A multi-modal fusion intelligent park energy management and control system

Abstract

A multimodal fusion smart park energy management system includes: a data acquisition unit that collects multi-source energy consumption data; a decision optimization engine with a built-in dynamic energy path optimizer based on multi-agent reinforcement learning, which performs fusion analysis on multi-source data and generates optimization decision instructions; a communication guarantee component that dynamically allocates network resources according to energy transmission priority; a digital twin platform that constructs a virtual image of the smart park energy system and realizes online simulation and closed-loop control; in the simulation phase, the platform runs multi-step look-ahead simulation based on historical data and real-time multi-source energy consumption data to generate virtual energy flow states; in the decision-making phase, the decision optimization engine calculates the priority scores of each energy transmission path based on the virtual state, generates dynamic optimization instructions, and the communication guarantee component performs predictive resource scheduling. This invention significantly improves the efficiency, robustness, and adaptability of smart park energy management through the synergy of digital twin and intelligent decision-making.

Description

Technical Field

[0001] This invention relates to the field of intelligent park management technology, specifically to a multimodal integrated smart park energy management system. Background Technology

[0002] Traditional smart park energy transmission systems typically employ fixed parameter configurations, making it difficult to adapt to dynamic changes in energy demand and resulting in low energy utilization efficiency. At the data acquisition level, traditional sensors have limited sampling frequencies, hindering accurate energy consumption monitoring; in terms of communication, uneven bandwidth resource allocation severely impacts the real-time performance and reliability of data transmission.

[0003] Therefore, traditional smart park energy transmission systems have obvious technical shortcomings, mainly manifested in rigid system architecture, insufficient data acquisition accuracy, and high communication latency. These defects not only restrict the level of intelligence in energy management, but also increase the system's operating costs and security risks. Summary of the Invention

[0004] To address the shortcomings of existing smart park energy systems in multi-source data fusion, transmission optimization, and communication assurance, this technical solution provides a multimodal fusion smart park energy management and control system. Through deep integration of digital twins and multi-agent reinforcement learning, it integrates high-precision data acquisition, intelligent path optimization, and predictive communication scheduling to achieve dynamic optimization and self-healing control of energy transmission paths, ultimately improving energy utilization efficiency, system robustness, and adaptability; effectively solving the technical problems.

[0005] This invention is achieved through the following technical solution:

[0006] A multimodal fusion smart park energy management system, comprising:

[0007] The data acquisition unit is used to collect multi-source energy consumption data;

[0008] The decision optimization engine has a built-in dynamic energy path optimizer based on multi-agent reinforcement learning, which is used to fuse and analyze multi-source energy consumption data and generate optimization decision instructions for energy transmission paths.

[0009] The communication assurance component integrates a predictive communication resource scheduling algorithm to dynamically allocate network resources based on energy transmission priorities, ensuring reliable and low-latency data transmission.

[0010] The digital twin platform, as the system's central hub, is coupled to the data acquisition unit, decision optimization engine, and communication support components. It is used to construct a virtual image of the smart park's energy system and to achieve online simulation and closed-loop control.

[0011] In the simulation phase, the digital twin platform runs multi-step look-ahead simulation based on historical data and real-time multi-source energy consumption data to generate a virtual energy flow state. In the decision-making phase, the decision optimization engine calculates the priority score of each energy transmission path based on the virtual energy flow state using a multi-agent reinforcement learning algorithm and generates dynamic optimization instructions. The communication assurance element performs predictive resource scheduling according to the priority score, ultimately realizing the overall dynamic optimization and self-healing control of the smart park energy transmission system.

[0012] Furthermore, the data acquisition unit includes a multi-energy access module, an intelligent routing device, and an electronic counting sensor module;

[0013] The data acquisition unit collects multi-source energy consumption data through a multi-energy access module and an intelligent routing device;

[0014] The electronic counting sensor module achieves microsecond-level data acquisition based on electronic energy parameters; the acquired energy value... The calculation method is as follows: ;

[0015] in is Planck's constant. At the speed of light, For electron wavelength, For a moment The number of electrons detected inside, Energy conversion efficiency.

[0016] Furthermore, the multi-source energy consumption data includes photovoltaic power, energy storage power, and heat pump power; the operation method of the data acquisition unit to acquire multi-source energy consumption data includes the following steps:

[0017] Step 1.1: Connect the multi-energy access module to photovoltaic energy, energy storage system, heat pump system and other distributed energy, and dynamically calculate the real-time power based on the physical parameters of each energy source;

[0018] The power calculation of photovoltaic energy depends on time. Light intensity Photovoltaic panel area and photoelectric conversion efficiency Its output power is: ;

[0019] Power calculations for energy storage systems are based on the rate of energy change. and charge / discharge efficiency Its output power is: ;

[0020] Power calculations for heat pump systems involve either cooling or heating capacity. and performance coefficient Its output power is: ;

[0021] Step 1.2: Couple the intelligent routing device to the multi-energy access module to dynamically integrate multi-source power data and calculate the total power by superposition. ;

[0022] in, Indicates the first Real-time power of each energy access point, including photovoltaic energy power. Energy storage system power Heat pump system power Other distributed energy power Its set is ;

[0023] The intelligent routing device receives real-time load forecast values ​​output from the load forecasting module. And based on the calculated real-time load demand changes ;in The load prediction value at the previous sampling time is used to dynamically select the data transmission path by comprehensively comparing the service priority of each data stream with the current network status, so as to minimize transmission delay and packet loss rate, thereby optimizing the data stream path.

[0024] Furthermore, the data acquisition unit collaborates with the digital twin platform through a timing synchronization mechanism, and during the simulation phase, the sampling interval... Upload multi-source energy consumption data according to the simulation requirements of the digital twin platform, and dynamically adjust the acquisition frequency according to optimization instructions during the decision-making stage. With routing strategies.

