Modular data center shelter collaborative communication method and system

Abstract

This invention discloses a modular data center container collaborative communication method and system, including receiving uploaded 3D drawings, identifying container categories and constructing digital twin models; collecting data from each container in real time, and updating the state of the digital twin model in real time based on the container data; predicting the future state of the container based on the real-time updated digital twin model; performing global optimization adjustments based on the container category and the predicted future state; issuing and executing global optimization commands, and repeating the above steps. The system includes a central digital twin system, edge computing nodes deployed in each container, and a communication network. By constructing a global digital twin model and combining it with a model predictive control framework, this invention achieves advanced prediction of IT load and cooling demand, as well as multi-objective global optimization scheduling, effectively solving the problems of control lag and low energy efficiency in existing container systems, and improving the overall operational energy efficiency and security of modular data centers.

Description

Technical Field

[0001] This invention relates to a modular data center container collaborative communication method and system, belonging to the field of communication control technology. Background Technology

[0002] With the rapid development of cloud computing, artificial intelligence, and big data, traditional data center construction models are no longer sufficient to meet the demands of explosive growth in computing power and rapid deployment. Modular data center modules (referred to as modules) have emerged to address this need. A module is a prefabricated modular unit that highly integrates the physical infrastructure required for a data center, such as IT cabinets, power distribution units, cooling equipment, fire protection, and security monitoring, within a factory. Once transported to the site, it can be put into use through simple assembly and pipeline connections. The main function of modules is to enable on-demand expansion of data centers, shorten construction cycles, and improve the utilization rate of physical space. Typical large-scale modular data centers usually contain heterogeneous modules performing different specialized tasks, such as IT computing power modules, integrated cooling station modules, hydraulic module modules, and power distribution modules.

[0003] Currently, existing data center container communication and control systems mostly adopt a "distributed + passive response" management architecture. Each container is typically equipped with an independent Business Management System (BMS), which relies on physical sensors (such as thermometers and flow meters) deployed within the container to collect data in real time and uses traditional PID control logic to locally regulate the equipment within the container. For example, when the sensors in the IT container detect that the cabinet intake air temperature has risen and exceeded a set threshold, the system will send a cooling command to the chiller container via the communication network, and the chiller will then increase its cooling capacity.

[0004] However, unlike the ideal situation, existing modular hospitals have the following obvious drawbacks in actual multi-module collaborative operation: Information silos exist between the various modules, and control strategies lag behind, resulting in low energy efficiency. The existing communication and control architecture lacks a global perspective for deep state awareness and forward-looking prediction. Due to the inherent physical delay in the transfer of cooling capacity from the chiller module to the IT module via hydraulic modules, traditional passive response control often only begins to increase cooling capacity after a sudden surge in IT load and temperatures exceeding limits, exhibiting significant lag. Furthermore, because it is impossible to predict IT load trends and cooling demand in advance, the system often keeps cooling equipment in a highly redundant and inefficient operating state for extended periods for safety reasons, resulting in a persistently high overall PUE (Power Usage Effectiveness) for the data center.

[0005] Lacking global dynamic constraints and fault-tolerance mechanisms, existing modular shelter systems exhibit poor robustness in handling complex operating conditions and extreme anomalies. When faced with multi-objective conflicts (such as balancing energy conservation with ensuring local temperatures do not exceed limits), they lack the ability to coordinate solutions comprehensively. Especially when encountering sudden extreme loads, cooling tower failures, or interruptions in inter-cabin communication networks, existing systems often only execute rigid hard alarms or partial shutdowns, failing to intelligently coordinate equipment in each cabin at the global level (such as dynamically relaxing temperature constraints, calling upon the power distribution compartment's battery to discharge, and initiating local autonomy). This greatly increases the risk of localized or even global service outages and downtime in the data center. Summary of the Invention

[0006] The technical problem that this invention aims to solve is that existing mobile shelters operate in isolation and lack a coordinated mechanism based on global state awareness and forward-looking prediction, resulting in lagging system power supply scheduling and a lack of effective global optimization constraints and fault-tolerant control capabilities when facing complex operating conditions.

[0007] To address the aforementioned technical problems, this invention proposes a modular data center container collaborative communication method, comprising the following steps: S1. Receive uploaded 3D drawings, identify the type of modular shelter, and build a digital twin model; S2. Collect data from each mobile shelter in real time and update the status of the digital twin model in real time based on the data. S3. Predict the future state of the cabin based on a real-time updated digital twin model; S4. Perform global optimization and adjustment based on the type of mobile cabin and prediction of its future status; S5. Issue and execute the global optimization instruction, and repeat S2 to S5.

[0008] The categories of mobile shelters identified in S1 include: Based on the drawing analysis results, three types of features are extracted: component features, spatial layout features, and textual semantic features. The three types of features are concatenated into a comprehensive feature vector, which is then input into a multimodal classification model composed of a CNN+RNN fusion network. The multimodal classification model outputs the probability of the container class based on a predefined class library; The predefined category library includes integrated cooling plant cabins, hydraulic module cabins, power distribution cabins, IT computing power cabins, and prefabricated component cabins.

[0009] The construction of the digital twin model in S1, as described above, includes: Retrieve the parametric template model corresponding to the category of the makeshift hospital. The parametric template model includes the geometric model skeleton, physical model equations, and data-driven model structure. Instantiate the template model based on the equipment model or dimensions in the drawings; Based on the pipeline topology in the drawings, connect the various equipment models to form a complete modular digital twin model; Perform consistency verification on the digital twin model, and assign a unique identifier to it and store it in the model library after it passes the verification.

[0010] The data collected in real time for each mobile shelter in S2, as mentioned above, includes: Deploy edge computing nodes within each shelter; the edge computing nodes are preloaded with a lightweight model summary of the target shelter. Edge computing nodes collect real-time data at preset intervals, and the lightweight model summary includes sensor locations, measurement ranges, and sampling frequencies. The collected data from the makeshift hospitals were preprocessed, including missing value imputation, noise filtering, outlier removal, unit conversion, and data compression. The preprocessed data is packaged according to a unified timestamp format and uploaded to the central digital twin system in real time via a communication network.

