A dormitory ai active safety and multi-function complementary cooperation system

Abstract

The application relates to a high school dormitory AI active safety and multi-energy complementary cooperation system, which breaks the data island of electricity "monitoring" and "control", realizes management and control integration, completes the upgrade from single power management to full-factor collaborative control, and builds a multi-dimensional electricity safety line to eliminate hidden dangers such as illegal electricity fire and long-term equipment load from the source. At the same time, the system can reduce the comprehensive energy consumption of the dormitory building by 15%-30% and improve the green electricity consumption rate by more than 50% through multi-energy complementation and AI optimization scheduling. The system also has self-evolution ability and can dynamically adapt to scene changes, minimize the long-term operation cost of the system while ensuring the indoor environment comfort, and has remarkable safety, economy and environmental benefits.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary fields of smart energy management, campus security and artificial intelligence. Specifically, it relates to a comprehensive monitoring and control system and method for university dormitory scenarios that integrates Internet of Things sensing, artificial intelligence decision-making and multi-energy complementary optimization scheduling. Background Technology

[0002] Existing university dormitory energy systems generally face the following structural and technical challenges: Energy supply is isolated and coordination is difficult: photovoltaic power generation, energy storage batteries, air source heat pumps, municipal power grids, and backup electric heating systems typically operate independently, lacking a unified intelligent dispatch center. Each system operates according to its own simple strategies (such as photovoltaic power generation and consumption on demand, and energy storage charging and discharging at fixed times), making it impossible to achieve global dynamic optimal matching of "source-grid-load-storage-loop" at the building level, resulting in low overall energy efficiency.

[0003] Insufficient accuracy in load and supply forecasting: Traditional energy management systems rely heavily on historical electricity data to forecast dormitory loads, severely neglecting key non-electrical physical parameters such as building thermal inertia, indoor occupant dynamics, and the opening and closing status of doors and windows. For example, opening doors and windows can significantly alter a room's heat load, but existing systems cannot detect this and adjust heating strategies accordingly, resulting in energy waste.

[0004] The control strategies are static and rigid, unable to adapt to complex markets: Operating strategies are mostly based on fixed rules or simplified day-ahead optimization, unable to respond in real time to dynamic time-of-use pricing, instantaneous fluctuations in photovoltaic output, and random load changes caused by student behavior. The system lacks online learning and adaptive adjustment capabilities, making it difficult to continuously approach the dynamic goal of "lowest operating cost."

[0005] The data dimensions are limited and the generalization ability of AI models is weak: Existing data analysis is mostly limited to power consumption and electricity, and does not perform spatiotemporal alignment and fusion analysis of multimodal data such as heat pump performance coefficient (COP), battery state of health (SOH), indoor and outdoor temperature difference, and building envelope status (doors and windows). As a result, AI models cannot learn the deep coupling rules of system operation, and the optimization effect is limited. Summary of the Invention

[0006] This invention relates to an AI-powered proactive safety and multi-energy complementary collaborative system for university dormitories. It achieves integrated management and control, upgrading from simple power management to comprehensive collaborative control, thus strengthening electrical safety from multiple dimensions and eliminating potential hazards such as fires caused by illegal electricity use and long-term equipment overload. Simultaneously, relying on multi-energy complementarity and AI-optimized scheduling, the system can reduce the overall energy consumption of the dormitory building. Specifically, the present invention relates to a dormitory AI proactive safety and multi-functional complementary collaborative system, which includes at least a physical device layer, an IoT sensing and data fusion layer, a digital twin and AI learning layer, and an edge control execution layer; wherein the digital twin and AI learning layer is used to generate the optimal action command at based on the data stream transmitted by the IoT sensing and data fusion layer, and to send the optimal action command at to the edge control execution layer.

[0007] Furthermore, the data streams received by the IoT sensing and data fusion layer include at least electrical data streams, thermal and environmental data streams, and market data streams. The IoT sensing and data fusion layer uses a data fusion engine to perform timestamp alignment, unit standardization, and spatial mapping on the above data in a 1-minute cycle to form a unified spatiotemporal matrix of building energy status.

[0008] Furthermore, the digital twin and AI learning layer includes at least an environment simulator and a deep reinforcement learning (DRL) agent. The environment simulator trains a recurrent neural network (RNN) model based on historical data to simulate system state transitions. The deep reinforcement learning (DRL) agent employs a deep deterministic policy gradient (DDPG) algorithm framework to generate optimal action instructions for continuous control. , Among them, P ess,set (t) represents the set power of the energy storage, P hp,set (t) represents the set power of the heat pump, T set , 1…n (t) represents the set temperature of the nth dormitory room, λ 1…n (t) is the flexible load adjustment coefficient allocated to the nth dormitory room.

