A cold storage dynamic cold accumulation control system based on load prediction

Abstract

The application discloses a cold storage dynamic cold accumulation control system based on load prediction and belongs to the technical field of cold storage control. The system adopts a three-layer hierarchical control architecture of "load prediction-dynamic optimization-collaborative execution": a data acquisition and perception layer acquires multi-source real-time operation data; a load prediction and state estimation layer integrates a hybrid prediction model of a multi-scale time sequence gate network and a graph neural network, predicts the future time domain cold storage heat load, and estimates the state of the cold accumulation device in combination with an extended Kalman filter; a dynamic optimization decision layer solves the optimal start-stop sequence of the refrigerating unit and the optimal charging and discharging cold power curve of the cold accumulation device based on a model predictive control algorithm with the minimum operation cost as the target; and a collaborative execution control layer adopts a priority dynamic allocation strategy to decompose the total refrigeration demand to each evaporator for execution. The application realizes forward scheduling and precise control of cold storage refrigeration, ensures stable storage temperature, and significantly reduces operation energy consumption.

Description

Technical Field

[0001] This invention relates to the field of cold storage control technology, and in particular to a dynamic cold storage control system for cold storage based on load prediction. Background Technology

[0002] Cold storage facilities, as core infrastructure in the cold chain logistics system, account for over 40% of the total energy consumption of the entire cold chain system. Currently, cold storage refrigeration systems generally adopt a start-stop control strategy based on temperature upper and lower limits. That is, the refrigeration unit starts when the temperature inside the storage exceeds the set upper limit and stops when it falls below the set lower limit. This control method has the following prominent problems: First, the control strategy is based on fixed thresholds and fails to consider the time-varying characteristics of heat load. It lags in response to sudden load changes, resulting in large fluctuations in storage temperature, affecting storage quality and increasing energy consumption. Second, the charging and discharging strategies of the cold storage system lack forward-looking optimization and are mostly executed according to fixed schedules or thresholds. They cannot be dynamically adjusted according to future loads, resulting in low cold storage utilization. Third, there is a lack of coordinated scheduling among multiple evaporators. When some evaporators defrost or malfunction, the remaining evaporators cannot dynamically adjust the load distribution, which can easily lead to local overshoot or insufficient cooling.

[0003] Therefore, it is necessary to design a dynamic cold storage control system based on load forecasting to achieve accurate prediction of cold storage heat load, dynamic optimization scheduling of cold storage devices, and coordinated control of multiple evaporators, thereby minimizing refrigeration energy consumption while ensuring the stability of cold storage temperature. Summary of the Invention

[0004] This invention proposes a dynamic cold storage control system for cold storage based on load forecasting. To address the problems of slow response, inefficient cold storage scheduling, and lack of coordination among multiple evaporators in existing cold storage control systems, a three-layer hierarchical control architecture of "load forecasting - dynamic optimization - coordinated execution" is constructed to achieve accurate forecasting and forward-looking scheduling of cold storage cooling supply.

[0005] This invention provides a dynamic cold storage control system for cold storage based on load forecasting. The system includes: a data acquisition and sensing layer, a load forecasting and state estimation layer, a dynamic optimization decision-making layer, and a collaborative execution control layer.

[0006] The data acquisition and sensing layer is deployed in the cold storage body, refrigeration unit and cold storage device, integrating temperature sensor, humidity sensor, pressure sensor, power monitoring module and door status sensor to acquire multi-source real-time operating data.

[0007] The load forecasting and state estimation layer is electrically connected to the data acquisition and sensing layer. It processes multi-source real-time operating data to obtain the current heat load status of the cold storage and the remaining cold storage capacity of the cold storage device, and predicts the cold storage heat load sequence in the future time domain. This layer integrates a data fusion and preprocessing module, a cold storage heat load forecasting module, and a cold storage device state estimation module.

[0008] Data fusion and preprocessing module: Establishes a unified spatiotemporal benchmark, performs time synchronization, outlier detection and missing value imputation on multi-source real-time running data, and generates a synchronized multivariate time series dataset;

[0009] Cold storage heat load prediction module: Based on a synchronized multivariate time-series dataset, this module employs a hybrid prediction model based on multi-scale time-series gating networks and graph neural networks to perform multi-step rolling predictions of the cold storage heat load within a preset time window, outputting the cold storage heat load prediction sequence and its prediction confidence interval. The specific prediction process of this module includes:

[0010] Step A1: Extract each variable from the synchronized multivariate time series dataset by sliding window according to the time window, and construct the input tensor. The variables include the temperature inside the warehouse, the temperature outside the warehouse, the humidity inside the warehouse, the evaporation temperature of the refrigeration unit, the condensation temperature, the opening and closing status of the warehouse door, and the recent goods entering the warehouse.