[0011] The above-mentioned updates to the digital twin model state in S2 include: Based on the unique identifier of the mobile cabin in the uploaded data packet, load the corresponding digital twin model instance from the model library; Associate and map real-time data with variables in the model; Key states that cannot be directly measured are estimated using state estimation algorithms; Update the mapping data and state estimation results to the state variable set of the digital twin model; The updated model undergoes a consistency check, and if the error exceeds a preset threshold, an online model correction mechanism is triggered.

[0012] The predictions made in S3 regarding the future state of the enemy cabin include: A model predictive control framework is adopted, which uses the currently updated digital twin model as a basis to make rolling predictions of the state evolution in the future prediction time domain; Forecasts include at least one of IT load forecasts, cooling demand forecasts, system energy efficiency forecasts, and external environment forecasts; The prediction employs a multi-model ensemble strategy, outputting predicted values ​​and confidence intervals.

[0013] The global optimization adjustments made in S4 as described above include: Construct a multi-objective optimization function, which includes the objectives of minimizing energy consumption, minimizing temperature deviation, ensuring equipment stability, and imposing safety penalties. Set constraints, including equipment capacity constraints, temperature constraints, power constraints, and coupling constraints; A hierarchical solution strategy is adopted, including upper-level equipment combination optimization and lower-level continuous variable optimization; Set up an extreme case handling mechanism to implement a constraint relaxation strategy or switch to emergency operation mode when there is no feasible solution to the optimization problem.

[0014] The global optimization instructions issued and executed in S5, as mentioned above, include: The optimal control commands obtained from the optimization calculation are encapsulated in JSON format and distributed to the edge nodes of each module according to the modular unit grouping. After receiving the instruction, the edge node returns confirmation information, including confirmation that the instruction has been received, confirmation that the parsing was successful, and confirmation that the instruction has been executed. The system has a preset timeout threshold; if no confirmation is received, a retransmission mechanism is triggered. The edge nodes feed back the actual execution results to the system in real time. The system compares the deviation between the actual value and the instruction value. If the deviation exceeds the deviation threshold, an execution deviation alarm is recorded. When the optimization problem has no feasible solution, the system executes a constraint relaxation strategy or an emergency operation mode.

[0015] A modular data center container collaborative communication system uses a modular data center container collaborative communication method, including a central digital twin system, edge computing nodes, and a communication network; The central digital twin system is used to perform tasks such as identifying the type of mobile shelter, building and updating the digital twin model, predicting future states, and making global optimization adjustments, as well as issuing optimization instructions. Edge computing nodes are deployed in each shelter to collect shelter data in real time, perform data preprocessing, receive and execute optimization instructions, and switch to local autonomous mode when communication is interrupted. A communication network is used for bidirectional data interaction between the central digital twin system and edge computing nodes.

[0016] The aforementioned edge computing node includes a multi-protocol access module, an analog / digital input module, a network communication module, a local storage module, an edge computing module, and an instruction parsing and execution module; Multi-protocol access module, used to support connection of sensors and device controllers using multiple industrial protocols; Analog / digital input module, used to connect analog sensors and switch signals; The network communication module is used for bidirectional data interaction with the central system; Local storage module, used for data caching; The edge computing module is used to run data preprocessing algorithms and local emergency control logic; The instruction parsing and execution module is used to convert instructions into signals that the device can recognize and drive the device to perform actions, as well as to provide feedback on the actual execution results.

[0017] This invention has positive effects: (1) This invention extracts three types of features from the drawings and inputs them into a CNN+RNN fusion network to realize intelligent identification of the type of shelter and automated construction of digital twin model. On this basis, the key states that cannot be directly measured are calculated by combining the state estimation algorithm, breaking the physical space limitation and constructing a high-fidelity global data mirror of the shelter.

[0018] (2) The present invention adopts a multi-model integration strategy to make rolling predictions of IT load, cooling demand and system energy efficiency; by introducing cooling transmission delay modeling, the system can predict the future state evolution trend in advance, thereby guiding the early action of equipment such as cooling stations, effectively eliminating the control lag problem caused by physical delay and improving the overall energy utilization efficiency.

[0019] (3) The present invention constructs a multi-objective optimization function based on the model predictive control (MPC) framework for global optimization. It has a complete instruction issuance closed-loop mechanism. When faced with extreme situations such as no optimization solution or sudden failure, the system can intelligently execute constraint relaxation strategy or switch to fast extreme response mode, which greatly enhances the security of modular data centers under complex working conditions and communication anomalies. Attached Figure Description

[0020] The invention will now be further described with reference to the accompanying drawings.

[0021] Figure 1 This is a schematic diagram of the steps; Figure 2 This is a flowchart. Detailed Implementation

[0022] Example 1 See Figures 1 to 2 This embodiment includes the following steps: S1. Receive uploaded 3D drawings, identify the type of modular shelter, and build a digital twin model; S2. Collect data from each mobile shelter in real time and update the status of the digital twin model in real time based on the data. S3. Predict the future state of the cabin based on a real-time updated digital twin model; S4. Perform global optimization and adjustment based on the type of mobile cabin and prediction of its future status; S5. Issue and execute the global optimization instruction, and repeat S2 to S5.

[0023] The 3D drawings in S1 above are uploaded through the system, which automatically parses the drawing content. Supported drawing formats include IFC, RVT, DWG, and PDF containing vector information. If it is a BIM model (IFC / RVT), then directly extract structured data such as component type, geometric dimensions, spatial location, material properties, and equipment model; If it is DWG / PDF, then computer vision and OCR technology are used to identify graphic elements and text annotations, and a mapping between graphic elements and physical objects is established by combining a predefined legend library; The predefined legend library is a basic knowledge base used to map two-dimensional graphic symbols to actual physical devices or components. It includes various graphic symbols commonly used in data center modular design and their corresponding device types, attributes, and semantic information.

[0024] The specific categories of mobile shelters are identified as follows: Based on the analysis results, three types of features are extracted: component features, spatial layout features, and textual semantic features. The component characteristics mainly include the type and quantity of the main equipment inside the cabin; Spatial layout features are used to analyze equipment layout density and pipeline routing patterns; Text semantic features are extracted from the title bar and device labels, and then converted into semantic vectors through a word embedding model; The three types of features are then concatenated into a comprehensive feature vector and input into a multimodal classification model composed of a CNN+RNN fusion network for calculation. The multimodal classification model outputs the probability of the cabin category based on a predefined category library.