[0009] Furthermore, before receiving the data stream from the IoT sensing and data fusion layer, there is a training step for the environment simulator and the deep reinforcement learning (DRL) agent using real historical data from the past year; the training step includes an offline pre-training stage, an online rolling optimization and learning stage, and a safety constraint and expert rule embedding stage.

[0010] Furthermore, the deep reinforcement learning (DRL) agent includes at least a deep learning-based malignant load millisecond-level active defense system module, a multi-energy complementary dynamic optimization scheduling module, and a digital twin-based panoramic monitoring and simulation platform module.

[0011] Furthermore, the deep learning-based millisecond-level proactive defense system module for malicious loads includes a data sampling submodule, a feature extraction submodule, and a lightweight AI model. The data acquisition module is used to collect the original waveform data of current and voltage of each power supply circuit in the dormitory in real time through the modified high-frequency sampling intelligent miniature circuit breaker, and to build a waveform feature sample library covering normal electrical appliances and various malicious loads. The feature extraction submodule is used to extract the transient start-up features of the current waveform, the subtle ripple features of the period, and the odd and even harmonic weight ratios unique to nonlinear loads, and to build a high-dimensional feature vector that can accurately distinguish the nature of the load. The lightweight AI model is a hybrid model combining a one-dimensional convolutional neural network (1D-CNN) and a long short-term memory network (LSTM). The one-dimensional convolutional neural network (1D-CNN) is responsible for extracting local deep features from the original waveform, while the long short-term memory network (LSTM) captures the temporal dependencies before and after load switching.

[0012] Furthermore, the multi-energy complementary dynamic optimization scheduling module includes a multi-source data fusion sub-module, an optimization model module, and an intelligent scheduling strategy module. The multi-source data fusion sub-module is used to collect energy and environmental information and campus characteristic information. The optimization model module uses a deep reinforcement learning (DRL) algorithm to continuously optimize the scheduling strategy over the long term. The intelligent scheduling strategy module is used to output a rolling optimized scheduling plan for the next 24 hours.

[0013] Furthermore, the panoramic monitoring and simulation platform module of digital twin is used to synchronize the operating status and monitoring data of physical entities to the virtual model in real time, realizing the mirror mapping and two-way interaction between the physical system and the digital model. It simulates the operation and multi-dimensional effect evaluation of new AI scheduling strategies, extreme weather scenarios, equipment failure plans or emergency plans. After verifying the feasibility, economy and safety, the strategy is then sent to the physical system for execution.

[0014] Furthermore, the constraints of the reward function rt are minimizing electricity costs, reducing equipment wear and tear, and maintaining indoor comfort; the reward function rt is defined as: The physical parameters are defined as follows: Price(t) is the real-time electricity price at time t; Pgrid(t) is the grid power purchase; Δt is the control cycle duration; Pess(t) is the charging and discharging power of the energy storage system at time t; Tin,i(t) is the indoor temperature of the i-th dormitory at time t; Tcomfort is the target comfort temperature or the midpoint of the comfort temperature range, and 1{condition} is the indication function; Wi(t) is the door and window status of the i-th dormitory at time t; Paux,i(t) is the power of the electric auxiliary heating equipment in the i-th dormitory at time t; P threshold α, β, γ are the effective power thresholds for electric auxiliary heating and the weighting coefficients for electric auxiliary heating.

[0015] Furthermore, the active defense system module determines whether a load is malicious based on the confidence threshold of the malicious load. When the threshold is exceeded, the active defense system module sends a trip command to the corresponding smart circuit breaker and locks it within 10 milliseconds. At the same time, it uploads the event details to the cloud platform for alarm and performs intelligent early warning marking to achieve hierarchical management.

[0016] Compared with the prior art, this application has the following advantages: 1) Enhance security from multiple dimensions: Change the identification of malicious loads from "post-event alarm" to "in-event millisecond-level blocking", eliminate fires caused by illegal use of electricity or long-term loads caused by failure to shut down after use from the source, and it is expected to reduce such fires by more than 90%.

