An intelligent temperature and humidity control system inside a granary

Abstract

The present application relates to the technical field of temperature and humidity control, in particular to a kind of intelligent temperature and humidity control system inside granary, including depth time sequence prediction network construction neural dynamics observer, utilize space-time attention mechanism to capture the penetration law of external meteorology to grain pile, total disturbance state is predicted in advance;Combining reinforcement learning generates feedforward decoupling compensation component, constructs temperature and humidity independent control loop, logically peels off the strong coupling between variables;At the same time, the system adopts multi-objective optimization neural network and multi-gate hybrid expert architecture, combines water potential gradient and air enthalpy value quantization energy state saturation, realizes the adaptive scaling of feedback gain, finally utilizes multi-head attention mechanism to project and map global instruction into the power distribution weight of each execution mechanism;The present application effectively inhibits the irrational respiration and moisture loss of grain, solves the hysteresis and equipment interference problems of traditional regulation and control, realizes the high steady state and energy efficiency optimization of warehouse environment.

Description

Technical Field

[0001] This invention relates to the field of temperature and humidity control technology, specifically to an intelligent temperature and humidity control system for grain storage. Background Technology

[0002] Currently, traditional management models mainly rely on manual inspections and experience-driven mechanical ventilation operations. Under this model, grain, as a biologically active organism, has its physiological activities directly regulated by environmental temperature and humidity. The complex nonlinear heat and mass exchange process poses a great challenge to precise control.

[0003] Existing control systems face serious problems of variable coupling and control inaccuracy when handling such controlled objects. First, the cooling process is often accompanied by unexpected moisture loss from the controlled medium. This is because the internal environment of the medium is a typical nonlinear and unsteady-state process; there is a complex dynamic balance between its surface vapor pressure and ambient humidity. Traditional control logic is usually based on simple temperature and humidity threshold triggering. This threshold-based regulation lacks in-depth quantification of the air energy state of the controlled space. In actual operation, the control system has difficulty accurately defining whether the current control input is in the effective sensible heat exchange cooling range or in the latent heat exchange pumping range that would cause the controlled medium to lose too much water due to the high water potential difference, resulting in direct economic losses. In addition, excessive moisture loss can also lead to deterioration of processing quality and accelerated physiological aging. To address this, an intelligent temperature and humidity control system for grain storage is proposed. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent temperature and humidity control system for grain storage facilities to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: An intelligent temperature and humidity control system for grain storage includes: Data sensing module: Real-time collection of grain temperature, relative humidity, moisture content and environmental meteorological parameters at different spatial levels in the grain warehouse, and simultaneous calculation of vapor pressure deficit and basic water potential difference on the grain surface to establish an initial spatial vector; The disturbance decoupling module: learns from the initial spatial vector using a deep temporal prediction network to generate a total disturbance state; and uses reinforcement learning to generate feedforward decoupling compensation components based on the total disturbance state to construct a temperature control loop and a humidity control loop. Collaborative control module: Extracts the real-time water potential gradient and spatial air enthalpy value of the grain surface, and uses a multi-objective optimization neural network to perform nonlinear mapping on the real-time water potential gradient and spatial air enthalpy value in combination with the temperature control loop and humidity control loop to generate a dynamic energy state saturation index; calculates the feedback gain of the controlled process in real time based on the dynamic energy state saturation index to generate a global control law command. Signal generation module: Invokes the global control law instruction and the characteristic parameters of the actuator, uses the attention mechanism to dynamically decouple the temperature and humidity control vectors, transforms the control requirements into the power allocation weights of the actuator, and generates the drive signal.

[0006] Preferably, the specific process of simultaneously calculating the vapor pressure deficit and baseline water potential difference on the grain surface includes: The grain temperature is converted into the saturated vapor pressure that can be generated when the air reaches a moisture saturation state at the corresponding temperature. The actual vapor pressure of the current environment is calculated based on the relative humidity. The saturated vapor pressure is subtracted from the actual vapor pressure to generate a vapor pressure deficit value. Extract the grain moisture content and calculate the equilibrium absolute humidity in the grain pile pores based on the current grain temperature. Calculate the basic water potential difference using the logarithmic ratio of the equilibrium absolute humidity to the humidity of the air on the grain surface. Normalize and merge the vapor pressure deficit, basic water potential difference, and original temperature and humidity data to generate an initial spatial vector.

