Dual-pipeline liquid-cooled energy storage cabinet and load prediction-based regulation method thereof

Abstract

The application relates to the field of energy storage, in particular to a double-pipeline liquid cooling energy storage cabinet and a regulation and control method thereof based on load prediction. At least comprising: two groups of mutually independent liquid supply pipelines, respectively pipeline A and pipeline B; each group of pipelines comprises a variable frequency circulating pump, an electric flow regulating valve, a liquid supply temperature sensor, a pipeline pressure sensor, a liquid supply branch pipe and a liquid return branch pipe; the liquid supply branch pipe of pipeline A and the liquid supply branch pipe of pipeline B are connected into a liquid cooling plate water inlet arranged below a battery module through a water distributor, and the water outlet of the liquid cooling plate is connected with the liquid return branch pipe of pipeline A and the liquid return branch pipe of pipeline B through a water collector. The application can effectively compensate the response lag caused by thermal inertia by increasing the cooling capacity in advance through load prediction, can reduce the temperature peak of the battery cell during charging and discharging by 2-3 DEG C, and prolongs the cycle life.

Description

Technical Field

[0001] This application relates to the field of energy storage, and in particular to a dual-pipeline liquid-cooled energy storage cabinet and its control method based on load forecasting. Background Technology

[0002] Commercial and industrial energy storage units play multiple roles in the power system, including peak-valley arbitrage, demand-side response, and frequency regulation assistance. Their charging and discharging power exhibits dramatic planned fluctuations over time. Taking a typical 1C rate energy storage system as an example, during peak electricity price periods, continuous full-power discharge can generate tens of kilowatts of instantaneous heat from the battery cells, with a temperature rise rate of 3°C to 5°C per minute. The battery modules themselves have significant thermal inertia, and the PID control of the liquid cooling system, based on real-time temperature feedback, typically requires a lag of 1 to 3 minutes from issuing the adjustment command to achieving an effective cooling effect. This lag leads to significant overshoot in the cell temperature when the discharge power surges, with measured overshoots of 5°C to 8°C. This not only accelerates battery capacity degradation but also increases the risk of thermal runaway during long-term operation.

[0003] To address this issue, existing technical solutions propose using day-ahead load forecasts or electricity price curves to adjust the chiller outlet water temperature in advance, allowing the cooling system to enter a high-heat dissipation state before peak load, thus partially compensating for thermal inertia. For example, cooling demand for each period can be calculated based on predicted charge and discharge power curves, and the chiller outlet water temperature can be adjusted in advance. However, these solutions are all based on a single cooling circulation pipeline, and the adjustment methods only include pump speed and chiller outlet water temperature, resulting in low control flexibility. When load forecasting is inaccurate, the cooling capacity adjustment margin of a single pipeline is extremely small, often forcing a difficult choice between conservative high energy consumption and aggressive overheating risks, failing to balance economy and safety.

[0004] In terms of system reliability, high-reliability scenarios place higher demands on the fault tolerance of liquid cooling systems. In a single-pipe liquid cooling system, if the pump seizes up, the pipe becomes blocked, or the proportional valve fails, the entire thermal management loop will completely lose its cooling capacity. Therefore, a number of products and solutions employing a dual-pipe liquid supply architecture have emerged in the industry. Existing literature discloses a dual-pump redundant liquid cooling system that automatically switches to the backup pipe when the main pump experiences insufficient flow or abnormal pressure difference. However, this control strategy is essentially a passive switch after a fault; during normal operation, the backup pipe remains idle for extended periods, and the switching process can easily cause pressure and flow surges. Another solution involves distributing the total flow at a fixed ratio when the two pumps operate in parallel. However, this fixed-ratio distribution method does not dynamically optimize according to real-time cooling demands, resulting in high overall power consumption for the pump unit and making it unable to adapt to the drastic fluctuations in the thermal load of the energy storage cabinet.

[0005] Furthermore, current mainstream energy storage management systems are capable of predicting or planning charging and discharging power for the next few hours, and this planning information can be obtained in advance. However, existing liquid-cooled thermal management systems do not deeply integrate this predictive information with the redundancy and coordination capabilities of dual-pipeline systems, especially failing to utilize the uncertainty output by the predictive model itself as a basis for pipeline mode switching and redundancy deployment decisions. In actual operation, any predictive model has a certain degree of error. If energy-saving optimization is performed when the predictive reliability is high, and the redundancy level is proactively increased to withstand unexpected temperature rises when the reliability is low, the overall efficiency and robustness of the system will undoubtedly be significantly improved.

[0006] In summary, the main shortcomings of existing technologies can be summarized as follows: First, the cooling control suffers from thermal inertia lag, and the ability to suppress instantaneous temperature peaks in the battery cells is insufficient. The overshoot of PID control based on real-time temperature can reach more than 5°C. Second, single-pipe feedforward regulation control has low degrees of freedom and poor tolerance for prediction errors. Third, the dual-pipeline system has a single control strategy, which only achieves passive switching or fixed ratio allocation after a fault, and lacks the means to actively coordinate modes and dynamically optimize flow based on cooling demand and prediction confidence. Fourth, the energy storage scheduling plan is disconnected from the thermal management system, and the forecast uncertainty is not used for risk decision-making regarding pipeline redundancy. Summary of the Invention

[0007] The technical problem to be solved by this application is: in a liquid-cooled energy storage cabinet with dual-pipeline liquid supply capability, how to predict cooling demand in advance based on available load forecast information and its uncertainty, dynamically select the dual-pipeline working mode and achieve smooth transition between modes, while optimizing flow distribution, and effectively suppressing cell temperature peaks and reducing total system energy consumption while ensuring redundancy safety.

[0008] To achieve the above objectives, a dual-pipeline liquid-cooled energy storage cabinet and its control method based on load forecasting are provided.

[0009] In one aspect, this application discloses a dual-pipeline liquid-cooled energy storage cabinet, comprising two independent sets of liquid supply pipelines, namely pipeline A and pipeline B. Each set of pipelines includes a variable frequency circulating pump, an electric flow regulating valve, a liquid supply temperature sensor, a pipeline pressure sensor, a liquid supply branch pipe, and a liquid return branch pipe. The liquid supply branch pipes of pipeline A and pipeline B converge via a water distributor to the inlet of a liquid-cooled plate located below the battery module. The water outlet of the liquid-cooled plate is connected via a water collector to the liquid return branch pipes of pipeline A and pipeline B, respectively. After the liquid return branch pipes of pipeline A and pipeline B converge, the water enters the evaporator side of the chiller unit for recooling. The water outlet of the chiller unit is split to the inlet of the circulating pump of pipeline A and the inlet of the circulating pump of pipeline B. The dual-pipeline liquid-cooled energy storage cabinet also includes a centralized controller, which receives power scheduling information from the energy storage converter, battery state of charge (SOC) and state of health (SOH) provided by the battery management system, and data from ambient temperature and humidity sensors.

[0010] Secondly, this application discloses a load prediction and control method based on a dual-pipeline liquid-cooled energy storage cabinet.

[0011] The control method is executed by a centralized controller and includes the following steps: Step 1: Load Forecasting and Heat Production Calculation The centralized controller acquires the planned charge and discharge power sequence P(t) within the future time window T received by the energy storage converter in real time, t∈[t0, t0+T]. The value of T is generally in the range of 2 to 4 hours, preferably 2 hours, and the sampling period is 1 minute. At the same time, it collects the current battery SOC, SOH, temperature of each measuring point of the battery module, ambient dry bulb temperature and relative humidity.