[0025] Furthermore, the decision optimization engine incorporates a multi-agent reinforcement learning framework. Through the coordinated operation of five stages—state perception, action generation, reward optimization, policy update, and closed-loop verification—it achieves the precise generation and issuance of dynamic optimization instructions for energy transmission paths. The specific execution method includes the following steps:

[0026] Step 5.1: In the state awareness phase, each agent... Each corresponds to an energy access point. Energy access point Each intelligent agent includes photovoltaic units, and / or energy storage systems, and / or heat pump units. state space Defined as:

[0027] ;

[0028] in, For real-time power data, For load demand forecasting, For network topology information, the state space ensures that each agent... Both can make locally optimal decisions based on comprehensive environmental information;

[0029] The state space Integrates real-time power data collected by the data acquisition unit Load demand forecasting based on historical data and weather information using GRU neural networks. and network topology information obtained in real time by the SDN controller. This provides a global environment awareness foundation for instruction generation;

[0030] Step 5.2: In the action generation phase, the engine outputs power allocation instructions for each energy transmission path through the action space generation module based on the near-end policy optimization algorithm. Power allocation instructions for motion space The constraints are satisfied: ,in, Total power;

[0031] Step 5.3: After the instruction is generated, the decision optimization engine (5) adopts a multi-objective reward function. Guiding agents to cooperate in optimization is defined as follows:

[0032] ;

[0033] in, For multi-objective weighting coefficients, satisfying ; This is the dynamic energy efficiency target value; Actual operating energy efficiency; For dynamic cost budgeting; Actual costs incurred; The rated frequency; This refers to the actual system frequency. For frequency tolerance;

[0034] Step 5.4: The decision optimization engine achieves online learning through the policy update module, policy function. parameters Updated using gradient ascent: ,in To accumulate rewards, The learning rate;

[0035] Step 5.5: Generate dynamic optimization instructions to achieve adaptive control of the energy transmission path; the dynamic optimization instructions are: Before the dynamic optimization instruction is issued, the simulation results are used to verify the feasibility of the decision instruction through the digital twin platform (7), and the optimization strategy is iterated according to the real-time data stream. If the simulation deviation exceeds the threshold, re-optimization is triggered to form a closed-loop control.

[0036] Furthermore, the communication assurance element performs the following steps:

[0037] Step 6.1: Construct a hybrid neural network model based on historical communication data and real-time multi-source energy consumption data. The hybrid neural network model includes an input layer, a hidden layer, and an output layer.

[0038] The input layer receives real-time parameters from the data acquisition unit. ) and historical communication traffic Real-time parameters ( Includes light intensity Energy storage charging and discharging status Heat pump operation mode Environmental parameters are used to characterize external operating conditions that affect communication requirements;

[0039] The hidden layer uses a bidirectional long short-term memory network to extract temporal features;

[0040] Output layer predicts the future Communication demand vector of each energy node within a given time period:

[0041] ,

[0042] in, , , These represent the predicted communication demand intensity of the photovoltaic, energy storage, and heat pump nodes at time t+Δt, respectively.

[0043] Step 6.2: Construct a weighted multi-objective function:

[0044] ;

[0045] in, This is a transmission delay optimization term, reflecting the degree of matching between the average delay of data packets and the priority of energy transmission; This is a bandwidth utilization optimization item to ensure that high-priority energy data does not occupy core bandwidth resources; This is an energy efficiency optimization item to reduce the energy consumption of the communication module; Let be the weighting coefficient, satisfying And it is dynamically adjusted using the entropy weight method to adapt to real-time operating conditions; Let be the communication priority weight of the i-th energy node; Let be the data transmission delay of the i-th node; The available bandwidth of the i-th node; To allocate bandwidth in practice; To ensure minimum bandwidth for each link; This is the sum of the total bandwidth capacity; The energy consumption of the kth communication node; N represents the total energy consumption; M represents the total number of energy nodes; K represents the total number of communication links; and K represents the total number of communication nodes (switches, gateways, etc.).

[0046] Step 6.3: Execute the predictive communication resource scheduling algorithm, which uses the communication demand prediction vector obtained in Step 6.1. Using real-time network topology state information as the core input, a constrained dynamic resource optimization problem is constructed; and the bandwidth allocation vector is iteratively solved on an embedded CPU using the gradient descent method. Minimize the weighted multi-objective function defined in step 6.2 Mathematical expression is The constraints include total bandwidth capacity limitations. and minimum bandwidth guarantee for each link Its mathematical expression is:

[0047] ;

[0048] in, Assign a vector to the bandwidth; The maximum allowable energy consumption; each element in the vector Representative assigned to the first The bandwidth of each communication link Index of communication links The algorithm uses gradient descent for iterative solution, and the update formula is:

[0049] ;

[0050] in, This represents the number of iterations. For the objective function The gradient; during the optimization process, online simulation is performed using a digital twin platform to verify the feasibility of the allocation strategy, and the optimization parameters are adjusted based on the simulation results;

[0051] The algorithm operates in a rolling time-domain control mode, re-predicting and optimizing at regular intervals; it updates the scheduling strategy based on the latest data and prediction results, forming a closed-loop feedback, and ultimately obtains the optimal objective function. ;

[0052] Then the optimal objective function Substituting these values ​​into the formula for iteratively solving the bandwidth allocation vector, the optimal bandwidth allocation vector is obtained. ;

[0053] Step 6.4: Issuance and execution of the optimized resource allocation scheme: The dynamic resource allocation module receives the optimal bandwidth allocation vector obtained from Step 6.3. Subsequently, the module will transfer the vector The configuration commands are converted into a set of bandwidth configuration instructions that can be recognized by specific network devices. These instructions are then distributed to the corresponding switches and routers in the campus network via the southbound interface of the software-defined network (SDN) controller. The network devices dynamically adjust their port or queue bandwidth allocation strategies according to the instructions, thereby ensuring that the data transmission links of energy nodes such as photovoltaic, energy storage, and heat pumps achieve optimal bandwidth allocation. The defined bandwidth resources ultimately achieve synergistic optimization of transmission delay, bandwidth utilization, and system energy efficiency.