[0007] Preferably, the deep temporal prediction network is a neurodynamic observer based on a cross-attention mechanism, including multi-scale convolutional layers, long short-term memory layers, cross-attention mechanism layers, nonlinear mapping layers, and residual extraction layers; the process of generating the total perturbation state includes: using multi-scale convolution to extract the instantaneous rate of change of the initial spatial vector, and using long short-term memory units to accumulate gradients of water content and water potential difference within historical sampling periods to generate water migration trend features; By matching environmental parameters with the internal state of the grain through a cross-attention mechanism, the infiltration interference weights of the external environment on different layers of the grain pile are calculated. The moisture migration trend characteristics and infiltration interference weights are input into a nonlinear mapping layer to generate an evolution benchmark. The residual is extracted from the instantaneous rate of change using the evolution benchmark to generate the total disturbance state.

[0008] Preferably, the specific process of constructing the temperature control loop and the humidity control loop is as follows: Using the total disturbance state as the environmental state input for reinforcement learning, a search is performed within the preset control parameter optimization range to determine the optimal decoupling strategy that matches the coupling characteristics of the current grain condition; a dynamic decoupling matrix is ​​constructed based on the optimal decoupling strategy, and a feedforward decoupling compensation component is generated in combination with the total disturbance state; the feedforward decoupling compensation component is deconstructed into a temperature compensation term and a humidity compensation term to establish a temperature control loop and a humidity control loop.

[0009] Preferably, the specific process of generating the dynamic energy state saturation index includes extracting the real-time water potential gradient and spatial air enthalpy value of the grain surface, simultaneously acquiring the current adjustment deviation and response rate of the temperature control loop and the humidity control loop, performing feature-level fusion with the real-time water potential gradient and spatial air enthalpy value, calculating the real-time urgency of the current physical feature vector converging to the dynamic equilibrium benchmark determined by the temperature control loop and the humidity control loop through a multi-objective optimization neural network, and performing nonlinear mapping in combination with the remaining adjustment space of the temperature control loop and the humidity control loop to generate the dynamic energy state saturation index.

[0010] Preferably, the multi-objective optimization neural network is a deep neural network based on a multi-gated hybrid expert architecture. The process of generating the dynamic energy state saturation index includes: concatenating the real-time water potential gradient, the spatial air enthalpy value, and the current adjustment deviation and response rate of the temperature control loop and the humidity control loop to construct the physical feature vector; setting up multiple parallel expert sub-networks to share the physical feature vector, extracting grain water retention constraint features and temperature and humidity regulation efficiency features respectively, setting up independent gating networks for convergence urgency prediction and dynamic energy state solution tasks respectively, dynamically calculating the activation weights of each expert sub-network based on the physical feature vector; fusing the expert layer weighted features, and performing nonlinear mapping in combination with the remaining adjustment space of the temperature control loop and the humidity control loop to generate the dynamic energy state saturation index.

[0011] Preferably, the specific process of generating the global control law instruction includes calculating the feedback gain matrix of the controlled process in the current energy state based on the dynamic energy state saturation index using a nonlinear mapping function, wherein the feedback gain matrix includes a temperature gain term and a humidity gain term; Obtain the real-time output limits of the temperature control loop and humidity control loop, and perform adaptive scaling on the feedback gain matrix in combination with the real-time urgency; multiply the calculated feedback gain with the current temperature and humidity adjustment deviation execution matrix, and superimpose the feedforward decoupling compensation component to generate a global control law instruction.

[0012] Preferably, the process of generating drive signals includes obtaining the physical coordinates, rated power, and response frequency of each actuator, and generating feature vectors of each actuator through linear mapping; the actuators include a temperature control unit, a humidity control unit, and a power coupling unit; using the global control law instruction as a query vector, and the feature vectors of each actuator as key vectors and value vectors, a multi-head attention mechanism is used to calculate the matching degree score of each actuator for the current temperature control requirements and humidity control requirements; Based on the matching score, the global control law instruction is subjected to orthogonal projection mapping, decoupled into temperature control vector and humidity control vector pointing to specific actuator nodes; the decoupled actuator control vectors are normalized to generate the power allocation weights of each actuator, and converted into drive signals.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This application elevates the control dimension from simple temperature and humidity readings to the dynamic nature of grain moisture migration by simultaneously calculating the vapor pressure deficit and basic water potential difference on the grain surface. Because the system can monitor the energy potential difference between the grain and the air in real time, it can effectively suppress irrational respiration and water evaporation in the grain. This physical mechanism-based control method avoids the excessive drying or localized condensation phenomena commonly found in traditional control methods, maintaining a stable storage environment while maximizing the preservation of the grain's moisture content, weight, and edible quality.

[0014] 2. This application constructs a neurodynamic observer using a deep temporal prediction network and employs a cross-attention mechanism to capture the penetration patterns of external meteorological parameters into the deep layers of the grain pile. Compared to the lag in traditional feedback control, which adjusts only after fluctuations occur, this module can generate the total disturbance state in advance and construct a feedforward decoupling compensation component by combining reinforcement learning. This prediction-driven control logic enables the system to pre-adjust the control loop before drastic changes in the short to medium term weather, effectively solving the regulation oscillations caused by strong coupling of temperature and humidity, and ensuring the smoothness of grain condition fluctuations under complex environments.