[0012] The aforementioned data, along with timestamp features including time period identifiers, weekday types, and season labels, are input into a pre-built and trained multivariate time series prediction model. This model employs a Transformer time series model incorporating convolutional self-attention mechanisms. Its encoder extracts features from historical power, SOC, and ambient temperature sequences, while the decoder takes future power plans as input and outputs the weighted average heat generation power Q(t) of the battery modules at each discrete time t within a future time period T, along with its prediction uncertainty σ(t).

[0013] The training labels for the prediction model are obtained as follows: using battery terminal voltage V, current I, and the SOC change rate dSOC / dt as observations, an extended Kalman filter is constructed to inversely calculate the generated heat power. The state vector x includes SOC, two polarization voltages V1 and V2, and ohmic internal resistance R0; the state equation is discretized based on the ampere-hour integral method and the polarization characteristic time constant; in the observation equation, the terminal voltage V_obs is obtained by subtracting the polarization voltage and ohmic voltage drop from the open-circuit voltage OCV, and the generated heat power Q is composed of ohmic heat and polarization heat, i.e., Q = I²·R0 + I·(V1+V2). The process noise covariance matrix and the observation noise covariance matrix are set based on offline battery test data and operational data statistics, and are updated online using an adaptive covariance matching algorithm to obtain the true generated heat power label value at each moment.

[0014] Then, the predicted heat production power sequence {Q(t)} and the corresponding uncertainty sequence {σ(t)} are output. σ(t) is the root mean square of the prediction variance obtained by the prediction model through multiple samplings using Monte Carlo Dropout, representing the confidence level of the model's prediction for that point.

[0015] Step 2: Dynamic Calculation of Cooling Demand A lumped-parameter thermal equivalent circuit model of the battery module was constructed. The model parameters included the heat source generated by the battery's internal resistance, the battery's thermal capacity C_b, the conductive thermal resistance R_cond from the battery to the liquid cooling plate surface, and the convective thermal resistance R_conv from the liquid cooling plate surface to the coolant. The initial values ​​of these model parameters were obtained through offline identification using multi-condition charge-discharge temperature rise experiments. To reflect the impact of battery aging on thermal characteristics, correction functions for SOH and thermal resistance were established: R_cond(SOH)=R_cond_0·[1+β·(1-SOH)], R_conv(SOH)=R_conv_0·[1+γ·(1-SOH)]. Where β and γ are empirical coefficients obtained by fitting thermal characteristic test data of the same battery model at different cycle aging stages; SOH is taken as the current capacity retention rate reported by the battery management system; C_b is considered not to change significantly with SOH.

[0016] Input the heat generation prediction sequence Q(t) into the thermal equivalent circuit model, and use the maximum allowable temperature of the battery cell T_max as the upper limit of the constraint. The default value of T_max is 38℃. Solve for the minimum liquid cooling power P_cool(t) required at each time t.

[0017] The P_cool(t) is converted into the total required flow rate F_req(t) and optimal supply temperature T_supply of the liquid cooling system. The conversion is based on: F_req(t) = P_cool(t) / (ρ·c_p·ΔT_design), where ρ is the coolant density, c_p is the coolant specific heat capacity, and ΔT_design is the designed standard temperature difference between the supply and return liquids, taken as 5℃. When P_cool(t) is in kW and c_p is in kJ / (kg·K), the calculated F_req(t) is in m³ / s, which is multiplied by 60000 in practical applications to convert to L / min. T_supply is selected based on the energy efficiency characteristic curve of the chiller unit, aiming for the lowest possible energy consumption of the chiller unit, while ensuring that the cooling capacity can cover P_cool(t). Finally, the total required flow rate sequence F_req(t) and the recommended supply temperature sequence T_supply(t) for the next T time period are output.

[0018] Step 3: Prediction and Switching of Dual-Pipeline Working Mode First, we define the method for determining the prediction uncertainty threshold σ_th: σ_th is the 90th percentile of the prediction error for the same period in the past 30 days multiplied by a coefficient of 1.2, and is updated daily.

[0019] Based on the total demand flow F_req(t) and the prediction uncertainty σ(t), the centralized controller determines the operating mode to be used at the current moment from the following three pre-set operating modes: Mode 1 is a single-pipe economic operation mode. Only pipeline A is in operation, while the circulation pump in pipeline B is stopped and its flow control valve is closed. The applicable conditions for this mode are: the total demand flow rate F_req(t) is not greater than 80% of the maximum allowable continuous flow rate F_max of a single pipeline, i.e., F_req(t) ≤ 0.8·F_max; the prediction uncertainty σ(t) is less than the threshold σ_th; and all prediction times within a preset forward look-ahead period ΔT_horizon (15 minutes) meet the above conditions. To avoid frequent mode switching, when switching from other modes to Mode 1, the above applicable conditions must be met continuously for at least 30 seconds; when switching from Mode 1 to other modes, the corresponding exit conditions must also be met continuously for at least 30 seconds. When it is predicted that the applicable conditions will no longer be met within the next ΔT_horizon, the mode is exited and transitions to another mode.

[0020] Mode 2 is a dual-path balanced flow sharing mode. Two pipelines operate simultaneously, sharing the total demand flow according to a single flow distribution coefficient α(t). Pipeline A provides F_A = α(t)·F_req(t), and pipeline B provides F_B = (1-α(t))·F_req(t), where 0 < α(t) < 1. This mode is applicable when the total demand flow satisfies 0.8·F_max < F_req(t) ≤ F_max, and the prediction uncertainty σ(t) is less than σ_th. Similarly, switching to this mode requires these conditions to be met continuously for 30 seconds. If the flow rate is within this range but σ(t) ≥ σ_th, for safety, the system switches to the active redundancy sub-mode in Mode 3.

[0021] Mode 3 is the primary / standby switching redundancy mode, which includes two sub-modes. **Limited Supplement Sub-mode:** Activated when F_req(t) > F_max. In this mode, pipeline A operates at full capacity (F_A = F_max), and pipeline B supplements the remaining flow with F_B = F_req(t) - F_max. The flow allocation coefficient is fixed at F_max / F_req(t). **Active Redundancy Sub-mode:** Activated when the total demand satisfies 0.8·F_max < F_req(t) ≤ F_max, but σ(t) ≥ σ_th, or when the system detects a pre-alarm signal in pipeline A. The pre-alarm signal includes a warning that the pump bearing vibration amplitude exceeds a preset vibration threshold. In this mode, the two pipelines are still allowed to share the total flow evenly, with the allocation coefficient α(t) calculated by subsequent optimization. However, a penalty term for the pump speed change rate is added to the optimization objective function to suppress frequent adjustments that may result from unreliable predictions. The weight of the penalty term is adaptively adjusted according to the magnitude of σ(t): when σ(t) is close to σ_th, the penalty weight is small, and when σ(t) is close to twice σ_th, the penalty weight increases sharply, thereby making the traffic distribution more uniform and using a conservative strategy to resist uncertain risks.