[0017] 2) Significant economic and environmental benefits: Through multi-energy complementarity and AI-optimized scheduling, it is expected to reduce the overall energy consumption of dormitory buildings by 15%-30%, increase the green electricity consumption rate by more than 50%, and significantly reduce carbon emissions and electricity costs.

[0018] 3) Achieve true integration of monitoring, management and control: Completely break down the data silos between "monitoring" and "control", and realize the leap from single power management to the coordinated control of all elements of "source-grid-load-storage-safety".

[0019] 4) Possesses self-evolution capability: The system's built-in AI model can continuously learn and optimize online based on new data, adapt to changes in student behavior patterns and devices, and dynamically evolve management strategies to suit different situations. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0021] The structures, proportions, sizes, etc., shown in the accompanying drawings are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed in the specification, and are not intended to limit the implementation conditions of this application. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size should still fall within the scope of the technical content disclosed in this application, provided that they do not affect the effects and purposes that this application can produce.

[0022] Figure 1 This is a schematic diagram of the overall system architecture of the present invention.

[0023] Figure 2 This is a system workflow diagram of the present invention.

[0024] Figure 3 This is a flowchart of the algorithm for the malicious load AI identification model of the present invention.

[0025] Figure 4 This is a schematic diagram illustrating the principle of the multi-energy complementary optimization scheduling method of the present invention.

[0026] Figure 5 This is a schematic diagram of the interface for simulation and deduction of the digital twin platform of the present invention. Detailed Implementation

[0027] The embodiments of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0028] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the following description is provided in conjunction with the appendix. Figure 1-2 The present application will be further described in detail with reference to specific embodiments.

[0029] This invention aims to overcome the aforementioned technical bottlenecks and provide a comprehensive intelligent energy control system for university dormitories that can achieve multi-energy physical coupling modeling, multi-modal data fusion sensing, and dynamic optimization across the entire state space. Its core objective is to automatically discover and continuously iterate the optimal collaborative control strategy through end-to-end deep reinforcement learning, minimizing the long-term operating cost of the system while meeting the hard constraints of indoor environmental comfort.

[0030] like Figure 1 The diagram shown is a schematic of the overall system architecture. The system consists of the following four layers: Physical equipment layer: including rooftop / facade photovoltaic arrays, cascaded energy storage battery packs, air source heat pump systems, independent electric auxiliary heating equipment for each dormitory (such as underground electric heaters or wall-mounted electric heaters), as well as the added door and window magnetic induction sensors and indoor temperature and humidity sensors.

[0031] IoT sensing and data fusion layer: Electricity consumption data stream: Smart meters collect photovoltaic power generation Ppv(t), energy storage charging and discharging power Pess(t), heat pump power consumption Php(t), total electricity consumption of each dormitory Pdorm,i(t), and electric auxiliary heating power Paux,i(t).

[0032] Thermal and environmental data streams: The heat pump system provides heating power Hhp(t) and real-time COP value COP(t); sensors in each dormitory provide indoor temperature Tin,i(t) and door and window open / close status Wi(t)∈{0,1}; the weather station provides outdoor temperature Tout(t) and solar irradiance G(t).

[0033] Market data stream: Access real-time or time-of-use electricity price Price(t).

[0034] Data fusion engine: Using a 1-minute cycle, the above heterogeneous data is timestamped, standardized in units, and spatially mapped (associated with specific dormitories) to form a unified "building energy status spatiotemporal matrix".

[0035] Digital Twin and AI Learning Layer (Core Innovation): Environment Simulator: This simulator trains a recurrent neural network (RNN) model based on historical data to simulate system state transitions. Its inputs are the current state `stst` and the control action `atat`, and its outputs are the predicted next state `st+1′` and the immediate cost `ctct`. This simulator allows agents to perform "sandbox simulations," significantly improving learning efficiency and safety.

[0036] Deep reinforcement learning (DRL) agents: employ the Deep Deterministic Policy Gradient (DDPG) algorithm framework, as it is suitable for continuous action spaces, such as adjusting power setpoints.