[0015] 3. This application utilizes an attention mechanism via a signal generation module to dynamically match and score the global control law commands with the physical feature vectors of each actuator. Through this orthogonal projection mapping, the system can automatically decouple and allocate optimal power weights based on the real-time performance and physical location of different fans and air conditioners, achieving on-demand allocation and optimal energy efficiency in heterogeneous scheduling. This not only avoids power cancellation and ineffective start-stop between devices, significantly reducing the overall energy consumption of warehouse control, but also significantly mitigates mechanical fatigue and wear of hardware equipment by smoothing power loads. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of a smart temperature and humidity control system for the inside of a grain warehouse. Figure 2 A schematic diagram of the framework structure of an intelligent temperature and humidity control system for the inside of a grain warehouse; Figure 3 This is a schematic diagram of the deep temporal prediction network architecture of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1: Please see Figure 1 and Figure 2 This invention provides an intelligent temperature and humidity control system for the interior of a grain warehouse, the technical solution of which is as follows: An intelligent temperature and humidity control system for grain storage includes: Data sensing module: Real-time collection of grain temperature, relative humidity, moisture content and environmental meteorological parameters at different spatial levels in the grain warehouse, and simultaneous calculation of vapor pressure deficit and basic water potential difference on the grain surface to establish an initial spatial vector; The disturbance decoupling module: learns from the initial spatial vector using a deep temporal prediction network to generate a total disturbance state; and uses reinforcement learning to generate feedforward decoupling compensation components based on the total disturbance state to construct a temperature control loop and a humidity control loop. Collaborative control module: Extracts the real-time water potential gradient and spatial air enthalpy value of the grain surface, and uses a multi-objective optimization neural network to perform nonlinear mapping on the real-time water potential gradient and spatial air enthalpy value in combination with the temperature control loop and humidity control loop to generate a dynamic energy state saturation index; calculates the feedback gain of the controlled process in real time based on the dynamic energy state saturation index to generate a global control law command. Signal generation module: Invokes the global control law instruction and the characteristic parameters of the actuator, uses the attention mechanism to dynamically decouple the temperature and humidity control vectors, transforms the control requirements into the power allocation weights of the actuator, and generates the drive signal.

[0019] The specific process of real-time collection of grain temperature, relative humidity, moisture content, and environmental meteorological parameters at different spatial levels within the grain warehouse includes dividing the grain warehouse vertically into surface, middle, and bottom layers, and horizontally deploying temperature and humidity sensors and moisture sensors at the center and around the perimeter to capture grain temperature, relative humidity, and moisture content at different spatial locations in real time. Simultaneously, environmental meteorological parameters, including outside air temperature, air humidity, and wind speed, are acquired through a meteorological station installed outside the warehouse. The vapor pressure deficit and basic water potential difference on the grain surface are calculated simultaneously to establish an initial spatial vector; the specific process of simultaneously calculating the vapor pressure deficit and basic water potential difference on the grain surface includes: The grain temperature is converted into the saturated vapor pressure that can be generated when the air reaches a moisture saturation state at the corresponding temperature. The actual vapor pressure of the current environment is calculated based on the relative humidity. The saturated vapor pressure is subtracted from the actual vapor pressure to generate a vapor pressure deficit value. Extract the grain moisture content and calculate the equilibrium absolute humidity in the grain pile pores based on the current grain temperature. Calculate the basic water potential difference using the logarithmic ratio of the equilibrium absolute humidity to the humidity of the air on the grain surface. Normalize and merge the vapor pressure deficit, basic water potential difference, and original temperature and humidity data to generate an initial spatial vector.

