deduction on the control actions corresponding to candidate multi-device joint control commands, and corrects the multi-device joint control actions based on the deduction results to obtain corrected control commands. The corrected control commands are then used to control home devices. The adaptive module is used to obtain the user intent corresponding to the data related to user behavior and compare it with the user intent corresponding to the user intent vector to obtain the deviation value. If the deviation value is found to exceed the threshold, the parameters of the local user habit model are adjusted.

[0005] Preferably, the multidimensional feature vector further includes a temporal coding vector, a device state vector, and a user's recent behavior sequence vector; The timing coding vector includes absolute time coding, relative time coding, and external environment information coding.

[0006] Preferably, in the intelligent learning module, the privacy protection method is as follows: based on historical data, a stochastic gradient descent algorithm that satisfies differential privacy is used to train and update the parameters of the local user habit model, and the local user habit model is encrypted and stored locally.

[0007] Preferably, the reinforcement learning strategy model uses the environment state vector and the user intention vector as the state space of the reinforcement learning model, the multi-device joint control action as the action space, the weighted summation function of energy saving, comfort, economic cost and user feedback as the reward function, and is trained to maximize long-term reward.

[0008] Preferably, the formula for calculating the reward function is: ; in, Energy-saving rewards are calculated based on the difference between the baseline power and the actual power. For comfort rewards, the deviation of the environmental state vector from the personalized preference baseline extracted from the user habit model is calculated. As an economic cost incentive, calculations are made based on real-time electricity prices and total device power consumption. Rewards for user feedback are determined based on whether the user manually corrects system commands and the deviation value of the correction. These are the weighting coefficients for energy-saving rewards, comfort rewards, economic cost rewards, and user feedback rewards, respectively, with a total of 1.

[0009] Preferably, the edge sensing module is further configured to map the electricity price into an electricity price vector; Furthermore, the intelligent learning module inputs the electricity price vector into the reinforcement learning strategy model, thus adding it to the state space of the reinforcement learning strategy model.

[0010] Preferably, in the learning reasoning module, the edges of the family heterogeneous graph include static edges and dynamic edges, wherein: The static edges are predefined based on the home's physical attribute topology; The dynamic edges are based on the correlation of historical operating data between devices, and their weights are dynamically calculated through an attention mechanism to characterize the coupling strength of functions between devices.

[0011] Preferably, the learning and reasoning module includes a message passing mechanism based on a graph neural network, wherein the update formula for the feature update of the i-th layer node in the graph neural network is: ; in, Let v be the feature vector of node v in the l-th layer. Let be the feature vector of node u in the (l-1)th layer, where node u is a neighbor node of node v. Let v be the initial feature vector of node v. and Let be the learnable weight matrix and bias vector of the l-th layer, respectively. Here, ⊙ represents aggregate functions, and ⊙ represents element-wise multiplication. These are the weight coefficients corresponding to the edge connecting node u and node v. Let v be the set of neighboring nodes. The ReLU function is used as the activation function.

[0012] Preferably, in the adaptive module, the deviation is determined by calculating the KL divergence between the distribution of recent user behavior data and the predicted distribution of the local user habit model.

[0013] The beneficial effects of this invention are as follows: In terms of technical effectiveness, this invention achieves refined and adaptive energy consumption management, significantly improving energy utilization efficiency and user comfort and satisfaction. Regarding user experience, it is ready to use out of the box without complex configuration and continuously learns and evolves. By understanding user intent and providing explainable decisions, it enhances users' sense of control and trust. By balancing real-time performance, privacy, and global optimization capabilities, the graph reasoning mechanism avoids policy conflicts, ensuring the stability and reliability of system operation. Furthermore, privacy protection eliminates concerns about user data leakage, and demand-side response capabilities provide a technical foundation for participation in virtual power plants and grid peak shaving, promoting the overall optimization and sustainable development of the energy system. Attached Figure Description

[0014] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0015] Figure 1 This is a schematic diagram of the system structure of the present invention; Figure 2 This is a schematic diagram illustrating the process by which the intelligent learning module of this invention trains model parameters using a stochastic gradient descent algorithm that satisfies differential privacy. Detailed Implementation