[0054] Furthermore, the execution steps of the digital twin platform include:

[0055] Step 7.1: Couple the digital twin platform to the data acquisition unit, decision optimization engine, and communication support components to build a multi-dimensional virtual image of the smart park energy system;

[0056] Step 7.2: In the simulation phase, the digital twin platform runs multi-step look-ahead simulations based on historical data and real-time multi-source energy consumption data. The simulation model integrates energy transmission dynamics equations and communication delay characteristics to generate future... Virtual energy flow state over time Multi-step simulation is achieved through iterative calculations, as shown in the formula:

[0057] ,

[0058] in, To control the input vector, It is a system dynamic function that encompasses changes in energy flow and network constraints; and The first Step and the first The virtual system state vector of each step contains information such as power, load, and network status of each node;

[0059] Step 7.3: In the decision-making phase, the digital twin platform and the decision optimization engine form a closed-loop interaction: First, the virtual energy flow state generated in Step 7.2 is... This serves as input to the multi-agent reinforcement learning algorithm within the decision optimization engine. Based on this input, the decision optimization engine calculates the priority scores for each energy transmission path. This calculation process is part of the multi-agent reinforcement learning algorithm, which evaluates the energy transmission paths in a virtual state. Different actions (power allocation instructions) will result in future cumulative rewards (determined by the reward function). The expected value of the definition is used to indirectly assign priority scores to each path, with higher-scoring paths corresponding to better power allocation strategies. At the same time, the platform verifies the feasibility of decision instructions through online simulation and dynamically adjusts virtual image parameters based on deviation detection to improve simulation accuracy and robustness.

[0060] Step 7.4: The digital twin platform supports the dynamic adaptive and self-healing control of the energy transmission system through continuous iterative optimization. The optimization objective is: ,in, This represents the actual system state.

[0061] The platform continuously compares the virtual state using a deviation detection algorithm. Compared with the actual state To minimize the difference between the virtual and actual states, if the deviation norm exceeds the threshold, the virtual mirror parameters are dynamically adjusted to improve simulation accuracy.

[0062] Ultimately, the platform uses the timing feedback mechanism in the model predictive control framework to send the optimized command timing to the physical system, forming a closed-loop feedback to support the system's dynamic adaptive and self-healing control. If an overload is predicted in the simulation, the energy routing is adjusted in advance through the decision engine.

[0063] Furthermore, the virtual image integrates physical entity mapping relationships, including energy equipment topology, energy flow dynamics, and communication network status, and maintains consistency with the physical system through a real-time data synchronization mechanism; the virtual image construction is based on multi-source heterogeneous data fusion, mathematically expressed as:

[0064] ,

[0065] in, For a moment The virtual mirror state vector, For communication data streams, This is a mapping function that enables precise projection of the physical system into the virtual space. The total power of the system. This refers to network topology information.

[0066] Beneficial effects

[0067] The multimodal fusion smart park energy management system proposed in this invention has the following advantages compared with existing technologies:

[0068] (1) This invention employs a high-precision data acquisition and intelligent routing device, which significantly improves the real-time performance and accuracy of energy consumption data; based on the collaborative mechanism of digital twin platform and multi-agent reinforcement learning, it realizes dynamic optimization and self-healing control of energy transmission paths; and the predictive communication resource scheduling algorithm further ensures low latency and high reliability of data transmission. While improving energy utilization efficiency and enhancing system robustness, this system effectively reduces operation and maintenance costs, providing key technical support for the sustainable development of smart parks.

[0069] (2) This invention, through the cooperation of a multimodal data acquisition unit, a multi-energy access module, and an intelligent routing device, can accurately collect energy consumption data from multiple sources such as photovoltaics, energy storage, and heat pumps in real time. Combined with a dynamic path optimizer based on multi-agent reinforcement learning, it can achieve real-time optimization and adjustment of energy transmission paths. This intelligent energy management mechanism effectively overcomes the problems of uneven energy distribution and large transmission losses in traditional systems, thereby comprehensively improving energy utilization efficiency and enhancing the system's adaptability to dynamic energy demands.

[0070] (3) The digital twin platform in this invention endows the system with stronger adaptability and robustness. By constructing a virtual mirror of the smart park energy system and realizing online simulation and closed-loop control, the system can perform multi-step look-ahead simulation to predict changes in energy flow state, and then realize dynamic calibration of the system state through a model predictive control framework. This mechanism enables the system to have self-healing capabilities in abnormal situations such as equipment failure or sudden load changes, ensuring the continuous stability of energy transmission and greatly improving the operational reliability of the system.

[0071] (4) The predictive communication resource scheduling algorithm in this invention achieves a significant breakthrough in transmission reliability. By constructing a weighted multi-objective function, the system dynamically allocates network resources according to energy transmission priority and intelligently optimizes transmission paths. This not only effectively reduces data transmission latency and improves bandwidth utilization efficiency, but also ensures the priority transmission of critical energy control commands, providing strong communication support for the smart park energy management system and significantly improving the overall performance of the system. Attached Figure Description

[0072] Figure 1 This is a schematic diagram of the overall system architecture in this invention.

[0073] Figure 2 This is a schematic diagram of the specific structure of the present invention.

[0074] Figure 3 This is a flowchart of the predictive communication resource scheduling algorithm in this invention.

[0075] Figure 4 This is a performance comparison analysis diagram of the multimodal fusion smart park energy system of the present invention.

[0076] Figure 5 This is a diagram illustrating the energy flow and structural optimization analysis of the present invention.

[0077] Figure 6 This diagram illustrates the system robustness and adaptive capability verification of the present invention. Detailed Implementation

[0078] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. The described embodiments are merely some embodiments of the present invention, and not all embodiments. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the design concept of the present invention should fall within the protection scope of the present invention.

[0079] Example 1:

[0080] like Figure 1 and Figure 2 As shown, a multimodal fusion smart park energy management system includes: a data acquisition unit 1, a decision optimization engine 5, a communication support component 6, and a digital twin platform 7.