[0020] In calculating the vapor pressure deficit, the system first acquires the real-time grain temperature data collected by the data sensing module at each layer. Utilizing the principles of water vapor dynamics, the grain temperature at each measuring point is converted into the saturated vapor pressure value that can be generated when the air is fully saturated with moisture under those temperature conditions. Simultaneously, the system extracts the relative humidity parameter for the corresponding layer and calculates the actual vapor pressure value in the current environment based on this relative humidity, specifically by multiplying the saturated vapor pressure by the relative humidity. By subtracting the actual vapor pressure from the saturated vapor pressure, the system generates the vapor pressure deficit value, reflecting the driving force of moisture evaporation from the grain surface. This process elevates the control dimension from simple environmental humidity to the dynamic essence of moisture migration on the grain surface. In the process of calculating the basic water potential difference, the system extracts the real-time monitored grain moisture content and combines it with the current grain temperature. It then uses the grain equilibrium moisture relationship to calculate the equilibrium absolute humidity inside the grain pile pores under the current state. The system retrieves the real-time humidity data of the air in the space above the grain surface, calculates the ratio between the above equilibrium absolute humidity and the humidity of the air in the grain surface, and performs logarithmic transformation on the ratio to obtain the basic water potential difference value. The basic water potential difference value can quantify the gradient of the migration of moisture inside the grain to the space air from the perspective of energy potential energy, thereby defining whether the control input is in the effective sensible heat exchange cooling range or the latent heat exchange range that will lead to excessive water loss, effectively avoiding the unexpected loss of grain moisture caused by traditional threshold regulation. The system aggregates real-time collected raw temperature and humidity data, moisture content data, and environmental meteorological parameters with synchronously calculated vapor pressure deficit and basic water potential difference. To eliminate the impact of dimensional differences between different physical parameters on the subsequent deep prediction network, the system performs normalization and merging processing on these data through the existing mapping, uniformly mapping the values ​​of each dimension to the standard range of 0 to 1; according to the spatial physical layer coordinates inside the grain warehouse, the normalized multidimensional parameters are arranged in a matrix to establish an initial spatial vector; The spatial physical layer coordinates are based on the geometric position index established in the three-dimensional space inside the grain warehouse. They are used to mark the physical nodes of the sensors at different height layers (such as the surface layer, middle layer, and bottom layer) and horizontal orientations of the grain pile. The spatial physical layer coordinate system can correspond the collected multidimensional parameters to specific spatial locations one by one. This invention delves into the dynamics of grain moisture migration by calculating vapor pressure deficit and water potential difference, effectively suppressing moisture evaporation and preserving grain quality. At the same time, it utilizes spatial strata coordinates to achieve precise sensing, avoiding excessive drying or condensation caused by traditional control methods, and significantly improving the economic efficiency of storage and control.

[0021] See Figure 3 The system learns from the initial spatial vector using a deep temporal prediction network to generate a total perturbation state. The deep temporal prediction network is a neurodynamic observer based on a cross-attention mechanism, including multi-scale convolutional layers, long short-term memory layers, cross-attention mechanism layers, nonlinear mapping layers, and residual extraction layers. The process of generating the total perturbation state includes: extracting the instantaneous rate of change of the initial spatial vector using multi-scale convolution, and using long short-term memory units to accumulate gradients of water content and water potential difference within historical sampling periods to generate water migration trend features. By matching environmental parameters with the internal state of the grain through a cross-attention mechanism, the infiltration interference weights of the external environment on different layers of the grain pile are calculated. The moisture migration trend characteristics and infiltration interference weights are input into a nonlinear mapping layer to generate an evolution benchmark. The residual is extracted from the instantaneous rate of change using the evolution benchmark to generate the total disturbance state.

[0022] Specifically, the constructed initial spatial vector is input into the deep temporal prediction network, receiving an initial spatial vector matrix with a dimension of 6×15, where 6 represents the six feature dimensions: vapor pressure deficit, basic water potential difference, temperature, humidity, moisture content, and environmental meteorological parameters, and 15 represents the number of spatial layer nodes in the grain silo (including 5 sampling points each for the surface, middle, and bottom layers). Multi-scale convolutional layers perform operations on the input matrix, using three sets of parallel one-dimensional convolutional kernels with kernel sizes of 3, 5, and 7, and a stride of 1. Same padding is used to maintain consistent sequence length. Each set of convolutional kernels contains 32 kernels. The convolution operation extracts instantaneous rate of change features at different time scales along the feature dimension. Finally, the three outputs are concatenated along the channel dimension to form a feature map with a dimension of 96×15, and the ReLU activation function is used for non-linear processing. The output of the convolutional layer is flattened along the spatial dimension and then input into the long short-term memory layer. A bidirectional LSTM structure is adopted, with the number of hidden units set to 128. Gradient accumulation is specifically performed on the water content sequence and water potential difference sequence within the historical 48 time steps (i.e., the data of the past 12 hours, with a sampling interval of 15 minutes). The unit state update adopts the Tanh function, and the final output is a 256-dimensional water migration trend feature. The cross-attention mechanism layer adopts a 4-head attention mechanism. The system maps environmental meteorological parameters into query vectors of dimension 128 through a fully connected layer, and maps the internal state features of the grain silo into key vectors and value vectors of dimension 128. The system performs scaling dot product operation based on the dimension of the key vector to calculate the penetration interference weight of the external environment on different layers, reflecting the impact of external meteorology on the deep layers of the grain pile. The nonlinear mapping layer receives a concatenated vector of water migration trend features and infiltration disturbance weights. This layer consists of three fully connected neural networks with node numbers set to 256-128-64 respectively, and uses a LeakyReLU activation function with a negative slope of 0.2. Through complex nonlinear mapping, it outputs a 6-dimensional evolution baseline prediction value, representing the expected change trend of each physical dimension under the current disturbance. The residual extraction layer performs calculations, subtracting the evolution baseline prediction value from the instantaneous change rate extracted by multi-scale convolution, element-wise, to obtain a 6-dimensional residual vector, i.e., the total disturbance state. Historical operational data from real grain depots were used as the training set, covering a time span of 12 months and including samples from different seasons and weather conditions. The label data was the actual grain condition change value for the next hour. The loss function used was mean squared error loss, the optimizer was Adam, the initial learning rate was set to 0.001, and it decayed to 0.5 times the original rate every 50 epochs, with a total of 300 training epochs. The batch size was 32, and an early stopping strategy was adopted during training, stopping training when the validation set loss did not decrease for 20 consecutive epochs.