[0011] Step A2: A multi-scale temporal gating network is used to extract temporal features from the input tensor. The multi-scale temporal gating network includes short-time memory gating units and long-time dependency gating units. The outputs of the two gating units are weighted and fused by adaptive fusion weights to generate fused temporal features.

[0012] Step A3: Divide the internal space of the cold storage into several sub-regions according to the coverage of the evaporator. Use each sub-region as a graph node and the heat transfer correlation strength between the sub-regions as the edge weight to construct a thermal correlation graph of the cold storage.

[0013] Step A4: Use a graph convolutional network to perform spatial feature aggregation on the fused temporal features generated in step A2 to obtain enhanced feature representations for each sub-region;

[0014] Step A5: Input the enhanced feature representation of each sub-region into the attention-based decoder. The decoder models the temporal dependency of each sub-region through a multi-head self-attention mechanism and outputs the predicted heat load value of each sub-region within a future preset time window.

[0015] Step A6: Spatially aggregate the predicted heat load values ​​of each sub-region to obtain the total heat load prediction sequence of the cold storage, and calculate the prediction confidence interval based on the statistical distribution of the prediction residuals;

[0016] Cold storage unit status estimation module: Based on synchronized multivariate time-series data and cold storage heat load prediction sequence, a state observer based on extended Kalman filter is used to estimate the remaining cold storage capacity, cold storage medium temperature distribution and available cooling rate of the cold storage unit in real time, and output a cold storage unit status estimation report.

[0017] The dynamic optimization decision layer, electrically connected to the load forecasting and state estimation layer, pre-stores real-time electricity price information obtained based on the national time-of-use electricity pricing policy. It takes the cold storage heat load forecast sequence, the cold storage device state estimation report, and real-time electricity price information as input, and aims to minimize the weighted sum of refrigeration unit operating costs and cold storage device depreciation costs. Employing a model predictive control-based dynamic optimization algorithm, it calculates the optimal start-stop sequence of the refrigeration unit and the optimal charging / discharging power curve of the cold storage device within the future control time domain, while satisfying constraints on cold storage temperature, refrigeration unit start-stop, and cold storage device charging / discharging rates. It then outputs a set of control commands. The specific optimization process of this layer includes: constructing the prediction and control time domains; establishing a discrete-time state-space model of the cold storage thermal dynamics; constructing an optimization objective function that includes electricity costs, depreciation costs, and cold storage temperature deviation penalties; setting upper and lower limits for cold storage temperature, unit start-stop constraints, and cold storage device constraints; using a sequential quadratic programming algorithm to obtain the optimal control sequence; and outputting the first control command and continuously optimizing in the next time step.

[0018] The collaborative execution control layer, electrically connected to the dynamic optimization decision layer, takes control command sets as demand input. Based on the current cooling capacity, defrosting status, operational health status, and temperature field distribution within the evaporator, it employs a priority-based dynamic allocation multi-evaporator collaborative scheduling strategy. This strategy decomposes the total cooling demand into cooling power setpoints for each evaporator and sends them to each evaporator controller via fieldbus for execution. The specific scheduling process at this layer includes: obtaining the current maximum cooling power, actual cooling power, and defrosting status of each evaporator; obtaining the temperature deviation and deviation rate of change for each sub-region; calculating the scheduling priority index for each evaporator based on the temperature deviation, deviation rate of change, and defrosting status; determining whether there is a capacity gap based on the sum of the total cooling demand and the available maximum cooling power, allocating cooling power according to the scheduling priority index, or triggering emergency cooling release from the cold storage device; and sending the allocated setpoints to each evaporator controller for execution.

[0019] The refrigeration unit and the cold storage device are connected in parallel to each evaporator through a cold supply pipeline. The cold storage device is charged with cold during the operation of the refrigeration unit and released to supplement the cold supply when the refrigeration unit is shut down or when the cooling capacity is insufficient.

[0020] Furthermore, the multi-scale temporal gating network includes short-term memory gating units and long-term dependency gating units. The bias of the short-term memory gating unit is initialized to a negative value to quickly forget transient fluctuations, while the bias of the long-term dependency gating unit is initialized to a positive value to maintain long-term dependencies. The outputs of the two gating units are weighted and fused by adaptive fusion weights to generate a comprehensive gating output.

[0021] Furthermore, the dynamic optimization decision layer adopts a dynamic optimization algorithm based on model predictive control. Its optimization objective function includes three items: the electricity cost of the refrigeration unit, the depreciation cost of the cold storage device, and the penalty term for temperature deviation in the storage room.