[0081] The data acquisition unit 1 includes a multi-energy access module 2, an intelligent routing device 3, and an electronic counting sensor module 4. The data acquisition unit 1 collects multi-source energy consumption data such as photovoltaic power, energy storage power, and heat pump power through the multi-energy access module 2 and the intelligent routing device 3. As the physical sensing layer of the system, the key to the implementation of the data acquisition unit lies in the high-precision, real-time acquisition and preliminary fusion of multi-source heterogeneous data.

[0082] The multi-energy access module 2 uses an embedded microprocessor (STM32H7 series) based on ARM architecture as the main control chip, and connects an external high-precision current / voltage sensor (Honeywell CSNL series) to collect electrical parameters of the photovoltaic array, lithium-ion energy storage system and air source heat pump in real time.

[0083] The photovoltaic power calculation module collects data from the light intensity sensor via the ADC module in 1-second cycles, and combines this data with the preset photovoltaic panel area (…). ) and efficiency parameters ( ), calculate its output power in real time.

[0084] The intelligent routing device 3 is essentially an industrial-grade edge computing gateway (such as the Huawei AR502 series). Its built-in time-series database (InfluxDB) is responsible for caching data streams from each access point and running lightweight data fusion algorithms.

[0085] The electron counting sensor module 4 uses an energy spectrum analyzer based on the photoelectric effect (Hamamatsu H11901 photomultiplier tube). Its microsecond-level acquisition capability is achieved through FPGA to realize high-speed electron pulse counting and energy value... The calculations strictly follow the quantum physics formulas in the documentation. The electron counting sensor module achieves microsecond-level data acquisition based on electron energy parameters, and the acquired energy value... The calculation method is as follows: ;

[0086] in is Planck's constant. At the speed of light, For electron wavelength, For a moment The number of electrons detected inside, Energy conversion efficiency.

[0087] The operation method of data acquisition unit 1 for acquiring multi-source energy consumption data includes the following steps:

[0088] Step 1.1: Connect the multi-energy access module to photovoltaic energy, energy storage system, heat pump system and other distributed energy, and dynamically calculate the real-time power based on the physical parameters of each energy source;

[0089] The power calculation of photovoltaic energy depends on time. Light intensity Photovoltaic panel area and photoelectric conversion efficiency Its output power is: .

[0090] Power calculations for energy storage systems are based on the rate of energy change. and charge / discharge efficiency Its output power is: .

[0091] Power calculations for heat pump systems involve either cooling or heating capacity. and performance coefficient Its output power is: .

[0092] Step 1.2: Couple the intelligent routing device to the multi-energy access module to dynamically integrate multi-source power data and calculate the total power by superposition. .

[0093] in, Indicates the first Real-time power of each energy access point, including photovoltaic energy power. Energy storage system power Heat pump system power Other distributed energy power Its set is .

[0094] The intelligent routing device receives real-time load forecast values ​​output from the load forecasting module. And based on the calculated real-time load demand changes ;in The load prediction value at the previous sampling time is used to dynamically select the data transmission path by comprehensively comparing the service priority of each data stream with the current network status, so as to minimize transmission delay and packet loss rate, thereby optimizing the data stream path.

[0095] The data acquisition unit collaborates with the digital twin platform through the OPC UA protocol, a timing synchronization mechanism. During the simulation phase, the sampling interval... Based on the simulation requirements of the digital twin platform, the system dynamically configures and uploads multi-source energy consumption data within a range of 100ms to 10s, and dynamically adjusts the acquisition frequency according to optimization instructions during the decision-making phase. With routing strategies; ensure the real-time nature of data acquisition and the balance of system load.

[0096] The decision optimization engine 5 incorporates a dynamic energy path optimizer based on a multi-agent reinforcement learning framework. This optimizer is used to fuse and analyze multi-source energy consumption data and generate optimization decision instructions for energy transmission paths. Through the coordinated operation of five stages—state perception, action generation, reward optimization, policy update, and closed-loop verification—it achieves the accurate generation and issuance of dynamic optimization instructions for energy transmission paths. The specific execution method includes the following steps:

[0097] Step 5.1: In the state awareness phase, each agent... Each corresponds to an energy access point. Energy access point Each intelligent agent includes photovoltaic units, and / or energy storage systems, and / or heat pump units. state space Defined as:

[0098] ;

[0099] in, For real-time power data, For load demand forecasting, For network topology information, the state space ensures that each agent... Both can make locally optimal decisions based on comprehensive environmental information.

[0100] The state space Integrates real-time power data collected by the data acquisition unit Load demand forecasting based on historical data and weather information using GRU neural networks. and network topology information obtained in real time by the SDN controller. This provides a global environment awareness foundation for instruction generation.

[0101] Step 5.2: In the action generation phase, the engine outputs power allocation instructions for each energy transmission path through the action space generation module based on the near-end policy optimization algorithm. Power allocation instructions for motion space The constraints are satisfied: ,in, This represents the total power.

[0102] Step 5.3: After the instructions are generated, the decision optimization engine adopts a multi-objective reward function. Guiding agents to cooperate in optimization is defined as follows:

[0103] ;

[0104] in, For multi-objective weighting coefficients, satisfying ; This is the dynamic energy efficiency target value; Actual operating energy efficiency; For dynamic cost budgeting; Actual costs incurred; The rated frequency; This refers to the actual system frequency. This represents the allowable frequency deviation.

[0105] Step 5.4: The decision optimization engine achieves online learning through the policy update module, policy function. parameters Updated using gradient ascent: ,in To accumulate rewards, This is the learning rate.

[0106] Step 5.5: Generate dynamic optimization instructions to achieve adaptive control of the energy transmission path; the dynamic optimization instructions are: Before the dynamic optimization command is issued, the feasibility of the decision command is verified by using simulation results through the digital twin platform. The strategy is then iteratively optimized based on the real-time data stream. If the simulation deviation exceeds the threshold, re-optimization is triggered to form a closed-loop control.