[0023] This invention constructs a neurodynamics observer to achieve logical decoupling between external environmental infiltration interference and the internal autonomous evolution process of grain. By using residual extraction to accurately lock the net interference signal, the system can generate the total disturbance state in advance before drastic short- to medium-term weather changes, transforming passive regulation into prediction-driven regulation. This effectively solves the regulation oscillation caused by strong coupling of temperature and humidity, and significantly enhances the smoothness and stability of grain condition regulation under complex environments.

[0024] Based on the total disturbance state, a feedforward decoupling compensation component is generated using reinforcement learning to construct a temperature control loop and a humidity control loop; the specific process of constructing the temperature control loop and humidity control loop is as follows: Using the total disturbance state as the environmental state input for reinforcement learning, a search is performed within the preset control parameter optimization range to determine the optimal decoupling strategy that matches the coupling characteristics of the current grain condition; a dynamic decoupling matrix is ​​constructed based on the optimal decoupling strategy, and a feedforward decoupling compensation component is generated in combination with the total disturbance state; the feedforward decoupling compensation component is deconstructed into a temperature compensation term and a humidity compensation term to establish a temperature control loop and a humidity control loop.

[0025] The disturbance decoupling module uses the generated total disturbance state as the environmental state input for reinforcement learning. The reinforcement learning deep deterministic policy gradient algorithm dynamically adjusts the weight factors in the decoupling matrix by performing a large number of simulations and trial-and-error iterations within the preset control parameter optimization range. In this process, reinforcement learning takes minimizing the real-time deviation of temperature and humidity control and optimizing system energy consumption as reward objectives. By continuously evaluating the system's suppression effect on disturbances under different parameter configurations, it automatically searches for and determines the optimal decoupling strategy that matches the current grain condition coupling characteristics and total disturbance state. The optimal decoupling strategy is used to construct a 2×2 dynamic decoupling matrix. The rows of the dynamic decoupling matrix represent the compensation output dimension, with the first row being the temperature compensation output row and the second row being the humidity compensation output row. The columns represent the interference input dimension, with the first column being the heat interference input column and the second column being the moisture interference input column. The elements in the matrix represent the cross-influence factors between temperature and humidity variables, which are used to quantitatively describe the influence of temperature regulation on the humidity field and the influence of humidity regulation on the temperature field. The total disturbance state is input into the dynamic decoupling matrix and multiplication is performed. Each disturbance component is multiplied and accumulated with its corresponding decoupling coefficient to generate a feedforward decoupling compensation component. The feedforward decoupling compensation component, as a kind of advanced prediction signal, can dynamically balance the coupling effect between variables before the deviation occurs, and logically separates the nonlinear relationship between temperature and humidity. The feedforward decoupling compensation component is further decomposed into temperature compensation term and humidity compensation term, and superimposed on independent temperature control loop and humidity control loop respectively. Through this loop construction method, the system realizes logical offsetting of control commands at the execution level, so that humidity fluctuations caused by temperature adjustment actions can be offset in real time.

[0026] The feedforward decoupling compensation component is further deconstructed into temperature compensation and humidity compensation terms. This process involves mapping the total disturbance vector to the control compensation space through matrix operations, and extracting the correction increments for the temperature control loop and the humidity control loop based on the row vector distribution of the decoupling matrix. At the logic execution level, the system superimposes the calculated temperature compensation component with the original temperature adjustment command to counteract the cross-interference caused by moisture changes in the thermal field. Simultaneously, the humidity compensation component is superimposed on the humidity adjustment command to balance the impact of temperature fluctuations on the moisture balance within the chamber. This deconstruction process restores the composite compensation signal to specific actuator actions, ensuring that the temperature and humidity control loops can mutually cancel each other out in terms of physical output.

[0027] The feedforward compensation mechanism enables logical hedging of control commands, eliminating nonlinear coupling interference between temperature and humidity before deviations occur, thus avoiding adjustment oscillations caused by frequent fluctuations in the actuator. At the same time, the introduction of reinforcement learning enables the system to have adaptive optimization capabilities, dynamically locking the optimal control path for different grain conditions, effectively reducing the system's operating energy consumption while ensuring a highly stable environment inside the warehouse.