[0016] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0017] Please see Figures 1-2 This embodiment provides a smart home energy consumption management system. The system includes an edge sensing module, an intelligent learning module, an intelligent decision-making module, a learning inference module, and an adaptive module connected by communication. The intelligent learning module, intelligent decision-making module, learning inference module, and adaptive module are integrated and deployed on a local edge computing node. The functions of each module work together as follows: The edge perception module is used to collect environmental status data, device status data, user interaction commands, and electricity prices in real time, and to construct a multi-dimensional feature vector, which includes an environmental status vector. Environmental status data is collected through deployed sensors for temperature, humidity, light intensity, human presence / movement, and CO2. Its function is to provide the system with a baseline of the physical state of the living space, serving as a direct basis for judging comfort and implementing environmental control measures. Device status data is collected in real time, including power consumption, power on / off status, and operating mode, through smart sockets, smart meters, and the device's own communication interfaces (such as Wi-Fi / Zigbee). User interaction commands, by recording all non-automated operations performed by users through physical switches, mobile apps, voice assistants, etc., are used to determine whether the automation strategy matches the user's true intentions and are the core data driving model optimization. Electricity prices are obtained in real time or time-of-use from the power grid or energy service providers via Internet APIs. External economic signals are introduced into the system, directly linking energy-saving behavior with users' economic interests and achieving optimization under cost constraints. The collected multi-source data is normalized. Continuous data is normalized using min-max normalization, while discrete data is normalized using one-hot encoding. A multi-dimensional feature vector is constructed by concatenating 5-dimensional sub-vectors. This multi-dimensional feature vector includes an environmental state vector, a time-series encoded vector, a device state vector, and a recent user behavior sequence vector. The environment state vector and the device state vector are vector values ​​obtained after normalizing the environment state data and device state data, respectively. The time-series coding vector is used to represent absolute time coding and relative time coding. Absolute time coding is the one-hot coding corresponding to the acquisition time of each data point, while relative time is the normalized value of the time difference between the current user interaction command and the previous user interaction command in the user interaction command. The user's recent behavior sequence vector includes a sequence of recent user interaction commands extracted through a sliding window. After encoding the command type, target device, and control parameters in the sequence, the sequence features are extracted through an LSTM network to obtain a fixed-dimensional behavior sequence vector. The edge perception module transmits the constructed multidimensional feature vectors to the intelligent learning module and the intelligent decision-making module in real time, while storing the raw collected data in a local database for subsequent model training.

[0018] The intelligent learning module is used to periodically train the local user habit model in a privacy-preserving manner by constructing a local dataset based on historical interaction data extracted from the local database. The multi-dimensional feature vector is input into the local user habit model, and the output is a user intent vector. Here, the user intent vector is an intent probability distribution vector, which is the probability value that the user may perform a certain action. The local user habit model is constructed using a layered structure combined with LSTM model architecture and trained using a local dataset. The model architecture using this layered structure combined with LSTM model architecture includes: Input layer: Receives multi-dimensional feature vectors transmitted by the edge sensing module; LSTM layer: Extracts temporal dependencies from feature vectors; Fully connected layer: Maps the output of the LSTM layer to the user intent vector as the intent dimension, such as comfort preference, energy saving tendency, usage time preference, etc. The output layer outputs a user intent vector, which is used to quantify the user's personalized needs for the home environment.

[0019] Because the computing power of local edge computing nodes is generally limited, learning models are typically trained in the cloud. However, training local user habit models requires data from home devices and user interactions. Therefore, to prevent the leakage of user privacy data, a stochastic gradient descent algorithm that satisfies differential privacy is used to train the model parameters. The local user habit model is then encrypted and stored locally to ensure that all data sources originate from sensors and devices within the home, absolutely not including any external home data, and that the original data never leaves the local storage area. The specific implementation process is as follows: Step 1: Initialize model parameters: Initialize model parameters, set privacy budget, noise coefficient, gradient clipping threshold, and model learning rate; Step 2: Randomly sample batch data from the local database; Step 3: Use the mean squared error loss function to calculate the gradient of the loss function with respect to the model parameters for the batch data, perform L2 norm clipping on the gradient, and ensure that the clipping value of L2 norm clipping is less than or equal to the set gradient clipping threshold. Step 4: Add Gaussian-distributed noise to the clipped gradient, where the noise variance is determined based on the privacy budget and the degree of execution. Step 5: Update the model parameters of the local user habit model using the gradient with added noise; Step 6: Repeat steps 2-5 until the model converges or the loss function value is lower than the preset threshold; The trained local user habit model is symmetrically encrypted using the AES encryption algorithm, and the encryption key is stored in the security chip of the edge gateway to ensure that the model is not accessed illegally. During the inference phase, the real-time multi-dimensional feature vector transmitted by the edge perception module is input into the encrypted model (decrypted through the security chip before inference), and the user intent vector is output and transmitted to the intelligent decision-making module.