[0107] The core of Communication Assurance Component 6 is the hardware deployment and optimization solution of an integrated predictive communication resource scheduling algorithm, which is used to dynamically allocate network resources according to energy transmission priority to ensure the reliability and low latency of data transmission. This component can be integrated into a campus core switch (Huawei CE6857 series) that supports SDN (Software Defined Networking). Its specific implementation aims to build a low-latency, highly reliable predictive communication network.

[0108] like Figure 3 As shown, the communication assurance element performs the following steps:

[0109] Step 6.1: Construct a hybrid neural network model based on historical communication data and real-time multi-source energy consumption data. The hybrid neural network model includes an input layer, a hidden layer, and an output layer.

[0110] The input layer receives real-time parameters from the data acquisition unit. ) and historical communication traffic Real-time parameters ( Includes light intensity Energy storage charging and discharging status Heat pump operation mode Environmental parameters are used to characterize external operating conditions that affect communication requirements.

[0111] The hidden layer uses a bidirectional long short-term memory network to extract temporal features.

[0112] Output layer predicts the future Communication demand vector of each energy node within a given time period:

[0113] ,

[0114] in, , , These represent the predicted communication demand intensity of the photovoltaic, energy storage, and heat pump nodes at time t+Δt, respectively.

[0115] Step 6.2: Construct a weighted multi-objective function:

[0116] ;

[0117] in, This is a transmission delay optimization term, reflecting the degree of matching between the average delay of data packets and the priority of energy transmission; This is a bandwidth utilization optimization item to ensure that high-priority energy data does not occupy core bandwidth resources; This is an energy efficiency optimization item to reduce the energy consumption of the communication module; Let be the weighting coefficient, satisfying And it is dynamically adjusted using the entropy weight method to adapt to real-time operating conditions; Let be the communication priority weight of the i-th energy node; Let be the data transmission delay of the i-th node; The available bandwidth of the i-th node; To allocate bandwidth in practice; To ensure minimum bandwidth for each link; This is the sum of the total bandwidth capacity; The energy consumption of the kth communication node; N represents the total energy consumption; M represents the total number of energy nodes; K represents the total number of communication links; and K represents the total number of communication nodes (switches, gateways, etc.).

[0118] Step 6.3: Execute the predictive communication resource scheduling algorithm, which uses the communication demand prediction vector obtained in Step 6.1. Using real-time network topology state information as the core input, a constrained dynamic resource optimization problem is constructed; and the bandwidth allocation vector is iteratively solved on an embedded CPU using the gradient descent method. Minimize the weighted multi-objective function defined in step 6.2 Mathematical expression is The constraints include total bandwidth capacity limitations. and minimum bandwidth guarantee for each link Its mathematical expression is:

[0119] ;

[0120] in, Assign a vector to the bandwidth; The maximum allowable energy consumption; each element in the vector Representative assigned to the first The bandwidth of each communication link Index of communication links The algorithm uses gradient descent for iterative solution, and the update formula is:

[0121] ;

[0122] in, This represents the number of iterations. For the objective function The gradient is calculated; during the optimization process, an online simulation is performed using a digital twin platform to verify the feasibility of the allocation strategy, and the optimization parameters are adjusted based on the simulation results.

[0123] The algorithm operates in a rolling time-domain control mode, re-predicting and optimizing at regular intervals; it updates the scheduling strategy based on the latest data and prediction results, forming a closed-loop feedback, and ultimately obtains the optimal objective function. .

[0124] Then the optimal objective function Substituting these values ​​into the formula for iteratively solving the bandwidth allocation vector, the optimal bandwidth allocation vector is obtained. ;

[0125] Step 6.4: Issuance and execution of the optimized resource allocation scheme: The dynamic resource allocation module receives the optimal bandwidth allocation vector obtained from Step 6.3. Subsequently, the module will transfer the vector The configuration commands are converted into a set of bandwidth configuration instructions that can be recognized by specific network devices. These instructions are then distributed to the corresponding switches and routers in the campus network via the southbound interface of the software-defined network (SDN) controller. The network devices dynamically adjust their port or queue bandwidth allocation strategies according to the instructions, thereby ensuring that the data transmission links of energy nodes such as photovoltaic, energy storage, and heat pumps achieve optimal bandwidth allocation. The defined bandwidth resources ultimately achieve synergistic optimization of transmission delay, bandwidth utilization, and system energy efficiency.

[0126] The Digital Twin Platform 7 serves as the system's central hub, and its implementation is manifested in the construction of virtual images, the integration of multi-step look-ahead simulation, and closed-loop control throughout the entire process. This platform can be deployed on high-performance workstations (HP Z8 G4) or private cloud platforms to build virtual images and synchronize data in real time with the underlying physical system (data acquisition unit) via API interfaces.

[0127] The digital twin platform 7 is coupled to the data acquisition unit, decision optimization engine, and communication support components to construct a virtual image of the smart park energy system and achieve online simulation and closed-loop control. The virtual image integrates physical entity mapping relationships, including the energy equipment topology, energy flow dynamics, and communication network status, and maintains consistency with the physical system through a real-time data synchronization mechanism. The virtual image construction is based on multi-source heterogeneous data fusion, mathematically expressed as:

[0128] ,

[0129] in, For a moment The virtual mirror state vector, For communication data streams, This is a mapping function that enables precise projection of the physical system into the virtual space. The total power of the system. This refers to network topology information.

[0130] In the simulation phase, the digital twin platform runs multi-step look-ahead simulation based on historical data and real-time multi-source energy consumption data to generate a virtual energy flow state. In the decision-making phase, the decision optimization engine calculates the priority score of each energy transmission path based on the virtual energy flow state using a multi-agent reinforcement learning algorithm and generates dynamic optimization instructions. The communication assurance element performs predictive resource scheduling according to the priority score, ultimately realizing the overall dynamic optimization and self-healing control of the smart park energy transmission system.

[0131] The execution steps of the digital twin platform include:

[0132] Step 7.1: Couple the digital twin platform to the data acquisition unit, decision optimization engine, and communication support components to build a multi-dimensional virtual image of the smart park energy system.