[0025] Predefined categories include: integrated cooling plant cabins, hydraulic module cabins, power distribution cabins, IT computing power cabins, and prefabricated component cabins.

[0026] A multimodal classification model consists of an input layer, hidden layers, and an output layer. The input layer receives the fused feature vectors, the hidden layers are several fully connected layers, the activation function is ReLU, and Dropout is added to prevent overfitting. The output layer is a softmax layer, which outputs the probability value of each predefined category.

[0027] When a feature vector of a new drawing is input, the multimodal classification model performs forward calculations, and the softmax layer outputs a probability distribution; for example, the distribution includes 0.92 for the cooling station, 0.05 for the hydraulic module, 0.02 for the power distribution cabin, 0.01 for the IT computing power cabin, and 0.00 for the prefabricated components; the highest probability value is the confidence level that the cabin belongs to the corresponding category.

[0028] The system sets a confidence threshold. If the highest probability is greater than the confidence threshold, the category is directly used as the recognition result.

[0029] If the highest probability does not reach the confidence threshold, but the difference between the highest probability and the second highest probability is greater than or equal to the preset error value, the category will still be used as the recognition result.

[0030] When the highest probability is less than the confidence threshold, it means that the model is not confident enough in the recognition result, and the manual review process is automatically triggered: the drawings are pushed to the manual review interface, where designers can manually select the type of shelter and correct the annotation information in the drawings as needed.

[0031] The construction of the digital twin model is as follows: The system has built-in parametric template models for various types of modular shelters, including: geometric model skeleton, physical model equations, and data-driven model structure.

[0032] The geometric model skeleton includes the general shape of the equipment and the location of connection points; The physical model equations include the chiller COP characteristic curve, the pump head-flow equation, the thermodynamic balance equation, and the equation coefficients are undetermined parameters, etc. The data-driven model structure includes the number of neural network layers, the number of neurons, and the weights to be trained.

[0033] Then, based on the equipment model or size in the drawings, the template is instantiated. If a specific model is available, the actual performance parameters are retrieved from the equipment library and filled into the physical model equations. If no model is available, the equipment capacity is estimated based on the dimensions, and the default typical curve is used. The geometric model is scaled and positioned according to the actual layout dimensions. Based on the pipeline topology in the drawings, each equipment model is connected into a complete modular digital twin model. Finally, the model is verified for consistency. Once the consistency verification is passed, a unique identifier is assigned and stored in the model library.

[0034] S2. Collect data from each mobile shelter in real time and update the status of the digital twin model in real time based on the pre-processed mobile shelter data; The real-time data collection from the mobile hospital is as follows: Extract the sensor list and communication protocols for each modular unit from the digital twin model built by S1, such as Modbus address and BACnet object ID.

[0035] Edge computing nodes are deployed in each modular shelter. The edge computing nodes are preloaded with a lightweight model summary of the target modular shelter. The lightweight model summary includes sensor locations, ranges, sampling frequencies, and real-time data is collected at preset intervals. The IT modular housing mainly collects data on rack intake and exhaust temperatures, server power consumption, CPU utilization, and fan speed. The main data collection equipment in the chiller plant container includes chilled water supply and return temperatures, cooling water supply and return temperatures, chiller operating status and power, and water pump frequency.

[0036] The hydraulic module container mainly collects flow rate, pressure, and bypass valve opening.

[0037] The power distribution cabin mainly collects voltage, current, power factor, and UPS status.

[0038] The aforementioned edge computing nodes include the following functions: Multi-protocol access, supports common industrial protocols, and can connect to multiple types of sensors and device controllers simultaneously; Analog / digital inputs, such as 4-20mA and 0-10V analog sensors, as well as switch signals, can be directly connected through the I / O module; Network communication capabilities include wired Ethernet, 5G, and WiFi-6, supporting two-way data interaction with the central system. Local storage, equipped with large-capacity storage for data caching; Edge computing capabilities enable the execution of data preprocessing algorithms and local emergency control logic.

[0039] The collected data from the mobile shelter were preprocessed as follows: According to the data collection configuration file, the edge nodes acquire the following data through polling, subscription, and hardwired methods respectively: For protocols such as Modbus and BACnet, edge nodes actively poll each device at a set period to read register values.

[0040] For protocols that support publish or subscribe modes, such as OPC UA and MQTT, edge nodes subscribe to the required data points, and the devices actively push data.

[0041] For 4-20mA analog sensors, voltage and current signals are directly read through the analog acquisition module and converted into engineering values.

[0042] Different types of data use different sampling frequencies; For rapidly changing quantities such as voltage, current, power, and pressure, the sampling frequency is relatively high, generally set to 0.5 seconds; for slow variables such as temperature, humidity, and flow rate, the sampling frequency can be appropriately reduced, generally set to 5 seconds; for status quantities such as equipment start / stop status and fault signals, event triggering is generally used.

[0043] First, missing values ​​are handled. For missing values ​​caused by instantaneous sensor interruption or packet loss, a sliding window is filled based on the average of the first three sampled values ​​in the historical pattern. If the continuous missing time exceeds the missing threshold, the data point is marked as unavailable and a sensor health status alarm is triggered. Then, preliminary processing is performed to filter sensor noise; outliers that clearly exceed the range or physical limits are identified and removed; raw digital quantities, such as 4-20mA, are converted into actual physical values, such as temperature (°C) and pressure (kPa); and data with small continuous changes are compressed and encoded to reduce network transmission volume.

[0044] Finally, the preprocessed data is packaged according to a unified timestamp format and uploaded to the central digital twin system in real time via the communication network. The system defaults to periodic uploading, meaning all data is uploaded once at a preset fixed time. When network conditions are poor, it automatically switches to variable-rate uploading, meaning data is only uploaded when changes exceed a set dead zone, reducing network load. If a network outage occurs, it automatically enables resume upload, meaning data is cached locally and automatically re-uploaded when the network is restored.