[0028] The real-time water potential gradient and enthalpy of the air on the grain surface are extracted. A multi-objective optimization neural network is then used to perform a nonlinear mapping of the real-time water potential gradient and enthalpy of the air, combining the temperature control loop and the humidity control loop, to generate a dynamic energy state saturation index. The specific process of generating the dynamic energy state saturation index includes extracting the real-time water potential gradient and enthalpy of the grain surface, simultaneously acquiring the current adjustment deviation and response rate of the temperature control loop and the humidity control loop, performing feature-level fusion with the real-time water potential gradient and enthalpy of the air, calculating the real-time urgency of the current physical feature vector converging to the dynamic equilibrium benchmark determined by the temperature control loop and the humidity control loop through the multi-objective optimization neural network, and performing a nonlinear mapping based on the remaining adjustment space of the temperature control loop and the humidity control loop to generate the dynamic energy state saturation index.

[0029] In the process of extracting the real-time water potential gradient on the grain surface, the system first obtains the real-time moisture content and temperature of the grain particles through high-precision moisture and temperature sensors preset inside the grain pile. It then calculates the equilibrium absolute humidity in the pores of the grain pile using the grain moisture balance isotherm equation. Simultaneously, it obtains the actual vapor pressure of the grain surface boundary layer using a spatial humidity sensor on the grain surface. The difference between the saturated vapor pressure converted from the equilibrium absolute humidity and the actual vapor pressure on the grain surface is calculated and divided by the characteristic radius of the grain particles. This allows for the quantitative calculation of the dynamic value of moisture migration from the inside of the grain particles to space, i.e., the real-time water potential gradient. During the acquisition of the enthalpy of the air in the storage space, the system utilizes a cluster of temperature and humidity sensors deployed at different heights within the storage facility to collect real-time data on the dry-bulb temperature and relative humidity at each point. The system calculates the sensible heat of the air based on the dry-bulb temperature, and combines this with the moisture content calculated from the relative humidity and the latent heat of vaporization of water vapor to obtain the latent heat. The sensible heat and latent heat are then summed to obtain the enthalpy of the air in the storage space, which characterizes the energy state of the air within the storage facility. This enthalpy reflects the heat capacity of the current storage environment and is a core parameter for evaluating energy conversion efficiency. In the nonlinear mapping process of feature-level fusion and multi-objective optimization neural network, the system aggregates real-time water potential gradient, spatial air enthalpy, and the current adjustment deviation and response rate of the temperature and humidity control loop to the data input layer. The multi-objective optimization neural network utilizes the nonlinear activation function of the hidden layer to model the correlation between the physical driving force of moisture migration and spatial heat energy distribution, analyzing the degree to which the current environmental parameters deviate from the target equilibrium point. By calculating the Euclidean distance of the feature vector to the equilibrium reference point in the state space, the real-time urgency of the current grain condition adjustment is obtained, which is used to measure the priority of the system in restoring steady state. In the final stage of generating the dynamic energy state saturation index, the system obtains the real-time current, frequency, and valve opening of the actuator through sensor feedback, calculates the ratio of the actual output power of the current temperature and humidity control equipment to the maximum rated power, and thus determines the remaining adjustment space of each loop; the multi-objective optimization neural network takes the real-time urgency as the input weight and the remaining adjustment space as the constraint boundary, performs normalized mapping calculation, and finally outputs a value between 0 and 1, which is the dynamic energy state saturation index; By integrating water potential gradient and air enthalpy in real time, the evolution trend of grain conditions can be predicted from the perspective of physical mechanisms, making the regulation commands more in line with the actual needs of biological respiration and energy conversion. At the same time, by using a multi-objective optimization neural network to dynamically assess the urgency of regulation and the remaining regulation space, the priority of temperature and humidity control can be intelligently balanced.

[0030] The multi-objective optimization neural network is a deep neural network based on a multi-gated hybrid expert architecture. The process of generating the dynamic energy state saturation index includes: concatenating the real-time water potential gradient, spatial air enthalpy, and the current adjustment deviation and response rate of the temperature control loop and humidity control loop to construct the physical feature vector; setting up multiple parallel expert sub-networks to share the physical feature vector, extracting grain water retention constraint features and temperature and humidity regulation efficiency features respectively, setting up independent gating networks for convergence urgency prediction and dynamic energy state calculation tasks respectively, dynamically calculating the activation weights of each expert sub-network based on the physical feature vector; fusing the weighted features of the expert layer, and performing nonlinear mapping in combination with the remaining adjustment space of the temperature control loop and humidity control loop to generate the dynamic energy state saturation index.