[0022] When switching from Mode 1 or Mode 2 to Mode 3, pipeline B must be started and the flow rate pre-established with a lead time Δt_pre. The lead time Δt_pre is determined by the following formula: Δt_pre = k·τ + ΔF_target / (dF_req / dt)_max Where τ is the thermal time constant of the battery module, ΔF_target is the required increase in flow rate, (dF_req / dt)_max is the maximum ramp-up rate of the required flow rate, and k is the safety factor, which is set to 1.2. The unit of Δt_pre is determined by the units of τ and ΔF_target / (dF_req / dt)_max. When τ is in seconds, ΔF_target is in L / min, and (dF_req / dt)_max is in L / (min·s), both are in seconds and can be directly added together. The switching between modes employs a smooth transition strategy: when it is determined that a transition from the current mode to the target mode is necessary, the controller enters a preset transition period T_trans, with T_trans set to 40 seconds. During the transition, a ramp function from the current opening to the target opening is applied to the flow regulating valve of pipeline B, and the speeds of the two pumps are adjusted synchronously according to a preset coordination curve; simultaneously, the rolling optimization process of the model predictive control is paused, and optimization resumes after the transition is complete. During the transition, if the temperature is detected to be close to the upper limit, the system directly switches to emergency mode.

[0023] Step 4: Joint optimization of flow distribution and liquid supply temperature A system control model is constructed. State variables include the battery average temperature T_batt, the temperature difference between the inlet and outlet of the cold plate ΔT_cp, the flow rates F_A and F_B of pipes A and B respectively, and the liquid supply temperature T_supply. Control variables include the pump speed n_A of pipe A, the pump speed n_B of pipe B, the proportional valve opening θ, and the chiller set temperature T_chiller_set. Disturbance inputs are the heat output Q(t) and the ambient temperature T_amb. Considering the hydraulic coupling of the pipe network formed by the distributor and collector, the coupling gain matrix from the two pump speeds to their respective flow rates is identified offline. Branch pressure balance conditions are introduced into the model predictive control constraints to ensure that the optimized flow distribution is physically feasible.

[0024] Model predictive control rolling optimization is performed within each control cycle Δt_ctrl, where Δt_ctrl is 30 seconds. Constraints are: battery temperature T_batt ≤ T_max; flow rates in each pipeline satisfy 0 ≤ F_A, F_B ≤ F_max; pump speed is between the minimum allowable speed and the rated speed; and the liquid supply temperature is within the adjustable range of the chiller unit. The optimization objective is to minimize the total system energy consumption over the next H control cycles, expressed as: min J = Σ_{k=1}^{H} [W_pumpA(k) + W_pumpB(k) + W_chiller(k)]·Δt_ctrl. Where W_pumpA and W_pumpB are the electrical power of pumps in pipelines A and B, respectively, calculated in real-time from the pump speed-flow-head characteristic curve and efficiency curve; W_chiller is the energy consumption of the chiller unit compressor, obtained based on the mapping relationship between cooling capacity, liquid supply temperature, and ambient temperature. H is the number of control cycles included in the prediction time domain, H=10. The nonlinear constraint optimization problem is solved online by the controller's built-in sequential quadratic programming algorithm to obtain the control sequence for the next H steps, and only the control command for the first step is issued and executed.

[0025] In the extreme redundancy sub-mode, F_A is fixed at F_max, and F_B is determined by F_req(t) - F_max. In this case, the optimization problem only addresses the pump speed and chiller temperature in pipe B. In other modes, the flow rates of both pipes are optimization variables, and in the active redundancy sub-mode, the objective function includes a penalty term for the rate of change of pump speed with adaptive weights.

[0026] The rolling optimization process described above is repeated in the next control cycle to achieve closed-loop correction.

[0027] Step 5: Safety Monitoring and Emergency Rollback The centralized controller monitors safety parameters at a high frequency with a monitoring cycle of 1 second. The monitored objects include: the maximum temperature T_batt_max at each measuring point of the battery module, the supply and return liquid temperature difference between pipeline A and pipeline B, the inlet and outlet pressure difference of the two circulating pumps, and the chiller unit fault signal.

[0028] The model predictive control mode will be immediately interrupted and switched to the preset real-time feedback emergency control when any of the following conditions occur: T_batt_max > T_max + δ1, where δ1 is 1℃; the supply and return liquid temperature difference of any pipeline exceeds 1.5 times the design standard temperature difference; the differential pressure of any circulating pump drops instantaneously by more than 30% of its rated value and lasts for more than 5 seconds; or a fault shutdown command is received from the battery management system or energy storage converter.

[0029] The emergency control strategy is as follows: both circulating pumps immediately switch to their rated maximum speed, the proportional valve is fully opened, and the chiller unit's set temperature drops to the preset minimum safety value T_chiller_min. Simultaneously, to avoid excessive internal temperature differences or condensation caused by sudden cooling, a soft limit is implemented on the cooling rate: when the average battery temperature drop rate exceeds 2℃ / min, the chiller unit's outlet water temperature setting is gradually adjusted to control the actual cooling rate within the range of 1.5 to 2℃ / min. Once the battery's maximum temperature drops below T_max - δ2 (where δ2 is 1℃), the supply and return liquid temperature difference returns to normal, and all alarm signals are cleared, the emergency mode is exited either manually or automatically after a delay. The automatic delay exit condition is: safety parameters such as temperature and pressure difference remain within the normal range for 5 consecutive minutes, after which the system reverts to the load forecast-based control mode.

[0030] The system records the time, type, and snapshots of the state before and after each emergency trigger event, which are used for subsequent model confidence assessment and adaptive adjustment of control parameters.

[0031] Compared with the prior art, this application has at least one of the following beneficial effects: Proactive cooling and temperature peak suppression. This application increases the cooling capacity in advance by predicting the load, effectively compensating for the response lag caused by thermal inertia, which can reduce the temperature peak of the battery cell during charging and discharging by 2 to 3°C and extend the cycle life.

[0032] Intelligent dual-pipeline coordination and energy efficiency optimization. This application enables economical operation of a single pipeline under low load, balanced distribution of power between the two pipelines under medium load to reduce the total power consumption of the pump set, and pre-activation of redundant pipelines under high load or unreliable forecasts. These three modes are linked with forecast information and uncertainties, taking into account energy efficiency, safety, and system reliability.

[0033] Deep coupling of energy storage and dispatch. This application fully utilizes the known power plans and prediction uncertainties of the energy storage system to synchronize the hierarchical decision-making of the thermal management system with actual operating conditions, resulting in more accurate response and stronger robustness.

[0034] Smooth switching and safety assurance. This application avoids pressure shocks and flow fluctuations caused by conventional dual-pipeline switching by predictive advance action, hysteresis judgment, and transition control; it is also equipped with an emergency backoff mechanism that limits the cooling rate, forming a safety closed loop from optimized operation to ultimate protection.

[0035] The beneficial effects listed above are not exhaustive of all advantages. Other potential beneficial effects and detailed technical implementation methods will be further disclosed in the embodiments or other descriptive sections of this application. Attached Figure Description

[0036] A better understanding of various aspects of this disclosure will be achieved by reading the following detailed description in conjunction with the accompanying drawings. The positions, dimensions, and extents of the structures shown in the drawings, etc., do not always represent actual positions, dimensions, and extents. In the drawings: Figure 1 This is a schematic diagram of the system structure of the dual-pipeline liquid-cooled energy storage cabinet in this application.

[0037] Figure 2 This is a flowchart illustrating the overall process of the control method described in this application.

[0038] Figure 3 This is a schematic diagram illustrating the conversion between load forecasting and cooling demand.

[0039] Figure 4 A logic block diagram for selecting the dual-circuit operating mode.