[0133] Step 7.2: In the simulation phase, the digital twin platform runs multi-step look-ahead simulations based on historical data and real-time multi-source energy consumption data. The simulation model integrates energy transmission dynamics equations and communication delay characteristics to generate future... Virtual energy flow state over time Multi-step simulation is achieved through iterative calculations, as shown in the formula:

[0134] ;

[0135] in, To control the input vector, It is a system dynamic function that encompasses changes in energy flow and network constraints; and The first Step and the first The virtual system state vector of the step contains information such as the power, load, and network status of each node.

[0136] Step 7.3: In the decision-making phase, the digital twin platform and the decision optimization engine form a closed-loop interaction: First, the virtual energy flow state generated in Step 7.2 is... This serves as input to the multi-agent reinforcement learning algorithm within the decision optimization engine. Based on this input, the decision optimization engine calculates the priority scores for each energy transmission path. This calculation process is part of the multi-agent reinforcement learning algorithm, which evaluates the energy transmission paths in a virtual state. Different actions (power allocation instructions) will result in future cumulative rewards (determined by the reward function). The platform uses the expected value of the definition to indirectly assign priority scores to each path, with higher-scoring paths corresponding to better power allocation strategies. At the same time, the platform verifies the feasibility of decision instructions through online simulation and dynamically adjusts virtual image parameters based on deviation detection to improve simulation accuracy and robustness.

[0137] Step 7.4: The digital twin platform supports the dynamic adaptive and self-healing control of the energy transmission system through continuous iterative optimization. The optimization objective is: ,in, This represents the actual system state.

[0138] The platform continuously compares the virtual state using a deviation detection algorithm. Compared with the actual state To minimize the difference between the virtual and actual states, if the deviation norm exceeds a threshold, the virtual mirror parameters are dynamically adjusted to improve simulation accuracy.

[0139] Ultimately, the platform uses the timing feedback mechanism in the model predictive control framework to send the optimized command timing to the physical system, forming a closed-loop feedback to support the system's dynamic adaptive and self-healing control. If an overload is predicted in the simulation, the energy routing is adjusted in advance through the decision engine.

[0140] Figure 4 Through a 30-day continuous experiment under the same park environment, the key performance indicators of the system of this invention and the traditional system were compared. In the experimental setup, the system of this invention employed a dynamic optimization mechanism based on multi-agent reinforcement learning, while the traditional system used fixed threshold control. The results show that, in terms of energy efficiency, the system of this invention achieves significantly higher actual operating energy efficiency. With target energy efficiency The deviation of this invention is consistently controlled within 5%, while the deviation of traditional systems reaches 28%; in terms of cost control, the actual operating cost of this invention is significantly lower. Compared to budget The system achieved a 18.3% reduction, while the traditional system exceeded the budget by 12.7%; regarding frequency stability, the frequency deviation of the system in this invention was... The frequency deviation remained consistently within the allowable range of ±0.15Hz, while the traditional system exhibited multiple instances of fluctuations exceeding the limit of ±0.5Hz. These data are clearly presented using a dual Y-axis line graph, with the left axis displaying the energy efficiency percentage and the right axis displaying the cost ratio. The frequency deviation range is also marked with shaded areas, visually verifying the synergistic optimization effect of the multi-objective reward function.

[0141] Figure 5Based on the multi-step look-ahead simulation function of the digital twin platform, the dynamic energy dispatch process during a specific time period (peak solar power output at midday in summer) was demonstrated. The simulation model integrates equipment such as a photovoltaic array (peak power 250kW), an energy storage system (capacity 500kWh), and a heat pump unit (cooling capacity 200kW), generating a virtual energy flow state for the next 15 minutes through iterative calculations. The diagram uses different colored streamlines to represent the direction of energy flow (red arrows represent solar power charging energy storage, green arrows represent energy storage supplying power to critical loads), with streamline width proportional to power output. Simulation results show that during peak solar power output (12:00-14:00), the system automatically dispatches 62% of the excess solar power (peak 180kW) to the energy storage unit for charging, while simultaneously increasing the energy storage discharge power to 85% of its rated value, reducing the amount of electricity purchased from the external grid by 43.2%. The diagram also uses a topology subgraph to show the path changes before and after optimization, demonstrating that digital twin-driven look-ahead dispatch can improve energy transmission efficiency by 31.5%.

[0142] Figure 6 The system response was tested by injecting three typical fault scenarios: a sudden disconnection of the photovoltaic array at t=5 min (power plummeting from 150kW to 0), a communication delay at t=12 min (increasing dramatically from 10ms to 3s), and a sudden load surge at t=20 min (35% increase over the baseline). Experimental data are presented as a cluster of time-series curves, including the norm of deviation between the virtual and actual states. Parameters such as critical line power fluctuations and system frequency changes were analyzed. Results showed that in a photovoltaic disconnection fault, the system re-optimized the path within 4.2 seconds through the decision engine, maximizing the energy storage discharge power, and the frequency deviation recovered to the allowable range within 8 seconds. In a communication delay scenario, the predictive communication scheduling algorithm completed bandwidth reallocation within 2.3 seconds, ensuring priority transmission of control commands. In a load surge test, the system proactively activated backup power through the advanced prediction of the digital twin platform, keeping voltage fluctuations within ±5%. Compared to traditional systems (which require manual intervention after a fault and have a recovery time exceeding 3 minutes), the system of this invention achieved self-healing recovery within 30 seconds in all test scenarios.

[0143] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered by the present invention.