[0022] Furthermore, the multi-evaporator collaborative scheduling strategy in the collaborative execution control layer calculates the scheduling priority index of each evaporator based on a comprehensive weighted average of temperature deviation, deviation change rate, defrosting status, and operational health status, and allocates cooling power from high to low according to the scheduling priority index.

[0023] Furthermore, the cold storage device is a phase change cold storage device, which is filled with phase change cold storage material with a phase change temperature of -5℃ to -15℃. The cold storage device is connected to the evaporator side of the refrigeration unit and the cooling supply side of each evaporator through a refrigerant circulation pipeline.

[0024] Furthermore, the system also includes a remote monitoring and management platform deployed on a cloud server. It communicates bidirectionally with the control host where the load forecasting and state estimation layer and the dynamic optimization decision layer are located via a cellular network, and integrates a historical data storage module, a forecasting model online update module, and a visualization monitoring module.

[0025] Furthermore, the temperature sensors in the data acquisition and sensing layer include an array of multi-point temperature sensors inside the cold storage and ambient temperature sensors outside the cold storage. The array of multi-point temperature sensors inside the cold storage is arranged according to a three-dimensional grid of the internal space of the cold storage. The measured values ​​of each temperature sensor are spatially interpolated by the data fusion and preprocessing module to generate a temperature field distribution map inside the cold storage.

[0026] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0027] This invention constructs a multi-scale temporal gating network, employing short-time memory gating units initialized with negative biases and long-time dependency gating units initialized with positive biases to capture the rapidly fluctuating components and slowly changing trend components in the heat load sequence, respectively. These components are then dynamically weighted and fused using adaptive fusion weights, achieving a coordinated representation of the cold storage heat load across different time scales. Furthermore, by constructing a thermal correlation graph of the cold storage and using a graph convolutional network for spatial feature aggregation, the invention effectively captures the heat transfer coupling relationships between sub-regions, enabling the model to simultaneously perceive local transient disturbances and global trend changes. Compared to existing control methods based on fixed thresholds, this invention can detect and respond before load mutations occur, preventing large fluctuations in storage temperature from the source.

[0028] This invention takes a cold storage heat load forecast sequence, a cold storage device status estimation report, and real-time electricity price information obtained based on the national time-of-use electricity pricing policy as inputs. The optimization objective is to minimize the weighted sum of refrigeration unit operating costs and cold storage device depreciation costs. A model predictive control algorithm is used for rolling optimization. Under the premise of satisfying cold storage temperature constraints, refrigeration unit start-up and shutdown constraints, and cold storage device charging and discharging rate constraints, the optimal start-up and shutdown sequence of the refrigeration unit and the optimal charging and discharging power curve of the cold storage device are solved in the future control time domain. This mechanism upgrades the cold storage device's charging and discharging strategy from a traditional fixed schedule or threshold rule to a dynamic forward-looking scheduling based on load forecasting and electricity price signals. It can proactively charge during periods of low electricity prices and prioritize discharging during periods of high electricity prices, significantly reducing system operating costs while ensuring stable cold storage temperature.

[0029] This invention, within the collaborative execution control layer, calculates the scheduling priority index for each evaporator based on its temperature deviation, deviation rate of change, defrosting status, and operational health status within the sub-regions covered by each evaporator, and allocates refrigeration power according to priority from high to low. This strategy solves the resource competition and action conflicts caused by the lack of coordination among multiple evaporators in existing cold storage facilities, achieving precise allocation of cooling resources to the most scarce areas. Simultaneously, by introducing an operational health status indicator into the scheduling priority index, the system can automatically isolate faulty evaporators from power allocation and automatically trigger the emergency cooling mode of the cold storage device to supplement the gap when refrigeration capacity is insufficient, significantly improving the system's operational reliability and cooling continuity under special operating conditions such as evaporator defrosting and malfunctions. Attached Figure Description

[0030] Figure 1 This is a system structure block diagram of a dynamic cold storage control system for cold storage based on load prediction proposed in this invention;

[0031] Figure 2 This is a schematic diagram of the structure of a multi-scale temporal gating network in an embodiment of the present invention. Detailed Implementation

[0032] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0033] Example 1:

[0034] according to Figure 1 This invention provides a dynamic cold storage control system for cold storage based on load forecasting, suitable for medium and large-sized cold storage facilities and cold chain logistics distribution centers. The system includes:

[0035] The data acquisition and sensing layer is deployed within the cold storage body, refrigeration units, and cold storage devices. Specifically, it includes: a PT100 platinum resistance temperature sensor array (measuring range -40℃ to +50℃, accuracy ±0.1℃), arranged in a 3D grid of 3 m × 3 m × 2 m within the cold storage space, totaling 36 measuring points; an external ambient temperature and humidity sensor (model HC2A-S3, temperature measuring range -40℃ to +60℃, accuracy ±0.2℃; humidity measuring range 0 to 100% RH, accuracy ±1% RH); and refrigeration unit suction pressure and exhaust pressure sensors (measuring range -0.1 to 3.5 m). MPa, accuracy class 0.5); power monitoring module (accuracy class 0.5, monitoring the real-time power of the refrigeration unit and evaporator fan); warehouse door status sensor (magnetic reed switch type, detecting the opening and closing status and opening duration of the warehouse door); goods entering the warehouse recorder (recording the type, quantity and initial temperature of the goods entering the warehouse via RFID reader); all sensor data are collected to the field data acquisition terminal via RS-485 bus (Modbus-RTU protocol), with a sampling cycle of 10 seconds / time;

[0036] The load forecasting and state estimation layer is electrically connected to the data acquisition and sensing layer and is deployed on a field industrial control computer (configured with an Intel Core i7 processor, 16GB of memory, a 512GB solid-state drive, and running a Linux real-time operating system). This layer integrates a data fusion and preprocessing module, a cold storage heat load forecasting module, and a cold storage device state estimation module.

[0037] Data fusion and preprocessing module: Based on the high-precision clock inside the field industrial control computer, it adds precise timestamps (accuracy to milliseconds) to the data of each sensor, uses an outlier detection algorithm based on local outlier factors to identify and mark outlier data points, and uses cubic spline interpolation to imput missing values ​​for the marked outliers, generating a synchronized multivariate time series dataset with an update frequency of 10 seconds / time.

[0038] Cold storage heat load prediction module: Based on a synchronized multivariate time series dataset, a hybrid prediction model based on multi-scale time series gating network and graph neural network is used to perform multi-step rolling prediction of the cold storage heat load for the next 4 hours (prediction step size 15 minutes / step), and output the cold storage heat load prediction sequence and its prediction confidence interval.

[0039] The cold storage unit state estimation module: Based on synchronized multivariate time-series data and cold storage heat load prediction sequences, a state observer based on extended Kalman filtering is used to estimate the remaining cold storage capacity, average temperature of the cold storage medium, and available cooling rate of the phase change cold storage unit in real time, and outputs a cold storage unit state estimation report with an update frequency of 1 minute / time; the state variables of the extended Kalman filter include the temperature and remaining cold storage capacity of the cold storage medium in each layer of the cold storage unit, the state transition equation is established based on the heat balance equation of the cold storage medium, and the observation equation takes the inlet and outlet refrigerant temperature and flow rate of the cold storage unit as input.

[0040] The dynamic optimization decision layer is electrically connected to the load forecasting and state estimation layer and deployed on the same industrial control computer. This layer pre-stores real-time electricity price information obtained based on the national time-of-use electricity pricing policy. Taking the cold storage heat load forecast sequence, the cold storage device state estimation report, and the real-time electricity price information as input, and minimizing the weighted sum of the operating cost of the refrigeration unit and the depreciation cost of the cold storage device as the optimization objective, a dynamic optimization algorithm based on model predictive control is adopted. Under the premise of satisfying the cold storage temperature constraint (set temperature ±1℃), the refrigeration unit start-up and shutdown constraint (minimum downtime 10 minutes, minimum running time 15 minutes), and the cold storage device charging and discharging rate constraint (maximum charging power 50 kW, maximum discharging power 40 kW), the optimal start-up and shutdown sequence of the refrigeration unit and the optimal charging and discharging power curve of the cold storage device in the future control time domain are calculated, and the control command set is output. The optimization cycle is 15 minutes / time.

[0041] The collaborative execution control layer is electrically connected to the dynamic optimization decision layer and is deployed in the field of each evaporator using an embedded controller (STM32F407ZGT6 main control chip). This layer takes the control instruction set as the demand input and, based on the current cooling capacity, defrosting status and temperature field distribution of each evaporator, adopts a priority-based dynamic allocation multi-evaporator collaborative scheduling strategy to decompose the total cooling demand into the cooling power set value of each evaporator and send it to the expansion valve controller and fan frequency converter of each evaporator via CAN fieldbus (communication rate 1 Mbps) for execution.