[0040] Figure 5 This is a schematic diagram illustrating the rolling optimization principle of model predictive control. Detailed Implementation

[0041] The technical solution of this application will be described in detail below with reference to the accompanying drawings. It should be stated that the battery modules, chiller units, variable frequency circulating pumps, electric proportional regulating valves, various temperature and pressure sensors, and centralized controllers involved in the following embodiments are all mature products in this field. Their basic structures, working principles, and conventional electrical connections and communication methods are all prior art. Here, only their specific connection relationships, parameter settings, and control coordination methods in this application will be fully described. Example

[0042] This embodiment uses a 500kW / 1MWh commercial and industrial lithium iron phosphate liquid-cooled energy storage cabinet as the implementation object. The energy storage cabinet adopts an outdoor cabinet structure and contains 8 battery modules. Each module consists of 32 280Ah lithium iron phosphate cells connected in series, with a rated voltage of 102.4V and a rated energy of approximately 28.7kWh. Each module has an aluminum alloy liquid cooling plate installed at its bottom. The liquid cooling plate has a serpentine flow channel with a rectangular cross-section of 8mm × 3mm and a total length of approximately 4.2 meters. Thermally conductive silicone grease is filled between the liquid cooling plate and the cells to reduce contact thermal resistance. The cooling medium is a 50% volume concentration ethylene glycol aqueous solution with a freezing point of approximately -35℃, a density ρ of approximately 1050 kg / m³ at 20℃, and a specific heat capacity c_p of approximately 3.5 kJ / (kg·K).

[0043] Regarding hardware connection and parameter configuration: Pipeline A and Pipeline B are two completely symmetrical independent loops. The pipeline material is 304 stainless steel, with a main pipe inner diameter of 25mm and a branch pipe inner diameter of 15mm. Each pipeline is connected in series with a variable frequency centrifugal circulating pump. The pump has a rated power of 1.5kW, a rated speed of 2900rpm, a rated flow rate of 40L / min, a rated head of 20m, and a speed adjustment range of 20% to 100% of the rated speed, corresponding to a flow rate range of approximately 8L / min to 40L / min. Based on pipeline hydraulic calculations and with a safety margin, the maximum allowable continuous flow rate F_max of a single pipe is determined to be 30L / min. Each pump outlet is equipped with an electric proportional regulating valve, a V-type ball valve, whose actuator accepts a 4-20mA signal to control the opening. The full stroke time is approximately 15 seconds. A PT100 platinum resistance temperature sensor (accuracy ±0.1℃) and a diffused silicon pressure sensor (range 0-1MPa, accuracy ±0.5%FS) are installed before the valve.

[0044] The supply branch pipes of pipeline A and pipeline B are each connected to a stainless steel manifold. Each manifold has one inlet and eight outlets, with the eight outlets connected to the inlets of the liquid-cooled plates of eight modules. The outlets of the liquid-cooled plates of each module are connected to another stainless steel collector, which has eight inlets and two outlets. The two outlets are connected to the return branch pipes of pipeline A and pipeline B, respectively. The return branch pipes of pipelines A and B converge and flow into the evaporator inlet of the chiller unit. The chiller unit is a 45kW rated capacity variable frequency screw chiller with adjustable compressor speed, an outlet water temperature setting range of 12℃ to 25℃, and a temperature control accuracy of ±0.5℃. The chiller unit outlet is branched through a tee to the circulation pump inlets of pipeline A and pipeline B.

[0045] The centralized controller uses an industrial embedded controller with a quad-core ARM Cortex-A72 processor (1.8GHz), 8GB of RAM, 128GB of solid-state storage, and a built-in GPU for neural network inference. The controller communicates with the energy storage converter and battery management system via a CAN 2.0B bus, reads data from environmental temperature and humidity sensors via an RS485 bus, acquires signals from various pressure and temperature sensors via a 4-20mA analog interface, controls the variable frequency circulating pump speed via PWM signals, and controls the opening of the electric proportional regulating valve via a 4-20mA output. The controller runs a Linux real-time operating system, with predictive model inference, MPC online optimization, and safety monitoring tasks running in independent software threads.

[0046] The above hardware connection relationship and basic performance parameters of the equipment are a preferred example of this application. The installation position of each sensor, the selection of the manifold, the specific frame format of the communication protocol, and the internal refrigeration cycle and compressor control algorithm of the chiller unit are all existing technologies.

[0047] Based on the hardware described above, the detailed construction and online operation method of the prediction model in this embodiment is as follows: The construction and operation process of the Transformer prediction model described in step one is detailed below: Data Acquisition and Preprocessing: Operational data from the energy storage cabinet and similar cabinets over the past 12 months were selected for model training, with a data sampling period of 1 minute. Collected data items included: charging and discharging power reported by the energy storage converter (unit: kW, charging is positive, discharging is negative), SOC (percentage) reported by the battery management system, cell temperature of each module (128 measurement points in total), ambient dry-bulb temperature (°C), and ambient relative humidity (percentage). A timestamp feature was added to each sampling moment: Based on local electricity price periods, the day was divided into four periods: low, flat, high, and peak, coded from 0 to 3; weekday type was coded as weekday 0 and non-weekday 1; seasonal label was coded as spring 0, summer 1, autumn 2, and winter 3. All continuous features were Z-score standardized, and discrete features were encoded using one-hot encoding.

[0048] Training label calculation: Using battery terminal voltage V, current I, and SOC change rate dSOC / dt as inputs, the heat generation power Q of the cell at each moment is calculated online using an extended Kalman filter based on a second-order RC equivalent circuit model. The specific implementation of this extended Kalman filter belongs to a known algorithm in this field. Its key steps are briefly described as follows: Define the state vector x = [SOC, V1, V2, R0]^T, where V1 and V2 are the polarization voltages of the two RC loops, and R0 is the ohmic internal resistance; the state prediction equation updates SOC based on the ampere-hour integral method and updates polarization voltage based on the first-order difference equation; the observation equation is V_obs = OCV(SOC) -V1 - V2 - I·R0, and the OCV-SOC curve is provided by the battery manufacturer and calibrated through offline testing; the heat generation power calculation formula is Q = I²·R0 + I·(V1+V2), where the first term is ohmic heat and the second term is polarization heat. The initial values ​​of process noise covariance and observation noise covariance are obtained from offline mixed pulse power characteristic test data and are dynamically updated according to the innovation sequence through an adaptive covariance matching algorithm in each filtering iteration.

[0049] Model Structure and Training: The network structure of the prediction model is as follows: The encoder consists of four stacked Transformer encoding layers, each containing an 8-head self-attention mechanism and a feedforward network, with a hidden layer dimension of 256. The encoder input is a historical data sequence of the past 120 sampling times (i.e., 2 hours), containing 8 feature dimensions: power, SOC, ambient temperature, and time features. The decoder consists of two Transformer decoding layers, using the power plan sequence of the next 120 sampling times as the query input, and outputting the mean μ(t) and variance σ²(t) of the heat generation power corresponding to the 120 times. The loss function is negative Gaussian log-likelihood: Loss = 0.5·Σ[ln(σ²(t)) + (Qtrue(t)-μ(t))² / σ²(t)]. Training uses the Adam optimizer with an initial learning rate of 0.001, a batch size of 64, and 200 training epochs. The model parameters with the minimum loss on the validation set are saved. After training, the prediction uncertainty σ(t) is taken as the standard deviation of the predicted values ​​obtained by multiple Monte Carlo Dropouts (with Dropout enabled during inference and forward propagation repeated 50 times).