Claims

1. A multimodal fusion smart park energy management and control system, characterized in that, include: Data acquisition unit (1) is used to collect multi-source energy consumption data; The decision optimization engine (5) has a built-in dynamic energy path optimizer based on multi-agent reinforcement learning, which is used to perform fusion analysis on multi-source energy consumption data and generate optimization decision instructions for energy transmission paths. The communication guarantee component (6) integrates a predictive communication resource scheduling algorithm to dynamically allocate network resources according to energy transmission priority, ensuring the reliability and low latency of data transmission; The digital twin platform (7), as the system hub, is coupled to the data acquisition unit (1), the decision optimization engine (5) and the communication support element (6) to build a virtual image of the smart park energy system and realize online simulation and closed-loop control; In the simulation phase, the digital twin platform (7) runs multi-step forward simulation based on historical data and real-time multi-source energy consumption data to generate virtual energy flow status; During the decision-making phase, the decision optimization engine (5) calculates the priority score of each energy transmission path based on the virtual energy flow state through a multi-agent reinforcement learning algorithm and generates dynamic optimization instructions; the communication guarantee element (6) performs predictive resource scheduling according to the priority score, and finally realizes the overall dynamic optimization and self-healing control of the smart park energy transmission system.

2. The multimodal fusion smart park energy management system according to claim 1, characterized in that: The data acquisition unit (1) includes a multi-energy access module (2), an intelligent routing device (3), and an electronic counting sensor module (4). The data acquisition unit (1) collects multi-source energy consumption data through the multi-energy access module (2) and the intelligent routing device (3); The electronic counting sensor module (4) achieves microsecond-level data acquisition based on electronic energy parameters, and the acquired energy value The calculation method is as follows: ; in Let be Planck's constant. At the speed of light, For electron wavelength, For a moment The number of electrons detected inside, Energy conversion efficiency.

3. The multimodal fusion smart park energy management system according to claim 2, characterized in that: The multi-source energy consumption data includes photovoltaic power, energy storage power, and heat pump power; the operation mode of the data acquisition unit (1) to acquire multi-source energy consumption data includes the following steps: Step 1.1: Connect the multi-energy access module (2) to photovoltaic energy, energy storage system, heat pump system and other distributed energy, and dynamically calculate the real-time power based on the physical parameters of each energy source; The power calculation of photovoltaic energy depends on time. Light intensity Photovoltaic panel area and photoelectric conversion efficiency Its output power is: ; Power calculations for energy storage systems are based on the rate of energy change. and charge / discharge efficiency Its output power is: ; Power calculations for heat pump systems involve either cooling or heating capacity. and performance coefficient Its output power is: ; Step 1.2: Couple the intelligent routing device (3) to the multi-energy access module (2) to dynamically integrate multi-source power data and calculate the total power by superposition: ; in, Indicates the first Real-time power of each energy access point, including photovoltaic energy power. Energy storage system power Heat pump system power Other distributed energy power Its set is ; The intelligent routing device receives real-time load forecast values ​​output from the load forecasting module. And based on the calculated real-time load demand changes ;in The load prediction value at the previous sampling time is used to dynamically select the data transmission path by comprehensively comparing the service priority of each data stream with the current network status, so as to minimize transmission delay and packet loss rate, thereby optimizing the data stream path.

4. A multimodal fusion smart park energy management system according to claim 1, 2, or 3, characterized in that: The data acquisition unit (1) collaborates with the digital twin platform (7) through a timing synchronization mechanism. During the simulation phase, the sampling interval is... Upload multi-source energy consumption data according to the simulation requirements of the digital twin platform (7), and dynamically adjust the acquisition frequency according to the optimization instructions during the decision-making stage. With routing strategies.

5. The multimodal fusion smart park energy management system according to claim 1, characterized in that: The decision optimization engine (5) incorporates a multi-agent reinforcement learning framework. Through the coordinated operation of five stages—state perception, action generation, reward optimization, policy update, and closed-loop verification—it achieves the accurate generation and issuance of dynamic optimization instructions for energy transmission paths. The specific execution method includes the following steps: Step 5.1: In the state awareness phase, each agent... Each corresponds to an energy access point. Energy access point Each intelligent agent includes photovoltaic units, and / or energy storage systems, and / or heat pump units. state space Defined as: ； in, For real-time power data, For load demand forecasting, For network topology information, the state space ensures that each agent... Both can make locally optimal decisions based on comprehensive environmental information; The state space Integrates real-time power data collected by the data acquisition unit Load demand forecasting based on historical data and weather information using GRU neural networks. and network topology information obtained in real time by the SDN controller. This provides a global environment awareness foundation for instruction generation; Step 5.2: In the action generation phase, the engine outputs power allocation instructions for each energy transmission path through the action space generation module based on the near-end policy optimization algorithm. Power allocation instructions for motion space The constraints are satisfied: ,in, Total power; Step 5.3: After the instruction is generated, the decision optimization engine (5) adopts a multi-objective reward function. Guiding agents to cooperate in optimization is defined as follows: ； in, For multi-objective weighting coefficients, satisfying ; This is a dynamic energy efficiency target value; Actual operating energy efficiency; For dynamic cost budgeting; Actual costs incurred; The rated frequency; This refers to the actual system frequency. For frequency tolerance; Step 5.4: The decision optimization engine (5) implements online learning through the policy update module, and the policy function... parameters Updated using gradient ascent: ,in To accumulate rewards, The learning rate; Step 5.5: Generate dynamic optimization instructions to achieve adaptive control of the energy transmission path; the dynamic optimization instructions are: Before the dynamic optimization instruction is issued, the simulation results are used to verify the feasibility of the decision instruction through the digital twin platform (7), and the optimization strategy is iterated according to the real-time data stream. If the simulation deviation exceeds the threshold, re-optimization is triggered to form a closed-loop control.