[0020] The intelligent decision-making module, based on a pre-trained reinforcement learning strategy model, takes the environmental state and user intent vector as input, outputs joint control actions of multiple devices, and decomposes the control actions into control commands, which are then used to control each home device separately. The reinforcement learning policy model uses the environment state vector, user intent vector, and electricity price vector as the state space, and multi-device joint control actions as the action space. Each action space corresponds to a set of control parameters for smart home devices. The reward function is a weighted summation function of energy saving, comfort, economic cost, and user feedback, and training is performed to maximize long-term rewards. For the continuous action space, a Deep Deterministic Policy Gradient (DDPG) model is used; for the discrete action space, a Proximal Policy Optimization (PPO) model is used to ensure that the model can efficiently learn the globally optimal control policy. The formula for calculating the reward function is as follows: ; in, The energy-saving reward is calculated based on the difference between the baseline power and the actual power; the larger the difference, the higher the energy-saving reward. The comfort reward is calculated based on the deviation between the environmental state vector and the personalized preference baseline extracted from the user habit model. The formula for calculating the comfort reward is as follows: ; in, This is a baseline of personalized preferences extracted from the local user habit model (such as the user's ideal temperature and humidity). The distance is Euclidean distance (characterizing the deviation between the environmental state and the preference baseline). The comfort reward is set to a value between 0 and 1, with a smaller deviation resulting in a higher comfort reward. As an economic cost incentive, calculations are made based on real-time electricity prices and total device power consumption. To reward user feedback, a segmented function is used to determine whether the user manually corrects system commands and the deviation value of the correction. These are the weighting coefficients for energy-saving rewards, comfort rewards, economic cost rewards, and user feedback rewards, respectively, with a total of 1.

[0021] Home energy consumption control is a typical sequential decision problem. Current actions not only affect immediate energy consumption, but also change environmental conditions (such as temperature), affecting future decision space and comfort. Through a set reinforcement learning strategy model, the system learns strategies that maximize long-term cumulative rewards through continuous trial and error with the environment.

[0022] The learning and reasoning module is used to construct and update a heterogeneous family graph with devices and spaces as nodes and physical or functional relationships as edges. It uses a message passing mechanism to perform global impact deduction on the control actions corresponding to candidate multi-device joint control commands, and corrects the multi-device joint control actions based on the deduction results to obtain corrected control commands. The corrected control commands are then used to control home devices. The edges of a heterogeneous family graph include static edges and dynamic edges, where: Static edges are predefined based on the physical attributes of the home, such as the relationship between devices and their spaces (air conditioner → living room) and the physical connections between devices (router → smart socket). Dynamic edges are based on the correlation of historical operating data between devices, and the edge weights are dynamically calculated through an attention mechanism to characterize the coupling strength of functions between devices; Static edges are constructed once during system initialization, while dynamic edges are dynamically updated based on historical device operation data (such as being updated hourly) through an attention mechanism to ensure that the graph structure can reflect the functional coupling relationship between devices in real time. A heterogeneous graph neural network (HGNN) is used to implement the message passing mechanism, and differentiated message aggregation methods are designed for different types of device nodes and spatial nodes. The feature update formula for the l-th layer node of the graph neural network is: ; in, Let v be the feature vector of node v in the l-th layer. Let be the feature vector of node u in the (l-1)th layer, where node u is a neighbor node of node v. Let v be the initial feature vector of node v. and Let be the learnable weight matrix and bias vector of the l-th layer, respectively. This is an aggregation function that uses a combination of mean aggregation and maximum aggregation to aggregate feature information from neighboring nodes. Let v be the set of neighboring nodes, and ⊙ represent element-wise multiplication, used to fuse neighbor node features with edge weights. These are the weight coefficients corresponding to the edge connecting node u and node v. The ReLU function is used as the activation function to introduce non-linear features. Different weight matrices are used for different node types (devices, spaces) in the heterogeneous graph. This ensures that the model can effectively learn the feature representations of different nodes; The process of performing a global simulation and correcting control actions includes: Candidate action input: Input the multi-device joint control action output by the intelligent decision module into the home heterogeneous graph and update the feature vector of the corresponding device node, such as updating the set temperature feature of the air conditioner node to the corresponding value in the control action; Global impact projection: Through the message passing mechanism of graph neural network, the feature changes of device nodes are transmitted to their neighboring nodes (including other device nodes and space nodes). The feature change of each node is calculated. For example, the temperature of the air conditioner decreases, which leads to the temperature feature of the living room space node, and then the operating power feature of the humidifier node increases. Based on the node feature change, the impact of candidate actions on global energy consumption, comfort and economic cost is projected. Based on the global impact simulation results, the candidate control actions are revised. The revision rules are as follows: If the simulation results show that the global energy consumption exceeds the preset threshold, then reduce the operating power of high-power devices; If the simulation results show that the comfort level of the local space is lower than the user's preferred baseline, then adjust the control parameters of the corresponding equipment. If the simulation results show that the economic cost is too high (during peak electricity prices), then switch the equipment operation mode; Finally, the revised control commands are converted into control commands for each home appliance, enabling precise control of the devices.