[0050] Online model updates: After the initial deployment of the model, the model is incrementally fine-tuned every 30 days using 30 days of newly added running data. The learning rate is reduced to 0.0001, and the fine-tuning is performed for 20 rounds to maintain the model's adaptability to battery aging and seasonal environmental changes.

[0051] The following example, using a typical 24-hour operation on a summer day, details the execution process and parameter changes of each step in the control method.

[0052] The local electricity price period for that day was divided as follows: 00:00-08:00 was the off-peak period; 08:00-11:00 and 13:00-18:00 were the normal periods; 11:00-13:00 and 18:00-21:00 were the peak periods; and 21:00-23:00 was the super-peak period. Based on this, the energy storage management system formulated a charging and discharging plan, and the central controller began implementing the control method described in this application from 00:00 on that day.

[0053] Period 1 (00:00-07:45): Off-peak charging, operating in mode 1. Starting at 00:00, the energy storage cabinet enters the off-peak charging phase, with a planned charging / discharging power of 0.2C and an actual charging power of approximately 100kW. The central controller triggers the calculations for steps one and two every 15 minutes: Step one's Transformer model outputs a heat generation prediction sequence and an uncertainty sequence based on historical data from the last two hours and the power plan for the next two hours; Step two calculates the demand flow rate F_req(t) based on the heat generation prediction.

[0054] Throughout the off-peak period, the cell SOC gradually increased from 20% to 80%, with the predicted heat output Q(t) fluctuating between 0.6kW and 1.0kW. The corresponding demand flow rate F_req(t) ranged from 10L / min to 16L / min, consistently below 0.8·F_max (24L / min). The prediction uncertainty σ(t) remained stable between 1.5kW and 1.9kW during this period, while the adaptive threshold σ_th, based on statistics from the same period over the past 30 days, was 2.0kW, and σ(t) remained consistently below σ_th.

[0055] Step 3's mode determination module detected that the above conditions were continuously met for six consecutive 15-minute cycles (90 minutes in total), confirming that the system was stably in a state where it could enter Mode 1. From 00:30, the system switched to single-path economic operation mode: the variable frequency pump in pipeline A operated at approximately 1200 rpm, the proportional valve opening was approximately 35%, and the actual liquid supply flow rate was maintained at approximately 13 L / min; the chiller unit outlet water temperature was set to 18℃. The pump in pipeline B was completely stopped, the proportional valve was closed, pipeline B was filled with coolant but there was no flow, and the power consumption was zero. Throughout the off-peak period, the average cell temperature remained between 26℃ and 29℃, with the highest temperature not exceeding 30℃. The actual power consumption of the pump in pipeline A was approximately 0.32kW, the chiller unit power consumption was approximately 1.8kW, and the total power consumption of the liquid cooling system was approximately 2.12kW.

[0056] Time Segment Two (07:45-08:00): Mode Switching Transition During the periodic calculation at 07:45, the load forecasting module detected that the future power plan indicated the start of the morning peak discharge period at 08:00, with the discharge power planned to switch from 0.2C charging to 1C discharging, a power change of 600kW. The predicted heat generation Q(t) rapidly increased from approximately 1.2kW at 07:55, and is expected to reach a peak of approximately 2.8kW by 08:10. Correspondingly, the demand flow rate F_req(t) surged from 18L / min to approximately 35L / min. During the process of the demand flow rate increasing from 18L / min to 35L / min, the maximum rate of increase (dF_req / dt)_max occurred within the steepest approximately 1-minute interval. Calculations showed that the flow rate increased by approximately 12L / min during this interval; therefore, (dF_req / dt)_max = 12 L / min ÷ 60s = 0.2 L / (min·s). Since 35L / min exceeds the maximum allowable flow rate of a single pipe F_max=30L / min, the exit condition for mode one is triggered.

[0057] Step 3: Calculate the lead time Δt_pre: The battery module thermal time constant τ, as tested offline, is approximately 90 seconds. The required additional flow rate ΔF_target = 35 - 18 = 17 L / min. Therefore, Δt_pre = 1.2 × 90 + 17 / 0.2 = 108 + 85 = 193 seconds, approximately 3.2 minutes. Thus, the system initiates the switching process at approximately 07:56:48. Since the switching target is the extreme supplementary sub-mode (F_req > F_max and σ < σ_th), F_A is expected to be fixed at 30 L / min, and F_B is expected to gradually increase from 0 to approximately 5 L / min.

[0058] After the switchover starts, a smooth transition period of T_trans=40 seconds begins: From 0 to 10 seconds, the pump in pipeline B starts from standby, with its speed increasing linearly from 0 to 600 rpm, while the proportional valve remains closed; From 10 to 30 seconds, the proportional valve in pipeline B increases linearly from 0% opening to the target opening of approximately 15% using a ramp function, while the speed of pump A in pipeline A is gradually adjusted from 1200 rpm to approximately 2200 rpm (corresponding to 30 L / min), and the speed of pump B in pipeline B increases from 600 rpm to approximately 800 rpm (corresponding to 5 L / min); From 30 to 40 seconds, the speeds and valve openings of both pipelines are finely adjusted and stabilized to the target values, and MPC resumes optimized operation.

[0059] During the entire transition period, the chiller outlet water temperature was reduced from 18℃ to 16℃ to accommodate the upcoming increase in heat load. The total flow rate of both pipelines increased steadily during the transition, without any pressure surges. The pressure fluctuations at the outlets of pipelines A and B were controlled within ±5 kPa. The highest cell temperature during the transition period only increased from 29.8℃ to 30.2℃.

[0060] Period 3 (08:00-11:00): Morning peak discharge, Mode 3 extreme compensation sub-mode operation. At 08:00, the discharge power reached 1C, or 500kW, as planned, while the actual heat generation power reached approximately 2.8kW, as predicted. The system was in Mode 3, the limit compensation sub-mode: Pipeline A was operating at full capacity, F_A = 30L / min, and the pump speed was approximately 2200rpm; for pipeline B, F_B was fine-tuned by the MPC between 4.5 and 5.5L / min based on real-time demand, and the pump speed fluctuated between 750 and 850rpm. The chiller outlet water temperature was optimized and set by the MPC between 15℃ and 16℃. The highest cell temperature reached the highest value of this segment at 08:05, 33.5℃, and then gradually stabilized between 32.8℃ and 33.2℃ under continuous cooling. The total liquid supply flow rate of the two pipes was approximately 35L / min, and the supply and return liquid temperature difference was approximately 4.8℃. The total power consumption of the liquid cooling system at this stage was approximately 4.5kW, of which the two pumps combined accounted for approximately 1.2kW, and the chiller unit accounted for approximately 3.3kW.

[0061] In contrast, a similar energy storage cabinet with the same configuration and traditional single-circuit PID control, under the same operating conditions, saw its cell temperature rise to 42.8℃ at 08:03 due to PID response lag, only gradually dropping below 38℃ after 08:15, with the temperature exceeding 38℃ for approximately 12 minutes. The proposed solution, by pre-starting the backup pipeline approximately 3 minutes in advance, reduced the temperature peak by about 9.3℃, and the temperature remained within the 38℃ safety limit throughout the entire process.