[0037] State space SS: defined as an abstraction of the aforementioned "building energy state spatiotemporal matrix", specifically including:

[0038] Action space The agent outputs a continuous control command vector at every 15 minutes:

[0039] Where Pess,set(t) is the energy storage setpoint power (positive discharge, negative charge), Php,set(t) is the heat pump setpoint power, Tset,i(t) is the setpoint temperature of the i-th dormitory (i takes values ​​from 1 to n, affecting the priority of the heat pump and electric auxiliary heating), and λi is the flexible load adjustment coefficient allocated to the i-th dormitory (used for demand response). The reward function rt (the core of guiding the learning direction) is:

[0040] The reward function represents the costs of electricity purchase, storage losses, comfort penalties, and heating penalties when doors and windows are open; α, β, and γ are the weighting coefficients for each term. This function forces the agent to balance minimizing electricity costs, reducing equipment losses, and maintaining indoor comfort, and specifically penalizes the extremely inefficient behavior of "continuous high-intensity electric heating when doors and windows are open." The edge control execution layer converts the optimal action command (at) generated by the DRL agent into specific device control protocols (such as Modbus or BACnet) through the edge gateway and distributes it to the energy storage converter (PCS), heat pump controller, smart air conditioning socket, and environmental control system.

[0041] The definitions and calculation instructions for each parameter are as follows:

[0042] The core AI learning and decision-making process of this application mainly includes the following stages: 1) Offline pre-training stage: Using fused data from the past year, an environmental simulator RNN is trained to accurately predict changes in system state (especially temperature and SOC) under given control actions. Within the trained simulator, a DDPG agent (Actor and Critic networks) is initialized and trained through millions of simulation steps. The agent explores strategies within the simulator, rapidly learning preliminary cooperative rules through reward function feedback. For example, it may use energy storage and surplus photovoltaic power to raise indoor temperatures to the upper limit of the comfort range before peak electricity prices, thus reducing heat pump operation during peak periods.

[0043] 2) Online rolling optimization and learning phase: Every 15 minutes, the system collects the latest real state *st*. The Actor network of the DDPG agent deterministically controls the action *at* based on *st*. Action *at* is then executed. The system waits until the next time step, collects the new real state *st+1*, and calculates the real reward *rt* (where *t+1* represents the next time step). This experience tuple (*st*, *at*, *rt*, *st+1*) is stored in the experience replay buffer. Small batches of data are periodically sampled from the buffer to update the Critic network (evaluating the value of actions) and the Actor network (improving the policy). Online learning enables the agent to continuously adapt to long-term changes such as device performance degradation and shifts in student behavior patterns.

[0044] 3) Security constraints and expert rule embedding: To prevent unreasonable actions during the learning process, a safety filter is superimposed on the final output layer. For example, it may enforce compliance with the SOC (System-Oriented Computation) standard. min ≤SOC(t)≤SOC maxWhen Wi(t) = 1 is detected, the Tset,i of the dormitory is forcibly reduced or its electric auxiliary heating is turned off to ensure that the heat pump operates within its high-efficiency zone. Some proven best practices (such as "prioritizing charging energy storage when there is excess photovoltaic power") are injected into the agent as initial strategies to accelerate learning convergence.

[0045] like Figure 2 , 3 The diagram shown illustrates the core module's operational logic, which includes the following modules: Module A: Deep Learning-Based Malignant Load Millisecond-Level Proactive Defense System This module aims to prevent electrical fires caused by the use of malicious loads at their source. Its innovation lies in combining high-frequency sampling, advanced feature engineering, and a lightweight AI model to achieve a fundamental shift from "post-incident alarm" to "millisecond-level blocking during an incident." The algorithm flowchart for the AI ​​identification model that identifies malicious loads is shown below. Figure 3 As shown.

[0046] Data Acquisition and Preprocessing: Using the modified high-frequency sampling intelligent miniature circuit breaker (sampling rate ≥10kHz), the original waveform data of current and voltage of each power supply circuit in the dormitory are collected in real time, and a waveform feature sample library covering normal electrical appliances (such as computers and desk lamps) and various malicious loads (such as electric kettles, electric stoves, and illegally modified electrical appliances) is constructed.

[0047] Feature engineering: Breaking through the limitations of traditional solutions that rely solely on RMS values ​​and total harmonics, this approach focuses on extracting transient start-up features of the current waveform, subtle ripple features of the period, and the unique odd-even harmonic weighting ratio of nonlinear loads, thereby constructing a high-dimensional feature vector that can accurately distinguish the nature of the load.

[0048] Model Construction and Deployment: A hybrid model combining a one-dimensional convolutional neural network (1D-CNN) and a long short-term memory network (LSTM) is employed. The 1D-CNN is responsible for extracting local deep features from the original waveform, while the LSTM captures the temporal dependencies before and after load switching. After training and optimization in the cloud, the model is lightweighted and deployed to the edge smart gateways in each dormitory building to achieve localized computing.