[0023] The adaptive module is used to obtain user intent corresponding to user behavior data and compare it with the user intent corresponding to the user intent vector through a model deviation based on KL divergence. If the deviation value exceeds a threshold, it triggers the parameter adjustment of the local user habit model. The specific process includes: A sliding window (window size K, e.g., K=30 days) is used to extract recent user behavior data (such as the time, parameters, and frequency of users manually controlling devices). The probability distribution of the user behavior data is then constructed using kernel density estimation (KDE) or histogram statistics. Where x represents the user's behavioral characteristics; Input the multidimensional feature vector within the sliding window into the local user habit model to obtain the probability distribution of user behavior predicted by the model. ; Quantitative calculation using KL divergence (Kullback-Leibler Divergence) and The deviation between two probability distributions, KL divergence, also known as relative entropy, is used to measure the degree of difference between the two probability distributions. Model parameter adaptive adjustment triggering and execution: Bias Detection: The calculated KL divergence is compared with a preset threshold. If the KL divergence is greater than the preset threshold, the model prediction bias is determined to be too large, triggering parameter adjustment. An incremental training approach is adopted, using recent user behavior data within a sliding window as the training set, and the DP-SGD algorithm is called again to update the parameters of the local user habit model. During the update process, the privacy budget and gradient clipping threshold remain unchanged to ensure the strength of privacy protection. After the parameters are adjusted, the updated model is re-encrypted (AES encryption) and the old model stored in the edge gateway is replaced to achieve adaptive iteration of the model.

[0024] This invention introduces a collaborative intelligent architecture of edge-local-cloud, utilizing models and differential privacy technology to build and continuously update user habit models locally, achieving privacy-secure learning of personalized user needs. Through deep reinforcement learning and composite reward functions, the system can make multi-objective dynamic decisions based on real-time environment, user intent, and electricity price information, achieving a real-time trade-off between energy saving, comfort, and cost. By using graph neural networks to model and reason about complex relationships between devices, it achieves global impact assessment and conflict resolution of control commands, avoiding negative chain reactions caused by local optimization. Finally, through a meta-learning rapid adaptation mechanism, the system can quickly reconstruct the model when it detects structural changes in user habits, ensuring long-term effectiveness and adaptability. Through a collaborative closed-loop system, it solves the problem of traditional unidirectional execution modes based on fixed rules. Through continuous personalized learning under privacy protection, differential privacy is deeply integrated into the model learning client. While ensuring that user data never leaves the local machine, secure and efficient personalized habit modeling and group knowledge sharing are achieved. Multi-objective reinforcement learning decision-making, incorporating user intent, uses the intent vector output by the user habit model as the core part of the deep reinforcement learning state space. A composite reward function integrating energy consumption, personalized comfort, economic cost, and direct user feedback is designed, enabling the agent to make refined, anthropomorphic decisions. Based on global influence reasoning and pre-simulation using graph neural networks, a heterogeneous graph of home devices containing static and dynamic edges is constructed. A message passing mechanism is used to propagate and deduce control actions, achieving a leap from single-point control to system-level collaborative optimization. KL divergence is used to monitor habit drift and trigger rapid model reconfiguration, giving the system the ability to quickly learn in response to drastic changes in lifestyle.