[0062] Time Segment 4 (11:00-13:00): Midday Peak, Mode 2 Operation The midday peak period is from 11:00 to 13:00, with a discharge power of 0.7C (350kW), a predicted heat generation power of approximately 1.9kW, and a demand flow rate F_req of approximately 26L / min, falling between 0.8·F_max and F_max. During this period, σ(t) is approximately 1.6kW, lower than σ_th (2.1kW). After the mode switching conditions are met for 30 seconds, the system switches from Mode 3 to Mode 2 dual-path balanced load-sharing mode.

[0063] In Mode 2, MPC performs rolling optimization with a 30-second cycle, and the calculated flow distribution coefficient α(t) dynamically varies between 0.55 and 0.62. For example, at 11:05, pump A in pipeline operates at approximately 1750 rpm with a flow rate of 15 L / min, while pump B in pipeline operates at approximately 1550 rpm with a flow rate of 11 L / min, totaling 26 L / min. The chiller's outlet water temperature is 17℃. The fluctuation of α(t) stems from MPC continuously seeking a combination of the two pumps' efficient operating ranges while meeting temperature constraints: when pump A's efficiency is slightly higher in this flow range, α(t) tends towards 0.6; when pump B's efficiency is better, α(t) tends towards 0.55. The total power consumption of the two pumps in this mode is approximately 0.68 kW, which is about 17% more energy-efficient than a fixed 1:1 distribution in both pipelines (each at 13 L / min, both pumps operating at approximately 1600 rpm, total power consumption approximately 0.82 kW).

[0064] Phase 5 (Simulated scenario of increased prediction uncertainty): Triggering active redundancy sub-mode To demonstrate the operation of the Mode 3 active redundancy sub-mode, it is assumed that during the 13:00 forecast period, due to a temporary change in external grid dispatching, a significant unexpected adjustment to the power plan occurs, causing the forecast uncertainty σ(t) output by the Transformer model to rise to 2.8kW, exceeding σ_th (2.1kW). However, the demand flow rate F_req remains at approximately 26L / min, falling within the range of 0.8·F_max to F_max. The activation conditions for the Mode 3 active redundancy sub-mode are met during this period.

[0065] The system switches to an active redundancy sub-mode: both pipelines continue to operate, and the MPC still optimizes α(t) and pump speed with the goal of minimizing total system energy consumption, but a penalty term for the rate of change of pump speed is added to the objective function. The weight of the penalty term is adaptively calculated based on σ(t) = 2.8kW (approximately 1.33 times σ_th). At this point, the penalty weight is moderately large, constraining α(t) to a relatively small range of 0.48 to 0.52. The flow rates of the two pipelines are almost evenly distributed, avoiding the oscillations that may be caused by the MPC frequently and significantly adjusting the flow distribution to pursue energy efficiency when the prediction is unreliable. In this mode, the total power consumption of the liquid cooling system is approximately 3.2kW, slightly higher than the 2.9kW in mode two, but significantly lower than the approximately 5.1kW of dual-pipe full-speed operation, achieving a balance between conservative safety and energy consumption.

[0066] Period Six (21:00-23:00): Peak discharge, Mode Three limit compensation During the peak period from 21:00 to 23:00, the discharge power returns to 1C. Similar to the morning peak, the system re-enters the mode three extreme compensation sub-mode operation. The process is similar to that of period three and will not be described in detail here.

[0067] To verify the emergency backoff function in step five, a proportional valve jamming fault was simulated in the operating pipeline A in the test environment. The fault triggering process is as follows: Under normal operating conditions, the flow rate of pipeline A is 28 L / min, and the supply and return liquid temperature difference is approximately 4.5℃. When the proportional valve actuator jams at approximately 15% opening under external command, the flow rate of pipeline A drops sharply from 28 L / min to approximately 9 L / min within 3 seconds, and the supply and return liquid temperature difference expands dramatically to approximately 9.2℃, reaching 1.84 times the standard design temperature difference of 5℃. Simultaneously, the pump outlet pressure rises from the normal 0.35 MPa to 0.62 MPa, and the pressure difference increases abnormally, exceeding the 30% protection threshold of the rated pressure difference and lasting for more than 5 seconds.

[0068] The centralized controller's safety monitoring thread detected the above anomaly within a 1-second sampling period and immediately triggered emergency control: interrupting the MPC thread, sending a rated maximum speed command (2900 rpm) to both pumps, a full-open command (100% opening) to both proportional valves, and a minimum outlet water temperature command (12℃) to the chiller unit. Pump B in pipeline accelerated to its rated speed from standby mode within approximately 8 seconds, and the total liquid supply flow rate rapidly increased from approximately 35 L / min to approximately 65 L / min.

[0069] Under the powerful emergency cooling, although the highest cell temperature briefly rose to 37.2℃ after the fault occurred, it did not exceed the hard safety line of T_max+δ1=39℃. Simultaneously, the cooling rate monitoring module detected a temperature drop rate of approximately 2.3℃ / min, exceeding the soft limit trigger value of 2℃ / min. It automatically adjusted the chiller unit's set temperature from 12℃ to 13.5℃, resulting in a soft landing of the actual cooling rate at approximately 1.8℃ / min. This avoided the risk of excessive internal temperature gradients and condensation on the cold plates that could have been caused by sudden cooling. After the fault was manually resolved, the flow rate in pipe A was restored, and all temperature and differential pressure parameters remained within the normal range for 5 minutes, the system automatically exited emergency mode after a delay and resumed operation in the load forecast-based control mode.

[0070] To objectively evaluate the technical effectiveness of this application, two identical 500kW / 1MWh liquid-cooled energy storage cabinets were subjected to a 30-day comparative operation experiment at the same site. The first cabinet served as the experimental cabinet, operating the dual-pipeline load prediction and control scheme described in this application; the second cabinet served as the control cabinet, operating a conventional single-pipeline PID control scheme combined with day-ahead load pre-adjustment. During the 30-day experiment, the daily charge and discharge power plans for both cabinets were completely identical, and the initial state of equilibrium (SOH) of the battery modules was 98% for both.

[0071] The comparison of various indicators in the experimental results is as follows.

[0072] Regarding the average maximum cell temperature, the proposed solution achieves 34.1℃, while the comparative solution achieves 40.3℃, representing a 6.2℃ reduction in peak temperature. In terms of the percentage of total operating time where cell temperature exceeds 38℃, the proposed solution achieves only 0.3%, compared to 8.7% for the comparative solution, a 96.6% reduction in over-temperature time. Regarding the peak cell temperature, the proposed solution achieves 36.8℃ under the worst operating conditions, compared to 43.5℃ for the comparative solution, a 6.7℃ reduction in peak temperature. In terms of the average daily total energy consumption of the liquid cooling system, the proposed solution consumes 47.2 kWh, compared to 58.5 kWh for the comparative solution, representing a 19.3% energy saving. Regarding pipeline switching pressure surges, the proposed solution achieves zero pressure surges per day through a smooth transition strategy, while the comparative solution experiences an average of 1.2 significant pressure fluctuations per day during passive fault switching.

[0073] The comparison results above show that the technical solution of this application is significantly superior to the existing technology in many aspects, such as temperature control performance, system energy consumption, operational stability and battery life protection.

[0074] While exemplary embodiments of this disclosure have been described, those skilled in the art will understand that various changes and modifications can be made to the exemplary embodiments of this disclosure without departing from the spirit and scope thereof. Therefore, all changes and modifications are included within the scope of protection of this disclosure as defined by the claims. This disclosure is defined by the appended claims, and equivalents of those claims are also included.