[0045] The specific status of the digital twin model is as follows: After receiving a data packet uploaded by a mobile shelter, the central digital twin system first loads the corresponding digital twin model instance from the model library based on the unique identifier of the shelter in the data packet. This model instance includes the shelter's geometric model, physical model equations, data-driven model, and current state variables.

[0046] The received real-time data is associated and mapped with the variables in the model: Sensor data is directly assigned to the corresponding variables in the model; for example, the collected chilled water supply temperature value is directly assigned to the variables in the model.

[0047] If the data units are inconsistent with the internal units of the model, automatic conversion will be performed.

[0048] For spatial distribution data such as temperature field, the temperature values ​​are obtained from the geometric model of S1 based on the actual position of the sensor inside the cabin, and then assigned to the temperature variables of the corresponding spatial points in the model.

[0049] However, many critical states within a physical shelter, such as the evaporation temperature inside a chiller, the junction temperature of chips in IT equipment, and the temperature of the inner walls of pipes, cannot be directly measured by sensors. Digital twin models utilize physical equations and partial measurements to deduce these unmeasurable states through state estimation algorithms. Specific methods include: 1. A physical model-based observer uses the physical model of the chiller, inputs the measured evaporator inlet and outlet water temperatures, condenser inlet and outlet water temperatures, and chiller power, and estimates the evaporation and condensation temperatures in real time through a Kalman filter or a Luneburg observer. For IT equipment, the chip thermal resistance model is used, and the measured rack intake temperature and server power consumption are input to estimate the chip junction temperature.

[0050] 2. Data-driven compensation: For complex processes that are difficult to model precisely, a pre-trained regression model is used. Relevant measurements are input, and the estimated unmeasurable state is directly output. For example, based on chilled water flow rate, supply and return water temperature difference, and chiller power, the refrigerant flow rate or heat exchange efficiency inside the chiller is estimated by using a regression model.

[0051] 3. Mass balance and energy balance calculations: Based on measurable inlet and outlet parameters, the internal state is estimated.

[0052] For example, the actual cooling capacity of the chiller is calculated based on the inlet and outlet temperature difference and flow rate of the chilled water side and the cooling water side, and then the heat transfer coefficients of the evaporator and condenser are deduced.

[0053] Finally, the directly mapped data and the state estimation results are updated together in the state variable set of the digital twin model. At this point, all the key variables of the model are consistent with the current state of the physical shelter.

[0054] The current status of the physical container includes: temperature, pressure, flow rate, power, frequency, liquid level, equipment start / stop status, valve on / off status, fault status, operating mode, cumulative operating time, cumulative energy consumption, number of start / stop cycles, etc.

[0055] After the state update is complete, the system will perform a consistency check to verify whether the model accurately reflects the behavior of the physical entity. This is done by comparing certain measurable outputs calculated by the model with actual sensor values ​​to calculate the error. If the error exceeds a preset state error threshold, it indicates that the model may deviate from reality, triggering the online model correction mechanism. For physical models, algorithms such as recursive least squares or Kalman filtering are used to adjust key parameters in the model online, such as heat transfer coefficient and drag coefficient, so that the model output is closer to reality. For data-driven models such as neural networks, newly collected data is added to the training set to perform incremental learning or retraining of the model. If extreme cases are encountered, if the current model structure cannot accurately describe the actual process, such as equipment failure or performance degradation, the system can try to switch or combine different sub-models to best match the current operating conditions. The corrected model parameters are saved and used for subsequent prediction and optimization.

[0056] The updated digital twin model status will be displayed in real time through the visualization module, showing the color changes of each device in the 3D model, real-time data curves and dashboards, and alarm prompts for abnormal status. At the same time, the updated model status serves as the input for step S3 for future status prediction.

[0057] S3. Predict the future state of the cabin based on a real-time updated digital twin model; The specific predictions for the future state are as follows: The Model Predictive Control (MPC) framework is adopted, which uses the currently updated digital twin model as a basis to make rolling predictions of the state evolution of each cabin in the future time domain.

[0058] The forecast includes IT load, cooling demand, system energy efficiency, and external environment, as detailed below: 1. IT load forecasting The purpose of IT load forecasting is to estimate the power consumption and heat generation of IT equipment over a period of time in the future. Input historical IT load data and time characteristics; Historical IT load data includes the total power consumption of IT equipment, power consumption of each rack, and average CPU utilization over the past 24 hours; Time characteristics include the current time, the time of day, the day of the week, and whether it is a holiday; First, data preprocessing is performed by resampling historical IT load data at fixed time intervals to form a time series; outliers caused by sudden jumps due to sensor failures are removed and replaced with neighboring values; the data is normalized so that it falls within the [0,1] interval, which facilitates model training.

[0059] Then, feature engineering is performed to obtain the current hour (0-23), day of the week (0-6), whether it is a working day (0 / 1), and whether it is a holiday (0 / 1) to construct time features; obtain the load values ​​of the past 1 hour, 2 hours, and 24 hours to construct lag features; and obtain the load average, maximum, minimum, and rate of change of the past 1 hour to construct statistical features.

[0060] The system then runs multiple prediction models simultaneously, improving prediction accuracy through an integration strategy: Long Short-Term Memory (LSTM) network: Input the load sequence of the past 6 hours, the LSTM layer extracts the long-term dependencies of the time series, and the fully connected layer outputs the load sequence of the next 15 minutes to capture periodic patterns. Gated Recurrent Unit (GRU): The input is the same load sequence as LSTM. The GRU layer extracts features, and the fully connected layer outputs them for real-time prediction. Extreme Gradient Boosting (XGBoost): Input the current time features, lag features, and statistical features, integrate multiple decision trees, and directly output the average load for the next 15 minutes, used to process tabular features.

[0061] Next, the prediction results of multiple models are weighted and averaged. The weights are dynamically adjusted based on the recent prediction errors of each model; the smaller the error, the greater the weight.

[0062] For example, if the mean absolute percentage error of LSTM is 5% and that of XGBoost is 6% over the past hour, then the weights of LSTM are slightly higher.

[0063] Monte Carlo dropout is used for LSTM and GRU, and quantile regression is used for XGBoost to output the 10%, 50%, and 90% quantiles at each prediction time. The quantiles are the confidence intervals. For example, the median load prediction at the 5th minute in the future is 100kW, and the 90% confidence interval is [95kW, 105kW].