[0042] The refrigeration unit consists of a twin-screw refrigeration compressor (150 kW cooling capacity, 45 kW input power) and a phase change cold storage device (300 kWh cold storage capacity, -10℃ phase change temperature). These devices are connected in parallel to each evaporator via cooling pipelines. During the operation of the refrigeration unit, the cold storage device absorbs cold energy through the circulation of a refrigerant (30% ethylene glycol aqueous solution, freezing point approximately -15℃) for charging. During the shutdown of the refrigeration unit or when the cooling capacity is insufficient, the cold energy is released to the evaporator through the refrigerant for supplemental cooling.

[0043] Example 2:

[0044] This embodiment is based on Embodiment 1. In this embodiment, the cold storage heat load prediction module adopts a hybrid prediction model based on multi-scale temporal gating network and graph neural network to predict the cold storage heat load. The specific steps include:

[0045] Step A1: Extract each variable in the synchronized multivariate time series dataset by sliding window (window length is the past 6 hours, step size is 15 minutes, and a total of 24 time steps) and construct the input tensor. The variables include the average temperature inside the warehouse, the temperature outside the warehouse, the humidity outside the warehouse, the evaporation temperature of the refrigeration unit, the condensation temperature, the cumulative opening time of the warehouse door, and the heat of recent goods entering the warehouse.

[0046] Step A2: A multi-scale temporal gating network is used to extract temporal features from the input tensor. Its structure is as follows: Figure 2 As shown, the multi-scale temporal gating network includes short-term memory gating units and long-term dependency gating units. The bias of the short-term memory gating unit is initialized to a negative value, making it more sensitive to recent transient fluctuations and faster forgetting. The bias of the long-term dependency gating unit is initialized to a positive value, making it retain a stronger memory of long-term dependencies across time steps. The two gating units capture the fast-fluctuating components and the slow-changing trend components in the load sequence, respectively. The network further constructs an adaptive fusion gating system, dynamically calculating the fusion weights of the two gating units based on the current input features and historical hidden states. The two outputs are weighted and summed to generate a fused comprehensive gating output, which is then used as a forgetting gating signal to participate in cell state updates. Let the outputs of the short-term memory gating unit and the long-term dependency gating unit be respectively... and Adaptive fusion gating calculates fusion weights based on the current input and the hidden state of the previous time step. Then the comprehensive gating output for:

[0047] ;

[0048] In the above formula, When the value approaches 1, the network tends to favor short-term memory units, making it more sensitive to sudden fluctuations in load. As the load approaches zero, the network tends to favor long-term dependent units, maintaining a stronger ability to track the overall trend of the load. Building upon this, the network employs a coupled input and forgetting gating mechanism, coupling input gating with fusion gating output. It updates the cell state (i.e., the long-term memory maintained internally by the network) according to the cell state update rules of standard Long Short-Term Memory (LSTM) networks, while simultaneously outputting the hidden state at the current time step. This results in a fused temporal feature, which remains independent for each sub-region, forming a node feature matrix. 4 represents the number of sub-regions. To hide the state dimension;

[0049] Step A3: Divide the internal space of the cold storage into 4 sub-regions (corresponding to 4 evaporators) according to the coverage area of ​​the evaporators. Using each sub-region as a graph node and the heat transfer correlation strength between sub-regions as edge weights, construct a thermal correlation graph for the cold storage. The heat transfer correlation strength is a weighted composite of two parts: the first part is obtained by mapping the spatial distance between sub-regions using a Gaussian kernel function, reflecting the direct heat conduction effect of neighboring regions; the second part is based on the time-delay cross-correlation coefficient of historical temperature sequences between sub-regions, capturing the delayed heat diffusion driven by airflow and the temperature coupling relationship between regions; edge weights... Calculate using the following formula:

[0050] ;

[0051] in, and Index for sub-regions; sub-region sub-region Spatial distance between them (unit: meters); The distance attenuation scale parameter controls the range of influence of spatial proximity on edge weights, and is set to 5 meters. sub-region sub-region The absolute value of the cross-correlation coefficient of the historical temperature series under the optimal time delay reflects the synchronicity and lag correlation strength of temperature fluctuations between two regions. It is calculated by searching for the time delay that maximizes the absolute value of the correlation coefficient within the maximum time delay of 30 minutes. The weighting coefficient is 0.3, which adjusts the contribution ratio of the spatial distance term and the time delay related term. This hybrid adjacency matrix not only ensures strong connectivity between spatially adjacent regions, but also captures the thermal coupling relationship between non-adjacent regions caused by airflow organization or cargo stacking methods.