[0049] Proactive defense process: The edge gateway performs online inference on real-time waveform data. If the confidence level of a detected malicious load exceeds a set threshold (e.g., 95%), the system directly sends a trip command to the corresponding smart circuit breaker and locks it within 10 milliseconds. Simultaneously, it uploads event details (device type, time, location) to the cloud platform for alarm purposes. The platform can intelligently mark dormitories with multiple violations for tiered management.

[0050] Module B: A Multi-Energy Complementary Dynamic Optimization Scheduling Method Integrating Behavioral Characteristics The core innovation of this module is to integrate the unique spatiotemporal behavioral data of universities into the optimization model, solving the problems of inaccurate load prediction and rigid strategies in general scheduling algorithms in dormitory scenarios, and achieving a dynamic balance of multiple objectives such as economy, low carbon emissions, and comfort. Figure 4 As shown.

[0051] Multi-source data fusion: The scheduling model's inputs not only include energy and environmental information such as photovoltaic power generation forecasts, energy storage state of charge (SOC), time-of-use pricing, and meteorological data, but also uniquely incorporates behavioral characteristic data specific to universities, including: class schedules (used to predict dormitory vacancy rates), campus card access control / consumption data (used to infer personnel flow patterns), historical energy consumption patterns for the same period, and holiday arrangements. These data collectively constitute a precise profile of the load characteristics of the dormitory cluster.

[0052] Optimization Model: A multivariate optimization model is established with the objectives of "lowest total operating cost, lowest carbon emissions, and highest energy comfort". Deep reinforcement learning (DRL) algorithms, such as Proximal Policy Optimization (PPO), are employed to enable the system to autonomously learn and approximate the long-term optimal scheduling strategy through continuous interaction with the environment.

[0053] Intelligent scheduling strategy: The model outputs a rolling optimized scheduling plan for the next 24 hours. For example: during the midday peak of solar power generation on sunny days, priority is given to ensuring power supply to dormitory loads, with surplus electricity used to charge energy storage or supply heat pumps in public areas; during peak grid power consumption in the evening, energy storage discharge is coordinated to support part of the load, and the set temperature of air conditioners in public areas is dynamically adjusted to smooth the load curve; when the system recognizes that the entire building has entered a "sleep mode" low-load state, it automatically switches to an energy-saving cruise strategy.

[0054] Module C: Panoramic Monitoring and Simulation Platform Based on Digital Twin This module constructs a virtual space that is synchronized and interactive with the physical world in real time, providing a "sandbox" simulation capability for system decision-making, significantly improving the security and scientific nature of management decisions, such as... Figure 5 As shown.

[0055] 3D modeling: Based on Building Information Modeling (BIM), Geographic Information System (GIS) and equipment parameters, construct a high-fidelity digital twin that includes details of dormitory building structure, power supply network topology, and load distribution.

[0056] Virtual-real mapping: Through IoT data streams, the operating status and monitoring data of physical entities (such as photovoltaic inverters, energy storage systems, and sensors) are synchronized to the virtual model in real time, realizing the mirror mapping and two-way interaction between the physical system and the digital model.

[0057] Simulation and simulation: In a virtual space, new AI scheduling strategies, extreme weather scenarios, equipment failure contingency plans, or emergency plans (such as fire evacuation simulations) can be simulated and their effects evaluated from multiple dimensions. After verifying their feasibility, economy, and safety, the strategies are then deployed to the physical system for execution, thereby greatly reducing trial-and-error costs and operational risks.

[0058] The system workflow will be further explained below in conjunction with specific application scenarios.

[0059] like Figure 2 As shown, the system workflow is as follows.

[0060] Real-time sensing: Terminal sensing devices (smart meters, high-frequency circuit breakers, environmental sensors, access control, etc.) deployed in various locations continuously collect multi-dimensional data and upload it to the edge gateway and cloud data hub through the Internet of Things network.

[0061] Parallel processing: The edge gateway executes localized high-priority tasks in parallel, particularly the millisecond-level identification and proactive power-off of malicious loads in module A. The cloud platform synchronously runs the multi-energy complementary optimization algorithm of module B and the digital twin simulation of module C.

[0062] Decision-making and dispatch: The cloud-based intelligent agent (DRL) generates optimized scheduling instructions based on comprehensive global information. After simulation and verification by the digital twin platform, the safe and reliable instructions (such as energy storage charging and discharging plans and equipment start-up and shutdown commands) are dispatched to the corresponding edge gateways.