[0064] Finally, output the IT load sequence and confidence interval for the next 15 minutes.

[0065] 2. Cooling demand forecast The purpose of cooling demand forecasting is to estimate the amount of cooling required from the cooling plant to maintain the normal operating temperature of IT equipment.

[0066] Input the IT load forecast results, current temperature status, ambient temperature, air conditioning terminal status, and thermal characteristic parameters of the container; The IT load forecast result is derived from the IT heat generation sequence for the next 15 minutes from the previous step; The current temperature status includes the current temperature of each area inside the IT cabin, the supply air temperature, and the return air temperature. Ambient temperature refers to outdoor dry-bulb temperature, and relative humidity includes current and future forecast values. The status of the air conditioning terminal includes the current fan speed and valve opening. The thermal characteristics of the shelter are obtained from the digital twin model, including heat capacity, thermal resistance, and cold energy transfer delay time.

[0067] First, a simplified thermodynamic model is constructed. The system uses the lumped parameter method to establish the heat balance equation of the IT cabin, simplifying the entire cabin into several hot zones. The temperature change of each hot zone is determined by the amount of cold air entering the zone and the heat generated by the IT equipment in that zone. The heat exchange with adjacent hot zones and with the outside world determines the heat exchange.

[0068] Then, a model of the cold energy transfer delay is performed. It takes time for chilled water to be transported from the chiller station to the IT container. Therefore, the cold energy provided by the chiller station will not immediately affect the temperature of the IT container. The delay model is established by measuring or estimating the chilled water transport time from the chiller station to the IT container and using a first-order inertial plus pure time delay model to describe the dynamic process of cold energy transfer.

[0069] The calculation for minute t is performed as follows: The input comes from the IT load forecast of the IT heat generation at minute t. The input depends on the supply and delay model of the cooling station in the previous few moments and the actual cooling capacity delivered to the IT cabin at minute t. According to the heat balance equation, the temperature change at minute t+1 is calculated. If it is predicted that the future temperature will exceed the set range, the required increase in cooling capacity is calculated; if it is predicted that it will not exceed the set range, no increase or decrease is made.

[0070] The above heat balance equation is: C∙dT(t) / dt = Q it (t)-Q cool (t)-Q env (t) Where T(t) is the temperature of the thermal zone, C is the heat capacity, and Q is the amount of heat required to raise the temperature of the thermal zone by 1°C. it (t) represents the heat dissipation power of the IT equipment in kW and Q. cool (t) represents the air conditioning cooling power in kW; a positive value indicates cooling capacity, Q. env (t) represents the amount of heat exchanged with the outside in kW, with a positive value indicating that heat is transferred from indoors to outdoors.

[0071] Next, the cooling demand predicted by the physical model is compared with the actual historical cooling demand. A trained neural network correction model is used to add the correction value to the physical model prediction to obtain the final cooling demand prediction.

[0072] Finally, the system outputs the cooling demand sequence for the next 15 minutes and the temperature trajectory prediction (the expected temperature change curve of each area of ​​the IT cabin in the next 15 minutes). If it is predicted that the temperature will exceed the safety threshold at some point in the future, an early warning will be issued.

[0073] 3. System energy efficiency prediction The purpose of system energy efficiency prediction is to estimate the energy consumption of the entire refrigeration system, including chillers, water pumps, and cooling towers, under different operating conditions, so as to provide an energy efficiency assessment basis for subsequent optimization.

[0074] Input device current status, device performance curves, future operating condition predictions, and device combination schemes; The current status of the equipment includes the load rate, condensing temperature, and evaporating temperature of each chiller unit; the speed and flow rate of each water pump; and the speed and outlet water temperature of the cooling tower fan. The equipment performance curves include the performance parameters of each piece of equipment obtained from the digital twin model, namely the coefficient of performance (COP) curves of the chiller unit under different load rates and cooling water temperatures, the efficiency and power curves of the water pump under different flow rates, and the heat dissipation capacity curves of the cooling tower under different fan speeds and ambient temperature and humidity. Future operating conditions forecasts include IT load forecasts, cooling demand forecasts, and outdoor temperature and humidity forecasts. In order to determine which equipment is in operation and which is on standby, the equipment combination scheme can be predicted based on the current status during the forecasting phase.

[0075] First, based on the current operating status and future load forecast, it is estimated which equipment will be in operation at various times in the future. If the future cooling demand exceeds the total capacity of the currently operating units, it is predicted that more units will be turned on; if the demand is below a certain threshold, it is predicted that some units will be turned off. Then, for each operating device, the power is calculated based on its performance curve and predicted operating conditions. The power of the chiller unit is obtained by inputting the proportion of cooling capacity that the unit undertakes from the cooling demand and the cooling water temperature, and obtaining the COP curve to obtain the COP value under this operating condition. The power is calculated using the formula: Power = Cooling Capacity Undertaken / COP.

[0076] The power of the water pump is obtained by inputting the required heat dissipation through the cooling tower, which is equal to the heat on the condenser side of the chiller and the ambient wet-bulb temperature. The cooling tower performance curve is then obtained to determine the required fan speed or power.

[0077] The power of the water pump is calculated based on the input flow rate requirement, cooling demand, supply and return water temperature difference, head requirement, and the water pump performance curve obtained from the pipeline resistance curve. The efficiency and power at that flow rate are then obtained using the formula: Power = Flow Rate × Head × Medium Density / (Pump Efficiency × Motor Efficiency).

[0078] Then, starting from the current moment and moving forward 15 minutes, step 2 is repeated every minute to obtain the total cooling system power at each moment. The fixed power consumption of auxiliary equipment is also taken into account.

[0079] Next, PUE prediction is performed. The total power consumption is obtained by using the formula: Total power consumption = IT load power consumption + Cooling system power consumption + Auxiliary power consumption. Then, the predicted PUE of the system energy efficiency is obtained by using the formula: Predicted PUE = Total power consumption / IT load power consumption (take the average value or point value of the next 15 minutes).

[0080] Finally, the system outputs the power sequence of the cooling system for the next 15 minutes, the PUE sequence for the next 15 minutes, and the energy consumption ratio of each device. If it is predicted that a certain device will operate in an inefficient zone, an energy efficiency bottleneck will be indicated.