[0052] Step A4: A two-layer graph convolutional network is used to perform spatial feature aggregation on the fused temporal features. The first layer of graph convolution maps the feature dimension from the hidden state dimension to 64 dimensions, and the second layer of graph convolution maps the 64 dimensions to 32 dimensions, to obtain the enhanced feature representation of each sub-region. This feature representation not only integrates the historical temporal information of each sub-region, but also gathers the correlation information brought by the heat transfer between adjacent sub-regions.

[0053] Step A5: Input the enhanced feature representation of each sub-region into the attention-based decoder. The decoder performs parallel modeling of the temporal dependencies of each sub-region through a multi-head self-attention mechanism and outputs the heat load prediction value of each sub-region in the next 4 hours (16 time steps in total, 15 minutes per step).

[0054] Step A6: Spatially sum and aggregate the predicted heat load values ​​of each sub-region to obtain the total predicted heat load sequence of the cold storage, and calculate the 95% prediction confidence interval based on the statistical distribution of the historical prediction residuals (assuming that the residuals follow a normal distribution with a mean of zero).

[0055] Example 3:

[0056] This embodiment is based on Embodiment 2. In this embodiment, the process of the dynamic optimization decision layer calculating the control instruction set using a model predictive control-based dynamic optimization algorithm specifically includes the following steps:

[0057] Step C1: Starting from the current time, construct a prediction time domain Tp=4 hours and a control time domain Tc=1 hour, with a discretization step size of 15 minutes. That is, the prediction time domain contains 16 discrete time steps and the control time domain contains 4 discrete time steps.

[0058] Step C2: Establish a discrete-time state-space model of the cold storage's thermal dynamics. The model uses the storage temperature, the cold storage capacity of the cold storage device, and the operating status of the refrigeration unit as state variables, and the start / stop commands of the refrigeration unit and the charging / discharging power of the cold storage device as control variables. The state transition equation describes the evolution of the state variables between adjacent discrete time steps. The change in storage temperature is jointly determined by the cold supply provided by the refrigeration unit, the charging / discharging power of the cold storage device, and the predicted sequence of the cold storage's heat load. The change in cold storage capacity is determined by the cumulative effect of the charging / discharging power. The operating status of the refrigeration unit is switched according to the start / stop commands. The output equation describes the mapping relationship between the observed storage temperature and the state variables at the current moment.

[0059] Step C3: Construct the optimization objective function; this function consists of three parts: the first part is the electricity cost generated by the operation of the refrigeration unit in the prediction time domain, which is obtained by multiplying the real-time electricity price at each time point by the operating power of the refrigeration unit and then summing the results; the second part is the depreciation cost generated by the charging and discharging process of the cold storage device, which is obtained by multiplying the unit depreciation cost of the cold storage device by the absolute value of the charging and discharging power at each time point and then summing the results; the third part is the penalty term for the cold storage temperature deviating from the set value, which is obtained by summing the weighted square of the cold storage temperature deviation in the prediction time domain, constraining the cold storage temperature to be as close to the set value as possible; Optimize the objective function. As shown below:

[0060] ;

[0061] in, For discrete time step index; For the first The real-time electricity price at each time step is obtained by looking up the real-time electricity price information pre-stored by the dynamic optimization decision-making layer based on the national time-of-use electricity pricing policy; For the first The operating power of the chiller unit at each time step is determined by the unit's start-up and shutdown status at the current moment. It is 45 kW when the unit is running and 0 kW when it is stopped. The unit depreciation cost for charging and discharging the cold storage device is taken as 0.02 yuan / kWh; For the first The charging and discharging power of the cold storage device at each time step is represented by a positive value indicating charging and a negative value indicating discharging. The weighting coefficient for the temperature deviation penalty is set to 10. For the first Predicted temperature inside the warehouse at each time step; The storage temperature setting is determined based on the type of stored goods; in this embodiment, it is set to -18°C. The weighting coefficient is adjusted accordingly. and unit depreciation cost It allows for a flexible balance between energy saving and temperature control accuracy;

[0062] Step C4: Set optimization constraints, specifically including: upper and lower limit constraints on storage temperature to ensure that the storage temperature is always within the safe range allowed for the goods; minimum downtime and minimum running time constraints for refrigeration units to prevent frequent start-stop of units from causing equipment damage and energy efficiency reduction; maximum start-stop frequency constraints for refrigeration units within the prediction time domain to further limit the start-stop frequency; upper and lower limit constraints on the cold storage capacity of the cold storage device to prevent overcharging or venting from causing the cold storage medium to fail; and maximum charge / discharge rate constraints for the cold storage device, which are determined by the heat exchange capacity of the cold storage device.