[0063] Collaborative execution: The edge gateway coordinates the execution commands of smart circuit breakers, energy storage converters (PCS), and controllable loads on this floor to achieve precise control of energy flow and equipment status.

[0064] Evaluation and Iteration: The digital twin platform compares and evaluates the actual implementation effect of the strategy, and feeds the evaluation results and new operational data back to the AI ​​model for continuous training and optimization, thereby forming a self-evolving closed loop of "perception-decision-execution-optimization".

[0065] The invention will be further illustrated below with reference to an embodiment (a university dormitory area with a population of 5,000).

[0066] Hardware Deployment: Replace the intelligent miniature circuit breaker (with high-frequency sampling function) described in this invention in the distribution box of each dormitory room. Install a total of 500kWp photovoltaic array on the dormitory rooftop, and deploy a 500kWh / 250kW + 150kWh / 90kW phase change energy storage container + lithium iron phosphate battery energy storage system in the underground parking garage. Install one edge intelligent gateway (with GPU computing power) in each building. Deploy air source heat pump units to replace the original gas boilers for hot water supply. Software and Algorithm Implementation: Terminal platform: Deployed in the school's data center, it adopts a microservice architecture and includes a data middleware platform, an AI algorithm platform, a digital twin engine, and visualization applications.

[0067] Model Training: Waveform data of various electrical appliances were collected over a period of 3 months, labeled, and then an initial malicious load identification model was trained. After the accuracy (AUC) reached 0.992, it was lightweightly deployed to various edge gateways. The initial strategy of the multi-energy complementary optimization model was trained by integrating dormitory allocation tables, academic affairs system class schedules, campus card data, and access control data statistics, combined with historical energy data.

[0068] Digital twin construction: 3D model of the dormitory area is built using BIM and GIS data, and tags of all IoT devices are attached.

[0069] Example of running: Scenario 1 (Security Defense): A student in a dormitory plugs in a high-powered hot pot that is not permitted. The edge gateway identifies the characteristics from the current waveform within 8ms, immediately disconnects the circuit breaker of that circuit, and reports an alarm to the administrator's mobile app.

[0070] Scenario 2 (Optimized Scheduling): The platform predicts a sunny day tomorrow with a large-scale event in the afternoon (based on behavioral data). The scheduling strategy is as follows: surplus photovoltaic power is used to charge energy storage at midday; before the event begins in the evening, energy storage is activated to discharge in advance to cope with the peak load, and the air conditioning temperature settings in public areas are slightly increased (within a comfortable range), thus successfully avoiding peak grid electricity prices and ensuring power supply reliability.

[0071] It should be noted that, depending on the implementation needs, the various steps / components described in this application can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.

[0072] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

[0073] The various embodiments in this specification are described in a progressive, parallel, or combined manner. Each embodiment focuses on its differences from other embodiments, and similar or identical parts between embodiments can be referred to interchangeably. For the apparatuses disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant parts can be referred to the method section.

[0074] It should also be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or apparatus comprising a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or apparatus that includes the aforementioned element.

[0075] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A dormitory AI-powered proactive safety and multi-functional complementary collaborative system, characterized in that: It includes at least a physical device layer, an IoT sensing and data fusion layer, a digital twin and AI learning layer, and an edge control execution layer; wherein the digital twin and AI learning layer is used to generate the optimal action command at based on the data stream transmitted by the IoT sensing and data fusion layer, and to send the optimal action command at to the edge control execution layer.

2. The system according to claim 1, characterized in that: The data streams received by the IoT sensing and data fusion layer include at least electrical data streams, thermal and environmental data streams, and market data streams. The IoT sensing and data fusion layer uses a data fusion engine to perform timestamp alignment, unit standardization, and spatial mapping on the above data in a 1-minute cycle to form a unified spatiotemporal matrix of building energy status.

3. The system according to claim 1, characterized in that: The digital twin and AI learning layer includes at least an environment simulator and a deep reinforcement learning (DRL) agent. The environment simulator trains a recurrent neural network (RNN) model based on historical data to simulate system state transitions. The DRL agent employs a deep deterministic policy gradient (DDPG) algorithm framework to generate optimal action commands for continuous control. ， Among them, P ess,set (t) represents the set power of the energy storage, P hp,set (t) represents the set power of the heat pump, T set , 1…n (t) represents the set temperature of the nth dormitory room, λ 1…n (t) is the flexible load adjustment coefficient allocated to the nth dormitory room.