[0081] 4. External Environment Prediction The purpose of external environment forecasting is to obtain future changes in outdoor temperature and humidity as input for cooling demand and energy efficiency forecasting.

[0082] Input local weather station real-time data, weather forecast API, and historical weather data; Real-time data from local weather stations includes current outdoor temperature and humidity, wind speed, wind direction, and solar radiation intensity. The weather forecast API obtains hourly forecasts for the next few hours by connecting to the third-party weather service AccuWeather. Historical meteorological data includes local measured data from the past few days.

[0083] First, real-time weather station data and API forecast data are fused using a combination of Kalman filtering and simple weighted averaging.

[0084] Then, for the very short-term forecast of the next 15 minutes, the ARIMA time series method is used to correct the changes based on the changes in the last few minutes. Specifically, the actual temperature and humidity changes in the past 30 minutes are collected first, then a simple autoregressive model is established to predict the changes in the next 15 minutes, and finally the API forecast value is used as the external regression quantity. The two are combined to output the outdoor temperature series and the relative humidity series for the next 15 minutes.

[0085] S4. Perform global optimization and adjustment based on the type of mobile cabin and prediction of its future status; The specific global optimization adjustments are as follows: Using the Model Predictive Control (MPC) approach, the following operations are performed in each control cycle: 1. Obtain the current status of each cabin from S2; 2. Obtain future forecast data from S3; 3. Obtain the container type and coupling relationship from S1; 3. Taking into account the current state and future predictions, solve an optimization problem within a finite time domain, determine the optimal setpoint for each controllable device at each future moment, issue the setpoint for the first moment to each cabin for execution, and repeat the process in the next cycle to achieve rolling optimization.

[0086] 4. Based on the different types of makeshift shelters, the decision variables are defined and all decision variables are denoted as vector u(k), where k represents the k-th time step in the future (k = 0,1,....,N-1), and N is the length of the control suit.

[0087] 5. Construct a multi-objective optimization function and dynamically adjust the weights according to the operating mode, such as energy-saving mode and safety mode: 5.1 Minimize total energy consumption

[0088] Among them, P IT (k) represents the power consumption of the IT equipment;

[0089] P aux (k) represents the power consumption of the auxiliary equipment, which is predicted by S3; The power consumption models for each device are derived from the performance curves in the digital twin model; Chiller unit power: (P) chiller,i (k)=f chiller,i (Q chiller,i (k),T cw,i (k))), where Q chiller,i (k) represents the cooling capacity, T cw,i (k) represents the cooling water temperature; Water pump power: (P) pump,j (k)=f pump,j (m j (k),H j (k))), where m j (k) represents the flow rate, H j (k) represents the head.

[0090] 5.2 Minimize temperature deviation

[0091] Where M represents the number of temperature-controlled zones within the IT shelter; T IT,m (k) represents the predicted temperature of the m-th region at time k, obtained recursively from the heat balance equation; T set,m To set the temperature; Temperature prediction uses the discretized S3 heat balance equation:

[0092] Among them, Q cool,m (k) is obtained from the cold station supply after a delay.

[0093] 5.3 Equipment lifespan and adjustment stability

[0094] Among them, (∆U d (k)=u d (k)-u d (k-1)) represents the rate of change of the equipment control quantity, penalizing large fluctuations.

[0095] Start-stop penalty: For equipment such as chillers, a fixed penalty value (μ) is added each time the unit starts or stops. d To avoid frequent start-stop cycles.

[0096] 5.4 Security objectives under extreme circumstances When a risk is predicted, a safety penalty item is dynamically added. For example, high temperature risk: if the predicted temperature T in a certain area is... IT,m (k) Exceeds the safety threshold T high Then add a penalty (a∙ max(0,T)). IT,m (k)-T high ) 2 ).

[0097] If the power or current of a device is about to exceed its rated value, then add the corresponding constraints.

[0098] The final objective function is a weighted sum: J = w1j1 + w2j2 + w3j3 + wsafetyjsafety The weights are dynamically adjusted based on the operating mode: w1 is increased in normal energy-saving mode; wsafety is increased in safety mode.

[0099] 6. The specific constraints mentioned above are as follows: 6.1 Equipment Capacity Constraints Chiller cooling capacity range: Q min,i ≤Q chiller,i (k)≤Q max,i Pump speed range: n min,j ≤p ump,j (k)≤n max,j Battery charge / discharge power range: P charge,max ≤P battery (k)≤P discharge,max 6.2 Temperature Constraint IT equipment intake air temperature range: T min ≤T IT,m (k)≤T max The upper limit of the temperature corresponding to the condensing pressure of the chiller, etc.

[0100] 6.3 Power Constraints Total power consumption shall not exceed the transformer capacity:

[0101] The power of the equipment shall not exceed the rated value.

[0102] 6.4 Coupling Constraints Cooling balance: The total cooling capacity provided by all cooling stations equals the sum of the cooling capacity required by each IT compartment.

[0103] Where d m The number of delay steps from the cooling station to the IT compartment m, ( () represents the conveying efficiency.

[0104] Hydraulic balance: The flow rates of each branch satisfy the node flow balance equation.

[0105] Power balance: Total power supply equals total power consumption: P grid (k)+P battery (k)=P IT (k)+P cooling (k)+P aux (k) Where Pgrid(k) represents the power purchased from the mains, with no upper limit but must be non-negative.

[0106] 7. Solution Strategy Due to the large number of devices, the long time-domain scale, and the inclusion of mixed integers and continuous variables, a hierarchical solution strategy is adopted to adapt to different situations: 7.1 Upper Layer: Equipment Combination Optimization For discrete decisions such as chiller start-up and shutdown, filtering is based on preset rules. An optimal start-up and shutdown combination table is pre-calculated offline, and the table is then looked up online to match the current load.

[0107] 7.2 Lower Level: Optimization of Continuous Variables Given start-stop combinations, continuous variable optimization can be simplified to a convex optimization problem or nonlinear programming, which can be solved using sequential quadratic programming (SQP) or interior-point methods. For large-scale systems, the alternating direction multiplier method (ADMM) distributed optimization is used to decompose the problem into independent solutions for each module, and then coordinate the coupling constraints.