[0025] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A smart home energy consumption management system, characterized in that: The system includes: The edge perception module is used to collect environmental status data, device status data, user interaction commands, and electricity prices in real time, and to construct a multi-dimensional feature vector, which includes an environmental status vector. The intelligent learning module is used to periodically train a local user habit model based on historical interaction data stored locally at home, in a privacy-preserving manner. It inputs multi-dimensional feature vectors into the local user habit model and outputs user intent vectors. The intelligent decision-making module, based on a pre-trained reinforcement learning strategy model, takes the environmental state and user intent vector as input, outputs a multi-device joint control action, and controls each home appliance according to the multi-device joint control action. The learning and reasoning module is used to construct and update a heterogeneous family graph with devices and spaces as nodes and physical or functional relationships as edges. It uses a message passing mechanism to perform global impact deduction on the control actions corresponding to candidate multi-device joint control commands, and corrects the multi-device joint control actions based on the deduction results to obtain corrected control commands. The corrected control commands are then used to control home devices. The adaptive module is used to obtain the user intent corresponding to the data related to user behavior and compare it with the user intent corresponding to the user intent vector to obtain the deviation value. If the deviation value is found to exceed the threshold, the parameters of the local user habit model are adjusted.

2. The smart home energy consumption management system according to claim 1, characterized in that: The multidimensional feature vector also includes a temporal coding vector, a device state vector, and a user's recent behavior sequence vector; The timing coding vector includes absolute time coding, relative time coding, and external environment information coding.

3. The smart home energy consumption management system according to claim 1, characterized in that: In the intelligent learning module, the privacy protection method is as follows: based on historical data, the parameters of the local user habit model are trained and updated using a stochastic gradient descent algorithm that satisfies differential privacy, and the local user habit model is encrypted and stored locally.

4. The smart home energy consumption management system according to claim 1, characterized in that: The reinforcement learning strategy model uses the environment state vector and the user intention vector as the state space of the reinforcement learning model, the multi-device joint control action as the action space, the weighted summation function of energy saving, comfort, economic cost and user feedback as the reward function, and is trained to maximize long-term reward.

5. The smart home energy consumption management system according to claim 4, characterized in that: The formula for calculating the reward function is as follows: ； in, Energy-saving rewards are calculated based on the difference between the baseline power and the actual power. For comfort rewards, the deviation of the environmental state vector from the personalized preference baseline extracted from the user habit model is calculated. As an economic cost incentive, calculations are made based on real-time electricity prices and total device power consumption. Rewards for user feedback are determined based on whether the user manually corrects system commands and the deviation value of the correction. These are the weighting coefficients for energy-saving rewards, comfort rewards, economic cost rewards, and user feedback rewards, respectively, with a total of 1.

6. The smart home energy consumption management system according to claim 1, characterized in that: The edge sensing module is also used to map electricity prices into an electricity price vector. Furthermore, the intelligent learning module inputs the electricity price vector into the reinforcement learning strategy model, thus adding it to the state space of the reinforcement learning strategy model.

7. The smart home energy consumption management system according to claim 1, characterized in that: In the learning and reasoning module, the edges of the family heterogeneity graph include static edges and dynamic edges, wherein: The static edges are predefined based on the home's physical attribute topology; The dynamic edges are based on the correlation of historical operating data between devices, and their weights are dynamically calculated through an attention mechanism to characterize the coupling strength of functions between devices.

8. The smart home energy consumption management system according to claim 1, characterized in that: The learning and reasoning module includes a message passing mechanism based on a graph neural network, wherein the update formula for the feature update of the l-th layer node in the graph neural network is: ； in, Let v be the feature vector of node v in the l-th layer. Let be the feature vector of node u in the (l-1)th layer, where node u is a neighbor node of node v. Let v be the initial feature vector of node v. and Let be the learnable weight matrix and bias vector of the l-th layer, respectively. Here, ⊙ represents aggregate functions, and ⊙ represents element-wise multiplication. These are the weight coefficients corresponding to the edge connecting node u and node v. Let v be the set of neighboring nodes. The ReLU function is used as the activation function.

9. The smart home energy consumption management system according to claim 1, characterized in that: In the adaptive module, the deviation is determined by calculating the KL divergence between the distribution of recent user behavior data and the predicted distribution of the local user habit model.

Images