[0031] The underlying parameters fed back from the sensor array are retrieved in parallel via the data bus. Six core parameters—real-time water potential gradient on the grain surface, enthalpy of the surrounding air, temperature control loop adjustment deviation, humidity control loop adjustment deviation, temperature response rate, and humidity response rate—are aligned and normalized. The processed parameters are then concatenated into vectors according to a predetermined 1- to 6-bit order to construct a 6-dimensional physical feature vector, which is then input in real-time into a multi-objective optimization neural network as the global driving source for the entire expert architecture. In the expert network computation layer, four parallel expert sub-networks are set up, each consisting of a sub-model containing three layers of fully connected neurons. Among them, expert sub-networks 1 and 2 are trained on large-scale historical grain storage data and are specifically responsible for extracting the grain water retention constraint features, that is, analyzing the correlation between the grain water loss rate and temperature and humidity fluctuations under the current water potential gradient; expert sub-network 3 is responsible for calculating the energy weight required to change the air temperature in the warehouse by 1°C, while expert sub-network 4 simultaneously calculates the energy weight required to change the relative humidity of the air by 1%. Simultaneously, two independent gating networks are constructed. The first gating network dynamically calculates the contribution rate of the four expert subnetworks to the convergence urgency task based on the physical feature vector, generating a set of 4-dimensional probability weight distributions. The second gating network generates another set of 4-dimensional weight distributions for the dynamic energy state calculation task. For example, when the grain temperature deviation is detected to exceed the threshold of 2 degrees Celsius, the gating network increases the weight of expert #3, which is responsible for temperature regulation efficiency, to 0.75. When moisture fluctuations trigger an alarm, the weight of expert #4 is increased in real time. A multi-task learning framework is adopted, with loss functions for the two tasks: MSE (predicted convergence time, actual convergence time) and MSE (predicted energy efficiency ratio, actual energy efficiency ratio), respectively. The Adam optimizer is used with a learning rate of 0.001, a batch size of 64, and 500 training epochs. The outputs of each expert subnetwork are multiplied element-wise and summed with their corresponding gating weights to obtain the fused feature quantity. The real-time operating current of the temperature control device and the humidity control device is extracted, and the difference between the current power and the rated power is calculated to obtain the remaining adjustment space value between 0 and 100. The multi-objective optimization neural network performs a nonlinear mapping between the fused feature quantity and this space value, and finally outputs a dynamic energy state saturation index that fluctuates in the range of 0 to 1.

[0032] By precisely locating the energy weight required for unit temperature and humidity regulation through an expert sub-network, a deep alignment between control logic and physical mechanism is achieved; the dual-gating mechanism can dynamically switch the adjustment center of gravity according to grain temperature deviation or moisture risk, ensuring immediate response under high-risk conditions; at the same time, the index deeply integrates the remaining adjustment space of the equipment, effectively preventing command overshoot and energy waste.

[0033] The feedback gain of the controlled process is calculated in real time based on the dynamic energy state saturation index to generate a global control law instruction. The specific process of generating the global control law instruction includes calculating the feedback gain matrix of the controlled process in the current energy state based on the dynamic energy state saturation index using a nonlinear mapping function. The feedback gain matrix includes a temperature gain term and a humidity gain term. Obtain the real-time output limits of the temperature control loop and humidity control loop, and perform adaptive scaling on the feedback gain matrix in combination with the real-time urgency; multiply the calculated feedback gain with the current temperature and humidity adjustment deviation execution matrix, and superimpose the feedforward decoupling compensation component to generate a global control law instruction.

[0034] The dynamic energy state saturation index is processed using a preset inverse proportional nonlinear mapping function. When the dynamic energy state saturation index is in the low range of 0.1 to 0.3, the system sets the temperature gain term in the higher range of 1.5 to 2.0 and the humidity gain term in the range of 1.2 to 1.8 to enhance the system's sensitivity to capturing initial deviations. When the saturation index rises to the equilibrium range of 0.8 to 1.0, the system automatically adjusts the temperature gain term to 0.3 to 0.5 and the humidity gain term to 0.2 to 0.4. The system obtains the real-time output limits of the actuator through sensor feedback and performs adaptive scaling based on the real-time urgency. Specifically, the system monitors the current air conditioning system's output frequency limit as 15Hz to 50Hz and the dehumidifier's rated power duty cycle limit as 20% to 90%. The system extracts the real-time urgency value; when the urgency is in the high range of 70 to 100, the system multiplies the previously calculated feedback gain matrix by an amplification factor of 1.25; if the calculated control quantity exceeds the output limit, it is forcibly anchored to the upper limit of 50Hz or 90%. The quantized feedback gain matrix is ​​multiplied by the temperature and humidity regulation deviation vectors collected by the current sensor to obtain the feedback control reference. The temperature compensation term and humidity compensation term generated by the disturbance decoupling module are retrieved. The feedback control reference and the feedforward decoupling compensation component are merged by algebraic superposition to synthesize the final global control law instruction. The temperature compensation term in the instruction preemptively offsets the temperature fluctuation trend that may be caused by the humidity adjustment action, while the humidity compensation term immediately offsets the interference caused by the temperature adjustment action on the moisture balance in the chamber. This quantitative hedging process ensures that every bit of energy output by the actuator can be accurately applied to the target controlled variable, logically separating the nonlinear relationship between temperature and humidity.