[0108] 8. Extreme Situation Handling Mechanism 8.1 When the temperature is high or the cooling capacity decreases, increase the weight of safety targets, tighten the upper limit of temperature constraints, adopt a pre-cooling strategy, increase the cooling capacity reserve in advance, and link with the IT scheduling system to limit non-critical loads.

[0109] 8.2 If a device failure occurs in the optimization model, the faulty device will be removed, the constraints will be updated, the backup device will be enabled, the load will be redistributed, and the temperature requirements for some areas may be relaxed to focus on protecting critical areas.

[0110] 8.3 When a sudden load peak occurs, the cooling output should be increased in advance by taking advantage of the predictability. Within the allowable range, the equipment can be temporarily overloaded, or instantaneous energy can be released by energy storage or cold storage.

[0111] 8.4 When communication is interrupted, the edge node switches to local mode and runs based on the local model and default policy; after communication is restored, the system resynchronizes and updates the model.

[0112] S5. Issue and execute the global optimization instruction, and repeat S2 to S5.

[0113] S5 is used to send the optimal control commands calculated by S4 to the edge nodes of each compartment, drive the actual equipment to move, and ensure the reliability of command execution and closed-loop feedback. At the same time, it provides a corresponding handling mechanism for situations where the constraints may not be met during the S4 optimization process.

[0114] The instructions are encapsulated in JSON format for easy network transmission and parsing.

[0115] The S4 optimization calculation outputs a future-oriented control sequence, but in each control cycle, the system only issues instructions for the current moment. Each instruction includes the following: Command ID: A globally unique command identifier used for tracking and confirmation.

[0116] Container Identifier: The unique ID of the target container, consistent with the identifier assigned in S1.

[0117] Device Identifier: The unique ID of the target device in the digital twin model.

[0118] Command types include setting command, start / stop command, and mode switching command.

[0119] Timestamp: The time when the instruction was generated, used for timing verification.

[0120] Validity period: The validity period of the instruction, which is the current control cycle.

[0121] The instructions are shown in Table 1. Table 1 Instruction Content

[0122] After the S4 optimization phase is completed, the system distributes the current commands to the edge nodes of each module based on the module grouping. The system is responsible for publishing command information, and each edge node subscribes to the dedicated command topic for its assigned module. For commands with high reliability requirements, a TCP direct transmission mechanism is used for transmission, and the commands are sent in parallel via primary and backup dual communication links to ensure reliable delivery.

[0123] After receiving a command, the edge node sends an acknowledgment to the system. First, the node immediately replies "Received" upon receiving the command. Second, after parsing the command and verifying its format, the node replies "Parsing Successful." Finally, after driving the device to perform the action, the node replies "Executed" along with the actual execution result. The system has a preset timeout threshold; if no acknowledgment is received, a retransmission mechanism is triggered, with a maximum of three retries. If the attempt still fails, a command delivery failure alarm is recorded.

[0124] After receiving the instruction, the edge node converts the instruction into a specific signal that the device can recognize.

[0125] Edge nodes feed back the actual execution results (such as the actual frequency and valve opening) to the system in real time. The system compares the deviation between the actual value and the command value. If the deviation exceeds the deviation threshold, an execution deviation alarm is recorded, which may trigger optimization and adjustment in the next cycle.

[0126] When the optimization problem has no feasible solution and no control scheme can satisfy all constraints, the system adopts degradation or emergency strategies to ensure the safe operation of core functions, as follows: 1. Constraint relaxation strategy When S4 reports no feasible solution, the system gradually relaxes the constraints and re-solves the problem in the following priority order. 1.1 Relaxing temperature constraints The upper limit of the IT intake air temperature is gradually increased from the normal value to the safe upper limit, increasing by 0.5℃ each time, until a feasible solution is found.

[0127] Simultaneously, temperature constraint relaxation events are recorded to alert maintenance personnel.

[0128] If there is still no solution even after relaxing the limits to the safety limit, then proceed to 1.2.

[0129] 1.2 Relax equipment capacity constraints Short-term overload operation of equipment such as chillers and pumps is permitted, with the overload not exceeding 10% of the rated value and the duration not exceeding 15 minutes.

[0130] Add an overload penalty term to the objective function, but prioritize finding a feasible solution.

[0131] If there is still no solution, proceed to 1.3.

[0132] 1.3 Relaxing power constraints The total power is allowed to exceed the transformer capacity for a short period of time, but not more than 10%, and the duration shall not exceed 5 minutes.

[0133] Record power over-limit alarms to indicate the risk of triggering upstream switch protection.

[0134] 1.4 Relaxing the constraints on cooling capacity balance Allow some non-critical IT areas to temporarily not meet their cooling needs, prioritizing core areas.

[0135] Based on the cabin category information in S1, the system identifies which IT cabinets are critical and which are non-critical, and provides differentiated protection.

[0136] 2. In the event of a sudden and serious malfunction, the system will automatically switch to emergency operation mode. 2.1 Core Protection Mode Shut down non-critical equipment, such as some cooling tower fans and unnecessary water pumps.

[0137] Limit non-critical IT load and coordinate with the IT scheduling system.

[0138] We will do our utmost to ensure cooling and power supply to the core IT area.

[0139] 2.2 Pure Local Autonomy Model If communication is normal but optimization fails, the platform sends a command to each edge node to enter local autonomous mode.

[0140] Each edge node switches to its local preset emergency control strategy, such as simple adjustment based on local temperature, no longer relying on platform collaborative optimization.

[0141] The platform will continue to monitor the situation and attempt to switch back to collaborative mode once conditions are restored.

[0142] 2.3 Shutdown Protection Mode When the temperature approaches the hardware failure threshold or the power is severely exceeded, the system automatically triggers an emergency shutdown protection.

[0143] Shut down unnecessary devices safely in a preset sequence, and notify maintenance personnel to intervene manually if necessary.

[0144] 3. In case of extreme situations, the system is pre-configured with a fast extreme response mode as shown in Table 2. Table 2 Fast Extreme Response Mode

[0145] Once the factors causing the constraint failure disappear, the system automatically attempts to resume normal optimization. The platform continuously monitors actual operating data. If it finds that the normal constraints have been met, it restarts S4 optimization; using the current actual state as the initial condition, it re-solves for feasible solutions and smoothly switches back to collaborative optimization mode.