6. The multimodal fusion smart park energy management system according to claim 1, characterized in that: The communication protection element (6) performs the following steps: Step 6.1: Construct a hybrid neural network model based on historical communication data and real-time multi-source energy consumption data. The hybrid neural network model includes an input layer, a hidden layer, and an output layer. The input layer receives real-time parameters from the data acquisition unit (1). ) and historical communication traffic Real-time parameters ( Includes light intensity Energy storage charging and discharging status Heat pump operation mode Environmental parameters are used to characterize external operating conditions that affect communication requirements; The hidden layer uses a bidirectional long short-term memory network to extract temporal features; Output layer predicts the future Communication demand vector of each energy node within a given time period: ， in, , , These represent the predicted communication demand intensity of the photovoltaic, energy storage, and heat pump nodes at time t+Δt, respectively. Step 6.2: Construct a weighted multi-objective function: ； in, This is a transmission delay optimization term, reflecting the degree of matching between the average delay of data packets and the priority of energy transmission; This is a bandwidth utilization optimization item to ensure that high-priority energy data does not occupy core bandwidth resources; This is an energy efficiency optimization item to reduce the energy consumption of the communication module; For the weighting coefficients, satisfying And it is dynamically adjusted using the entropy weight method to adapt to real-time operating conditions; Let be the communication priority weight of the i-th energy node; Let be the data transmission delay of the i-th node; The available bandwidth of the i-th node; To allocate bandwidth in practice; To ensure minimum bandwidth for each link; This is the sum of the total bandwidth capacity; The energy consumption of the k-th communication node; N represents the total energy consumption; M represents the total number of energy nodes; K represents the total number of communication links; and M represents the total number of communication nodes. Step 6.3: Execute the predictive communication resource scheduling algorithm, which uses the communication demand prediction vector obtained in Step 6.

1. Using real-time network topology state information as the core input, a constrained dynamic resource optimization problem is constructed; and the bandwidth allocation vector is iteratively solved on an embedded CPU using the gradient descent method. Minimize the weighted multi-objective function defined in step 6.2 Mathematical expression is The constraints include total bandwidth capacity limitations. and minimum bandwidth guarantee for each link Its mathematical expression is: ； in, Assign a vector to the bandwidth; The maximum allowable energy consumption; each element in the vector Representative assigned to the first The bandwidth of each communication link Index of communication links The algorithm uses gradient descent for iterative solution, and the update formula is: ； in, This represents the number of iterations. For the objective function The gradient; during the optimization process, the digital twin platform (7) is used for online simulation to verify the feasibility of the allocation strategy, and the optimization parameters are adjusted according to the simulation results; The algorithm operates in a rolling time-domain control mode, re-predicting and optimizing at regular intervals; it updates the scheduling strategy based on the latest data and prediction results, forming a closed-loop feedback, and ultimately obtains the optimal objective function. ; Then the optimal objective function Substituting these values ​​into the formula for iteratively solving the bandwidth allocation vector, the optimal bandwidth allocation vector is obtained. ; Step 6.4: Issuance and execution of the optimized resource allocation scheme: The dynamic resource allocation module receives the optimal bandwidth allocation vector obtained from Step 6.

3. Subsequently, the module will transfer the vector The configuration commands are converted into a set of bandwidth configuration instructions that can be recognized by specific network devices. These instructions are then distributed to the corresponding switches and routers in the campus network via the southbound interface of the software-defined network (SDN) controller. The network devices dynamically adjust their port or queue bandwidth allocation strategies according to the instructions, thereby ensuring that the data transmission links of energy nodes such as photovoltaic, energy storage, and heat pumps achieve optimal bandwidth allocation. The defined bandwidth resources ultimately achieve synergistic optimization of transmission delay, bandwidth utilization, and system energy efficiency.

7. The multimodal fusion smart park energy management system according to claim 1, characterized in that: The execution steps of the digital twin platform (7) include: Step 7.1: Couple the digital twin platform (7) to the data acquisition unit (1), the decision optimization engine (5) and the communication support element (6) to build a multi-dimensional virtual image of the smart park energy system; Step 7.2: In the simulation phase, the digital twin platform (7) runs multi-step look-ahead simulations based on historical data and real-time multi-source energy consumption data. The simulation model integrates energy transmission dynamics equations and communication delay characteristics to generate future... Virtual energy flow state over time Multi-step simulation is achieved through iterative calculations, as shown in the formula: ， in, To control the input vector, It is a system dynamic function that encompasses changes in energy flow and network constraints; and The first Step and the first The virtual system state vector of each step contains information such as power, load, and network status of each node; Step 7.3: In the decision-making stage, the digital twin platform (7) and the decision optimization engine (5) form a closed-loop interaction: First, the virtual energy flow state generated in step 7.2 is... As input to the multi-agent reinforcement learning algorithm in the decision optimization engine (5); the decision optimization engine (5) calculates the priority score of each energy transmission path based on this input, and the calculation process is part of the multi-agent reinforcement learning algorithm, which evaluates the priority score of each energy transmission path in the virtual state. The platform indirectly assigns priority scores to each path by calculating the expected value of future cumulative rewards for different actions. Paths with higher scores correspond to better power allocation strategies. At the same time, the platform verifies the feasibility of decision instructions through online simulation and dynamically adjusts virtual image parameters based on deviation detection to improve simulation accuracy and robustness. Step 7.4: The digital twin platform supports the dynamic adaptive and self-healing control of the energy transmission system through continuous iterative optimization. The optimization objective is: ,in, This represents the actual system state. The platform continuously compares the virtual state using a deviation detection algorithm. Compared with the actual state To minimize the difference between the virtual and actual states, if the deviation norm exceeds the threshold, the virtual mirror parameters are dynamically adjusted to improve simulation accuracy. Ultimately, the platform uses the timing feedback mechanism in the model predictive control framework to send the optimized command timing to the physical system, forming a closed-loop feedback to support the system's dynamic adaptive and self-healing control. If an overload is predicted in the simulation, the energy routing is adjusted in advance through the decision engine.

8. The multimodal fusion smart park energy management system according to claim 7, characterized in that: The virtual image integrates physical entity mapping relationships, including energy equipment topology, energy flow dynamics, and communication network status, and maintains consistency with the physical system through a real-time data synchronization mechanism. The virtual image construction is based on multi-source heterogeneous data fusion, mathematically expressed as: ， in, For a moment The virtual mirror state vector, For communication data streams, This is a mapping function that enables precise projection of the physical system into the virtual space. The total power of the system. This refers to network topology information.

Images