Claims

1. A dual-pipeline liquid-cooled energy storage cabinet, characterized in that, include: Two independent liquid supply lines, namely line A and line B; Each pipeline includes a variable frequency circulating pump, an electric flow regulating valve, a liquid supply temperature sensor, a pipeline pressure sensor, a liquid supply branch pipe, and a liquid return branch pipe. The liquid supply branch pipes of pipe A and pipe B converge into the liquid cooling plate inlet located below the battery module via a water distributor. The water outlet of the liquid cooling plate is connected to the return branch pipes of pipe A and pipe B via a water collector. After the return liquid branch pipes of pipe A and pipe B merge, they enter the evaporator side of the chiller unit for recooling. The water outlet of the chiller is split to the inlet of the circulation pump of pipe A and the inlet of the circulation pump of pipe B. The system also includes a centralized controller that receives power scheduling information from the energy storage converter, battery state of charge (SOC) and state of health (SOH) data from the battery management system, and data from ambient temperature and humidity sensors. The controller has a built-in load prediction module and an optimization control solver.

2. A load forecast-based control method, applied to the dual-pipeline liquid-cooled energy storage cabinet of claim 1, characterized in that, The control method is executed by a centralized controller and includes the following steps: Step 1: Load forecasting and heat production calculation; Step 2: Dynamic calculation of cooling demand; Step 3: Predicting and switching between dual-pipeline operating modes; Step 4: Joint optimization of flow distribution and liquid supply temperature; Step 5: Safety monitoring and emergency rollback.

3. The control method based on load forecasting according to claim 2, characterized in that, In step one: The centralized controller acquires the planned charge and discharge power sequence P(t) within a future time window T received by the energy storage converter in real time, where t∈[t0, t0+T], and the value of T is in the range of 2 to 4 hours, with a sampling period of 1 minute; Simultaneously, the current battery SOC, SOH, temperature at each measuring point of the battery module, ambient dry-bulb temperature, and relative humidity are collected. The above data, along with timestamp features including time period identifiers, weekday types, and season labels, are input into a pre-built and trained multivariate time series prediction model. This model employs a Transformer time series model that incorporates a convolutional self-attention mechanism. Its encoder extracts features from historical power, SOC, and ambient temperature sequences, while the decoder takes future power plans as the query input and outputs the weighted average heat generation power Q(t) of the battery module at each discrete time t within the future time T and its prediction uncertainty σ(t). The training labels for the prediction model are obtained as follows: using battery terminal voltage V, current I, and SOC change rate dSOC / dt as observations, an extended Kalman filter is constructed to back-calculate the heat generation power. The state vector x includes SOC, two polarization voltages V1 and V2, and ohmic internal resistance R0. The state equation is discretized based on the ampere-hour integral method and the polarization characteristic time constant. In the observation equation, the terminal voltage V_obs is obtained by subtracting the polarization voltage and ohmic voltage drop from the open-circuit voltage OCV. The heat generation power Q is composed of ohmic heat and polarization heat, i.e., Q = I²·R0 + I·(V1+V2). The process noise covariance matrix and the observation noise covariance matrix are set according to the battery offline test data and operation data statistics, and are updated online using an adaptive covariance matching algorithm to obtain the true heat generation power label value at each moment. The output heat production power prediction sequence {Q(t)} and the corresponding uncertainty sequence {σ(t)} are generated, where σ(t) is the root mean square of the prediction variance obtained by the prediction model through multiple samplings using Monte Carlo Dropout, which represents the confidence level of the model in predicting the point.

4. The control method based on load forecasting according to claim 2, characterized in that, In step two: A lumped parameter thermal equivalent circuit model of the battery module is constructed. The model parameters include the heat source generated by the battery internal resistance, the battery thermal capacity C_b, the conduction thermal resistance R_cond from the battery to the surface of the liquid cooling plate, and the convection thermal resistance R_conv from the surface of the liquid cooling plate to the coolant. The initial values ​​of the above model parameters are obtained through offline identification of multi-condition charge and discharge temperature rise experiments. To reflect the impact of battery aging on thermal characteristics, a correction function for SOH and thermal resistance is established: R_cond(SOH)=R_cond_0·[1+β·(1-SOH)], R_conv(SOH)=R_conv_0·[1+γ·(1-SOH)], where β and γ are empirical coefficients obtained by fitting thermal characteristic test data of the same model of battery at different cycle aging stages, SOH is taken as the current capacity retention rate reported by the battery management system, and C_b is considered to be not significantly different with SOH; Input the heat generation prediction sequence Q(t) into the thermal equivalent circuit model, and use the maximum allowable temperature of the battery cell T_max as the upper limit of the constraint. The default value of T_max is 38℃. Solve for the minimum liquid cooling power P_cool(t) required at each time t. P_cool(t) is converted into the total required flow rate F_req(t) and optimal supply temperature T_supply of the liquid cooling system. The conversion is based on: F_req(t) = P_cool(t) / (ρ·c_p·ΔT_design), where ρ is the density of the coolant, c_p is the specific heat capacity of the coolant, and ΔT_design is the designed standard temperature difference between the supply and return liquids, which is taken as 5℃. When P_cool(t) is in kW and c_p is in kJ / (kg·K), the calculated unit of F_req(t) is m³ / s. In practical applications, it is multiplied by 60000 to convert it to L / min. T_supply is selected based on the energy efficiency characteristic curve of the chiller unit, with the goal of minimizing the energy consumption of the chiller, and on the premise that the cooling capacity can cover P_cool(t). The final output is the total demand flow sequence F_req(t) and the recommended liquid supply temperature sequence T_supply(t) within the next time period T.