[0146] Obviously, the above embodiments are merely examples to clearly illustrate the embodiments of the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively list all embodiments here. However, these obvious variations or modifications derived from the spirit of the present invention are still within the protection scope of the present invention.

Claims

1. A method for coordinated communication of modular data center shelters, comprising: Includes the following steps: S1. Receive uploaded 3D drawings, identify the type of modular shelter, and build a digital twin model; S2. Collect data from each mobile shelter in real time, and update the status of the digital twin model in real time based on the mobile shelter data; S3. Predict the future state of the makeshift hospital based on the real-time updated digital twin model; S4. Perform global optimization and adjustment based on the container category and the future state prediction; S5. Issue and execute the global optimization instruction, and repeat S2 to S5.

2. The modular data center container collaborative communication method according to claim 1, characterized in that, The categories of mobile shelters identified in S1 include: Based on the drawing analysis results, three types of features are extracted: component features, spatial layout features, and textual semantic features. The three types of features are concatenated into a comprehensive feature vector, which is then input into a multimodal classification model composed of a CNN+RNN fusion network. The multimodal classification model outputs the probability of the container class based on a predefined class library; The predefined category library includes integrated cold station cabins, hydraulic module cabins, power distribution cabins, IT computing power cabins, and prefabricated component cabins.

3. The modular data center container collaborative communication method according to claim 2, characterized in that, Building a digital twin model in S1 includes: Retrieve the parametric template model corresponding to the aforementioned container category. The parametric template model includes a geometric model skeleton, physical model equations, and a data-driven model structure. Instantiate the template model according to the equipment model or size in the drawing; Based on the pipeline topology in the drawings, connect the various equipment models to form a complete modular digital twin model; The digital twin model is subjected to a consistency check, and if it passes the check, it is assigned a unique identifier and stored in the model library.

4. The modular data center container collaborative communication method according to claim 3, characterized in that, The real-time data collection for each mobile shelter in S2 includes: An edge computing node is deployed in each modular shelter, and the edge computing node is preloaded with a lightweight model summary of the target modular shelter; The edge computing node collects real-time data at a preset period, and the lightweight model summary includes sensor locations, measurement range, and sampling frequency. The collected data from the makeshift hospitals were preprocessed, including missing value imputation, noise filtering, outlier removal, unit conversion, and data compression. The preprocessed data is packaged according to a unified timestamp format and uploaded to the central digital twin system in real time via a communication network.

5. The modular data center container collaborative communication method according to claim 4, characterized in that, The updating of the digital twin model state in S2 includes: Based on the unique identifier of the mobile cabin in the uploaded data packet, load the corresponding digital twin model instance from the model library; Associate and map real-time data with variables in the model; Key states that cannot be directly measured are estimated using state estimation algorithms; Update the mapping data and state estimation results to the state variable set of the digital twin model; The updated model undergoes a consistency check, and if the error exceeds a preset threshold, an online model correction mechanism is triggered.

6. The modular data center container collaborative communication method according to claim 5, characterized in that, The prediction of the future state of the enemy cabin in S3 includes: A model predictive control framework is adopted, which uses the currently updated digital twin model as a basis to make rolling predictions of the state evolution in the future prediction time domain; The forecasts include at least one of IT load forecasts, cooling demand forecasts, system energy efficiency forecasts, and external environment forecasts. The prediction employs a multi-model ensemble strategy, outputting predicted values ​​and confidence intervals.

7. The modular data center container collaborative communication method according to claim 6, characterized in that, The global optimization adjustment in S4 includes: Construct a multi-objective optimization function, which includes an energy consumption minimization objective, a temperature deviation minimization objective, an equipment regulation stability objective, and a safety penalty objective; Define constraints, including equipment capacity constraints, temperature constraints, power constraints, and coupling constraints; A hierarchical solution strategy is adopted, including upper-level equipment combination optimization and lower-level continuous variable optimization; Set up an extreme case handling mechanism to implement a constraint relaxation strategy or switch to emergency operation mode when there is no feasible solution to the optimization problem.

8. The modular data center container collaborative communication method according to claim 7, characterized in that, The global optimization instructions issued in S5 include: The optimal control commands obtained from the optimization calculation are encapsulated in JSON format and distributed to the edge nodes of each module according to the modular unit grouping. After receiving the instruction, the edge node returns confirmation information, including confirmation that the instruction has been received, confirmation that the parsing was successful, and confirmation that the instruction has been executed. The system has a preset timeout threshold; if no confirmation is received, a retransmission mechanism is triggered. The edge nodes feed back the actual execution results to the system in real time. The system compares the deviation between the actual value and the instruction value. If the deviation exceeds the deviation threshold, an execution deviation alarm is recorded. When the optimization problem has no feasible solution, the system executes a constraint relaxation strategy or an emergency operation mode.

9. A modular data center container collaborative communication system using the collaborative communication method as described in claim 1, characterized in that, This includes a central digital twin system, edge computing nodes, and communication networks; The central digital twin system is used to perform container category identification, digital twin model construction and updating, future state prediction and global optimization adjustment, and to issue optimization instructions; The edge computing nodes are deployed in each shelter to collect shelter data in real time, perform data preprocessing, receive and execute optimization instructions, and switch to local autonomous mode when communication is interrupted. The communication network is used for bidirectional data interaction between the central digital twin system and the edge computing nodes.

10. The modular data center container collaborative communication system according to claim 9, characterized in that, The edge computing node includes a multi-protocol access module, an analog / digital input module, a network communication module, a local storage module, an edge computing module, and an instruction parsing and execution module; The multi-protocol access module is used to support the connection of sensors and device controllers using multiple industrial protocols; The analog / digital input module is used to connect to analog sensors and switch signals; The network communication module is used for bidirectional data interaction with the central system; The local storage module is used for data caching; The edge computing module is used to run data preprocessing algorithms and local emergency control logic; The instruction parsing and execution module is used to convert instructions into signals that the device can recognize and drive the device to act, as well as to provide feedback on the actual execution results.

Images