4. The system according to claim 3, characterized in that: Before receiving the data stream from the IoT sensing and data fusion layer, the training steps also include using real historical data from the past year to train the environment simulator and the deep reinforcement learning (DRL) agent. The training steps include, in sequence, an offline pre-training stage, an online rolling optimization and learning stage, and a safety constraint and expert rule embedding stage.

5. The system according to claim 4, characterized in that: A deep reinforcement learning (DRL) agent includes at least a deep learning-based malignant load millisecond-level active defense system module, a multi-energy complementary dynamic optimization scheduling module, and a digital twin-based panoramic monitoring and simulation platform module.

6. The system according to claim 5, characterized in that: The deep learning-based millisecond-level proactive defense system for malicious loads includes a data sampling submodule, a feature extraction submodule, and a lightweight AI model. The data sampling submodule is used to collect the original waveform data of current and voltage of each power supply circuit in the dormitory in real time through a modified high-frequency sampling intelligent miniature circuit breaker, and to build a waveform feature sample library covering normal electrical appliances and various malicious loads. The feature extraction submodule is used to extract the transient start-up features of the current waveform, the subtle ripple features of the period, and the odd and even harmonic weight ratios unique to nonlinear loads, and to build a high-dimensional feature vector that can accurately distinguish the nature of the load. The lightweight AI model is a hybrid model combining a one-dimensional convolutional neural network (1D-CNN) and a long short-term memory network (LSTM). The one-dimensional convolutional neural network (1D-CNN) is responsible for extracting local deep features from the original waveform, while the long short-term memory network (LSTM) captures the temporal dependencies before and after load switching.

7. The system according to claim 6, characterized in that: The multi-energy complementary dynamic optimization scheduling module includes a multi-source data fusion sub-module, an optimization model module, and an intelligent scheduling strategy module. The multi-source data fusion sub-module is used to collect energy and environmental information and campus characteristic information. The optimization model module uses the deep reinforcement learning (DRL) algorithm to continuously optimize the scheduling strategy over the long term. The intelligent scheduling strategy module is used to output a rolling optimized scheduling plan for the next 24 hours.

8. The system according to claim 4, characterized in that: The panoramic monitoring and simulation platform module of digital twin is used to synchronize the operating status and monitoring data of physical entities to the virtual model in real time, realize the mirror mapping and two-way interaction between the physical system and the digital model, simulate the operation and multi-dimensional effect evaluation of new AI scheduling strategies, extreme weather scenarios, equipment failure plans or emergency plans, and after verifying the feasibility, economy and safety, the strategy is then sent to the physical system for execution.

9. The system according to claim 4, characterized in that: The constraints of the reward function rt are minimizing electricity costs, reducing equipment wear and tear, and maintaining indoor comfort; the reward function rt is defined as: The physical parameters are defined as follows: Price(t) is the real-time electricity price at time t; Pgrid(t) is the grid power purchase; Δt is the control cycle duration; Pess(t) is the charging and discharging power of the energy storage system at time t; Tin,i(t) is the indoor temperature of the i-th dormitory at time t; Tcomfort is the target comfort temperature or the midpoint of the comfort temperature range, and 1{condition} is the indication function; Wi(t) is the door and window status of the i-th dormitory at time t; Paux,i(t) is the power of the electric auxiliary heating equipment in the i-th dormitory at time t; P threshold α, β, γ are the effective power thresholds for electric auxiliary heating and the weighting coefficients for electric auxiliary heating.

10. The system according to claim 5, characterized in that: The active defense system module determines whether a load is malicious based on the confidence threshold of the malicious load. When the threshold is exceeded, the active defense system module sends a trip command to the corresponding smart circuit breaker and locks it within 10 milliseconds. At the same time, it uploads the event details to the cloud platform for alarm and performs intelligent early warning marking to achieve hierarchical management.

11. A working method for a dormitory AI-driven proactive safety and multi-energy complementary collaborative system, specifically including the following steps: 1) Data acquisition and preprocessing; 2) Feature engineering; 3) AI model inference; 4) Judgment and Decision Execution; 5) Cloud-based collaboration and model optimization; in, The AI ​​model inference steps include: first, extracting local deep waveform features; then, capturing temporal dependencies; and finally, outputting the recognition results and confidence scores.

Images