5. The control method based on load forecasting according to claim 2, characterized in that, In step three: First, we define the method for determining the prediction uncertainty threshold σ_th: σ_th is the 90th percentile of the prediction error for the same period in the past 30 days multiplied by a coefficient of 1.2, and is updated daily. Based on the total demand flow F_req(t) and the prediction uncertainty σ(t), the centralized controller determines the operating mode to be used at the current moment from the following three pre-set operating modes: Mode 1 is a single-pipe economic operation mode, where only pipeline A is in operation, while the circulation pump of pipeline B is stopped and its flow regulating valve is closed. The applicable conditions for this mode are: the total demand flow rate F_req(t) is not greater than 80% of the maximum allowable continuous flow rate F_max of a single pipeline, i.e., F_req(t) ≤ 0.8·F_max; the prediction uncertainty σ(t) is less than the threshold σ_th; and all prediction times within a preset forward time period ΔT_horizon (15 minutes) meet the above conditions. To avoid frequent mode switching, when switching from other modes to Mode 1, the above applicable conditions must be met continuously for at least 30 seconds; when switching from Mode 1 to other modes, the corresponding exit conditions must also be met continuously for more than 30 seconds; when it is predicted that the applicable conditions will no longer be met within the future ΔT_horizon, the mode is exited and transitions to other modes. Mode 2 is a dual-path balanced load-sharing mode, where two pipelines operate simultaneously, sharing the total demand flow according to a flow distribution coefficient α(t). That is, pipeline A provides F_A = α(t)·F_req(t), and pipeline B provides F_B = (1-α(t))·F_req(t), where 0 < α(t) < 1. The applicable conditions for this mode are: the total demand flow satisfies 0.8·F_max < F_req(t) ≤ F_max, and the prediction uncertainty σ(t) is less than σ_th. Switching to this mode requires the conditions to be met continuously for 30 seconds. When the flow is within this range but σ(t) ≥ σ_th occurs, for safety reasons, it switches to the active redundancy sub-mode in Mode 3. Mode 3 is the primary / standby switching redundancy mode, which includes two sub-modes: The extreme replenishment sub-mode is activated when F_req(t) > F_max. In this mode, pipeline A operates at full capacity, i.e., F_A = F_max, and pipeline B replenishes the remaining flow with F_B = F_req(t) - F_max. The flow allocation coefficient is fixed at F_max / F_req(t). The active redundancy sub-mode is activated when the total demand satisfies 0.8·F_max < F_req(t) ≤ F_max, but σ(t) ≥ σ_th, or when the system detects a pre-alarm signal in pipeline A, is activated. The pre-alarm signal includes a warning that the vibration amplitude of the pump bearing exceeds a preset vibration threshold. At this time, the two pipelines are still allowed to share the total flow equally. The allocation coefficient α(t) is calculated by subsequent optimization. However, a penalty term for the rate of change of pump speed is introduced into the optimization objective function to suppress frequent adjustments that may be caused by unreliable predictions. The weight of the penalty term is adaptively adjusted according to the size of σ(t): when σ(t) is close to σ_th, the penalty weight is small; when σ(t) is close to twice σ_th, the penalty weight increases sharply, so that the flow distribution tends to be uniform, and a conservative strategy is used to resist uncertain risks. When switching from Mode 1 or Mode 2 to Mode 3, pipeline B needs to be started and the flow rate pre-established in advance by a lead time Δt_pre. The lead time Δt_pre is determined by the following formula: Δt_pre = k·τ + ΔF_target / (dF_req / dt)_max, where τ is the thermal time constant of the battery module, ΔF_target is the required increase in flow rate, (dF_req / dt)_max is the maximum ramp rate of the required flow rate, and k is the safety factor, which is taken as 1.

2. The unit of Δt_pre is determined by the units of τ and ΔF_target / (dF_req / dt)_max. When τ is in seconds, ΔF_target is in L / min, and (dF_req / dt)_max is in L / (min·s), both are in seconds and can be added directly. The switching between modes adopts a smooth transition strategy: when it is determined that it is necessary to switch from the current mode to the target mode, the controller enters a preset transition time T_trans, which is 40 seconds. During the transition, a ramp function from the current opening to the target opening is applied to the flow regulating valve of pipeline B. The speed of the two pumps is adjusted synchronously according to the preset coordination curve. At the same time, the rolling optimization process of the model predictive control is paused. Optimization is resumed after the transition is completed. During the transition, if the temperature is detected to be close to the upper limit, the emergency mode is directly switched.

6. The control method based on load forecasting according to claim 2, characterized in that, In step four: A system control model is constructed. The state variables include the average battery temperature T_batt, the temperature difference between the inlet and outlet of the cold plate ΔT_cp, the flow rates F_A and F_B of pipes A and B respectively, and the liquid supply temperature T_supply. The control variables include the pump speed n_A of pipe A, the pump speed n_B of pipe B, the proportional valve opening θ, and the chiller set temperature T_chiller_set. The disturbance inputs are the heat generation power Q(t) and the ambient temperature T_amb. Considering the hydraulic coupling of the pipeline network formed by the distributor and the collector, the coupling gain matrix from the speed of the two pumps to their respective flow rates is identified offline. The branch pressure balance condition is introduced into the model predictive control constraints to ensure that the optimized flow distribution is physically feasible. Model predictive control rolling optimization is performed within each control cycle Δt_ctrl, where Δt_ctrl is 30 seconds. The constraints are: battery temperature T_batt ≤ T_max; flow rates in each pipeline satisfy 0 ≤ F_A, F_B ≤ F_max; pump speed is between the minimum allowable speed and the rated speed; and the liquid supply temperature is within the adjustable range of the chiller. The optimization objective is to minimize the total system energy consumption over the next H control cycles. The objective function is expressed as: min J = Σ_{k=1}^{H} [W_pumpA(k) + W_pumpB(k) + W_chiller(k)]·Δt_ctrl, where W_pumpA and W_pumpB are the electric power of pumps in pipes A and B, respectively, which are calculated in real time from the pump speed-flow-head characteristic curve and efficiency curve. W_chiller is the energy consumption of the chiller compressor, which is obtained according to the mapping relationship between cooling capacity, liquid supply temperature and ambient temperature. H is the number of control cycles contained in the prediction time domain, H=10. This nonlinear constraint optimization problem is solved online by the sequence quadratic programming algorithm built into the controller to obtain the control sequence of the next H steps. Only the control command of the first step is issued and executed. When in the extreme redundancy sub-mode, F_A is fixed to F_max, and F_B is determined by F_req(t) - F_max. At this time, the optimization problem only solves for the pump speed and chiller temperature of pipe B. In other modes, the flow rates of both pipes are optimization variables, and when in the active redundancy sub-mode, the objective function adds the pump speed change rate penalty term with the adaptive weight. The rolling optimization process described above is repeated in the next control cycle to achieve closed-loop correction.

7. The control method based on load forecasting according to claim 2, characterized in that, In step five: The centralized controller monitors safety parameters at a high frequency with a monitoring cycle of 1 second. The monitored objects include: the highest temperature T_batt_max at each measuring point of the battery module, the supply and return liquid temperature difference between pipeline A and pipeline B, the inlet and outlet pressure difference of the two circulation pumps, and the chiller fault signal. The model predictive control mode will be immediately interrupted and switched to the preset real-time feedback emergency control when any of the following conditions occur: T_batt_max > T_max + δ1, where δ1 is 1℃; the supply and return liquid temperature difference of any pipeline exceeds 1.5 times the design standard temperature difference; the differential pressure of any circulating pump drops instantaneously by more than 30% of its rated value and lasts for more than 5 seconds; or a fault shutdown command is received from the battery management system or energy storage converter. The emergency control strategy is as follows: both circulating pumps immediately switch to their rated maximum speed, the proportional valve is fully opened, and the chiller set temperature is reduced to the preset minimum safety value T_chiller_min; at the same time, in order to avoid excessive internal temperature difference or condensation caused by sudden cooling, a soft limit is implemented on the cooling rate: when the average temperature drop rate of the battery exceeds 2℃ / min, the chiller outlet water temperature setting is gradually adjusted to control the actual cooling rate within the range of 1.5 to 2℃ / min; Once the battery's maximum temperature drops below T_max - δ2 (where δ2 is 1℃), the supply and return electrolyte temperature difference returns to normal, and all alarm signals are cleared, the emergency mode will exit after manual confirmation or automatic delay. The automatic delay exit condition is that safety parameters such as temperature and pressure difference are within the normal range for 5 consecutive minutes. After that, the system will return to the control mode based on load prediction. The system records the time, type, and snapshots of the state before and after each emergency trigger event, which are used for subsequent model confidence assessment and adaptive adjustment of control parameters.

8. The control method based on load forecasting according to claim 3, characterized in that, The value of T is 2 hours.

9. A control method based on load forecasting according to claim 3, characterized in that, The input sequence of the prediction model is the data from the past 2 hours, and the output is the predicted heat production power point and variance for each minute of the next 2 hours. The model is fine-tuned every 30 days using the latest data.

10. A control method based on load forecasting according to claim 5, characterized in that, The maximum permissible continuous flow rate F_max for a single pipe is 30 L / min.

Images