[0063] Step C5: The above optimization objective function is solved under various constraints using a sequential quadratic programming algorithm. This algorithm transforms the original optimization problem into a series of quadratic programming subproblems, and gradually approximates the optimal control sequence through iterative solutions, thereby obtaining the optimal start-stop command of the chiller unit and the optimal charging and discharging power of the cold storage device at each discrete moment in the control time domain.

[0064] Step C6: Output the first control command in the optimal control sequence (i.e., the start / stop command of the refrigeration unit and the charging / discharging power of the cold storage device at the current moment) as a control command set to the collaborative execution control layer; at the next sampling moment (15 minutes later), the system re-performs the entire optimization solution process based on the latest cold storage device state estimate and cold storage heat load prediction value to achieve rolling optimization.

[0065] Example 4:

[0066] This embodiment is based on Embodiment 3. In this embodiment, the multi-evaporator collaborative scheduling strategy based on priority dynamic allocation in the collaborative execution control layer specifically includes the following steps:

[0067] Step D1: Obtain the current maximum cooling power (kW), current actual cooling power (kW), defrosting status (defrosting in progress or not defrosting), and operating health status (normal or faulty) of the four evaporators via the CAN fieldbus; the current maximum cooling power is determined by the evaporator heat exchange area, fan speed, and refrigerant inlet temperature; the defrosting status indicator indicates whether the evaporator is in a defrosting cycle; the operating health status indicator indicates whether the evaporator is faulty, and faulty evaporators do not participate in subsequent power allocation, and their cooling power is considered zero;

[0068] Step D2: Obtain the deviation between the current temperature and the set temperature value of each sub-region covered by the evaporator and the rate of change of the deviation through the temperature sensor array in the warehouse; the temperature deviation reflects the degree of deviation of the region from the target, and the rate of change of the deviation reflects the trend and speed of temperature change;

[0069] Step D3: Calculate the scheduling priority index for each evaporator based on a comprehensive weighted average of temperature deviation, deviation change rate, defrosting status, and operational health status; let the first... The scheduling priority index of the evaporator is The calculation formula is as follows:

[0070] ;

[0071] in, For evaporator index, the value range is 1~4; for the evaporator index, the value range is 1~4. The current temperature deviation of the sub-region covered by the evaporator is equal to the current average temperature of the sub-region minus the storage temperature setpoint. The rate of change of temperature deviation is calculated from the difference between the two most recent sampled values. A positive value indicates that the temperature is rising and the deviation is worsening. This is a defrosting status indicator; a value of 1 indicates that the evaporator is defrosting, and a value of 0 indicates that it is not defrosting. This is an indicator of the evaporator's operational health status. A value of 1 indicates that the evaporator is operating normally, while a value of 0 indicates that the evaporator is malfunctioning. , , The weighting coefficients are 2.0, 1.5, and 3.0, respectively. The temperature deviation term reflects the urgency of the current cooling supply, the deviation change rate term reflects the trend of temperature deterioration, and the defrost status term will downgrade the priority of the evaporator that is defrosting to avoid assigning too heavy a load during the defrost period.

[0072] Step D4: Obtain the total cooling demand in the control command set and compare it with the sum of the current maximum cooling power of each normally operating evaporator; if the total cooling demand does not exceed the sum of the current available maximum cooling power, then allocate cooling power to each normally operating evaporator in descending order of scheduling priority index until the total demand is fully met, and the allocated power of a single evaporator does not exceed its current maximum cooling power; if the total cooling demand is greater than the sum of the current available maximum cooling power, then it is determined that there is a cooling capacity gap, and at this time the emergency cooling mode of the cold storage device is triggered, and the cold storage device supplements the gap with a maximum cooling power of 40 kW.

[0073] Step D5: The allocated cooling power setpoints for each evaporator are sent to the electronic expansion valve controller and fan inverter of each evaporator via the CAN fieldbus. The actuators then adjust the actual cooling power to track the setpoints in a closed loop, with an execution cycle of 15 seconds per cycle.

[0074] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

Claims

1. A dynamic cold storage control system for cold storage based on load forecasting, characterized in that, It includes a data acquisition and sensing layer, a load forecasting and state estimation layer, a dynamic optimization decision-making layer, and a collaborative execution control layer; The data acquisition and sensing layer is deployed in the cold storage body, refrigeration unit and cold storage device to acquire multi-source real-time operating data; The load forecasting and status estimation layer receives real-time operating data from multiple sources and generates cold storage heat load forecasting sequences and cold storage device status estimation reports. The dynamic optimization decision layer has pre-stored real-time electricity price information. It takes the cold storage heat load prediction sequence, the cold storage device status estimation report and real-time electricity price information as inputs, and takes the weighted sum of the operating cost of the refrigeration unit and the depreciation cost of the cold storage device as the optimization objective. It adopts a dynamic optimization algorithm based on model predictive control. Under the premise of satisfying the constraints of cold storage temperature, refrigeration unit start-up and shutdown constraints and cold storage device charging and discharging rate constraints, it calculates the optimal start-up and shutdown sequence of the refrigeration unit and the optimal charging and discharging power curve of the cold storage device in the future control time domain, and outputs the control command set. The collaborative execution control layer receives the control command set and, based on the cooling capacity, defrosting status, health status, and temperature field distribution of each evaporator, adopts a priority-based dynamic allocation multi-evaporator collaborative scheduling strategy to decompose the total cooling demand into the cooling power setpoints of each evaporator and issue them for execution. The refrigeration unit and the cold storage device are connected in parallel to each evaporator through a cold supply pipeline. The cold storage device is charged with cold during the operation of the refrigeration unit and released cold to supplement the cold supply when the refrigeration unit is shut down or when the cooling capacity is insufficient.

2. The system according to claim 1, characterized in that, The load forecasting and state estimation layer includes the following modules: Data fusion and preprocessing module: performs time synchronization and preprocessing on multi-source real-time running data to generate synchronized multivariate time series datasets; Cold storage heat load prediction module: Based on multivariate time series dataset, it adopts a hybrid prediction model based on multi-scale time series gating network and graph neural network to perform multi-step rolling prediction of cold storage heat load in the future time domain, and outputs cold storage heat load prediction sequence and prediction confidence interval. The cold storage device state estimation module uses a state observer based on extended Kalman filtering to estimate the remaining cold storage capacity, cold storage medium temperature distribution, and available cooling rate of the cold storage device in real time, and outputs a cold storage device state estimation report.

3. The system according to claim 2, characterized in that, The multi-scale temporal gating network includes short-term memory gating units and long-term dependency gating units; the bias of the short-term memory gating units is initialized with negative values ​​to quickly forget transient fluctuations; the bias of the long-term dependency gating units is initialized with positive values ​​to maintain long-term dependencies. The outputs of the two gating units are weighted and fused using adaptive fusion weights to generate a comprehensive gating output.

4. The system according to claim 2, characterized in that, The cold storage heat load prediction module further divides the internal space of the cold storage into several sub-regions according to the coverage of the evaporator. It constructs a cold storage thermal correlation graph with each sub-region as a graph node and the heat transfer correlation strength between sub-regions as edge weights. It also uses a graph convolutional network to aggregate the spatial features of the temporal features to obtain the enhanced feature representation of each sub-region. Finally, it outputs the predicted heat load value of each sub-region through a decoder based on a multi-head self-attention mechanism. The edge weights are composed of two weighted components: the first component is based on the Gaussian kernel function mapping of the spatial distance between sub-regions, and the second component is based on the time delay cross-correlation coefficient of the historical temperature series between sub-regions.

5. The system according to claim 1, characterized in that, The objective function of the dynamic optimization algorithm includes the electricity cost of the refrigeration unit, the depreciation cost of the cold storage device, and the temperature deviation penalty. The temperature deviation penalty constrains the temperature to be close to the set value.

6. The system according to claim 4, characterized in that, The priority-based dynamic allocation multi-evaporator collaborative scheduling strategy includes: calculating the scheduling priority index of each evaporator based on the temperature deviation, temperature deviation change rate, whether the evaporator is in defrosting state, and operating health status of the sub-region covered by each evaporator, and allocating cooling power from high to low according to the scheduling priority index; wherein, the operating health status identifier of the faulty evaporator sets its scheduling priority index to zero, so that it does not participate in the cooling power allocation.

7. The system according to claim 1, characterized in that, The cold storage device is a phase change cold storage device, which is filled with phase change cold storage material with a phase change temperature of -5℃ to -15℃. The cold storage device is connected to the evaporator side of the refrigeration unit and the cooling side of each evaporator through a refrigerant circulation pipeline.

8. The system according to claim 1, characterized in that, It also includes a remote monitoring and management platform, deployed on a cloud server, which communicates bidirectionally with the control host where the load forecasting and state estimation layer and the dynamic optimization decision layer are located via a cellular network; the remote monitoring and management platform integrates a historical data storage module, a forecasting model online update module, and a visualization monitoring module.

9. The system according to claim 1, characterized in that, The data acquisition and sensing layer includes an array of multi-point temperature sensors deployed according to a three-dimensional grid within the cold storage space; The data fusion and preprocessing module performs spatial interpolation on the measured values ​​of each temperature sensor to generate a temperature field distribution map within the storage area.

Images