[0035] The global control law instruction and actuator characteristic parameters are invoked, and the temperature and humidity control vectors are dynamically decoupled using an attention mechanism. The control requirements are transformed into power allocation weights for the actuators, and drive signals are generated. The process of generating drive signals includes obtaining the physical coordinates, rated power, and response frequency of each actuator, and generating feature vectors for each actuator through linear mapping; the actuators include a temperature control unit, a humidity control unit, and a power coupling unit; using the global control law instruction as the query vector, and the feature vectors of each actuator as the key vector and value vector, a multi-head attention mechanism is used to calculate the matching degree score of each actuator for the current temperature control requirements and humidity control requirements; Based on the matching score, the global control law instruction is orthogonally projected and mapped, decoupled into temperature control vectors and humidity control vectors pointing to specific actuator nodes; the decoupled actuator control vectors are normalized to generate power allocation weights for each actuator, and then converted into drive signals. In the process of constructing the feature vectors of the actuators, the physical coordinates of each actuator in the grain silo are obtained. The temperature control unit includes an industrial air conditioning unit, an air source heat pump, and a cooler; the humidity control unit includes a rotary dehumidifier, an ultrasonic humidifier, and a humidity-regulating material circulation device; the power coupling unit includes a variable frequency axial flow fan, a multi-split distribution valve, and a circulating air duct switching valve. The system obtains the rated power and the maximum supported response frequency by reading the nameplate parameters of each actuator, and uses a linear mapping algorithm to encapsulate the above spatial coordinates, power limit, and response frequency into the feature vectors of each actuator. The global control law command generated by the collaborative control module is used as the query vector, while the feature vectors of each actuator, such as the industrial air conditioning unit, rotary dehumidifier, and variable frequency axial flow fan, are used as the key vector and value vector, respectively. The multi-head attention mechanism calculates the dot product correlation degree between the query vector and each key vector in multiple subspaces to obtain the matching degree score of each actuator for the current temperature control requirement and humidity control requirement. During the execution of orthogonal projection mapping and vector decoupling, the global control law command is spatially mapped based on the matching degree score. By constructing an orthogonal basis, the system projects the composite global control requirements into mutually perpendicular decoupling spaces, thereby logically separating the temperature control vector and humidity control vector pointing to specific actuator nodes such as industrial air conditioning units, rotary dehumidifiers, and variable frequency axial flow fans. This orthogonal projection method ensures that temperature regulation actions and humidity regulation actions no longer interfere with each other at the vector level, realizing the accurate conversion from globally unified commands to local independent device control quantities. The control vectors of each decoupled actuator are normalized. By calculating the proportion of each vector in the total control demand, the power allocation weight of each actuator is generated. Based on this allocation weight and the communication protocol of each actuator, the signal generation module converts the weight value into a voltage signal, current signal, or pulse width modulation signal, which is then sent as the final drive signal to the specific machine in the temperature control unit, humidity control unit, and power coupling unit.

[0036] The signal generation module achieves deep decoupling between control commands and hardware execution characteristics. By utilizing a multi-head attention mechanism and orthogonal projection mapping, the system can accurately decompose abstract global control laws into independent commands that point to specific machines, completely eliminating logical interference when industrial air conditioning units and rotary dehumidifiers operate simultaneously. At the same time, the power distribution weight generated based on rated power and response frequency ensures that actuators such as variable frequency axial flow fans always operate within the optimal energy efficiency range.

[0037] This application significantly improves the collaborative accuracy and energy efficiency of grain warehouse temperature and humidity control by deeply integrating deep temporal prediction, reinforcement learning, and a multi-gated hybrid expert architecture. It breaks through the bottleneck of temperature and humidity interference in traditional control, and achieves physical-level decoupling of control vectors at the execution end by utilizing multi-head attention mechanism and orthogonal projection technology, ensuring logical counterbalancing of the actions of equipment such as industrial air conditioners and dehumidifiers. At the same time, by introducing water potential gradient, air enthalpy, and energy state saturation index, the control decision evolves from simple deviation correction to energy state prediction based on physical mechanisms. While ensuring a high degree of steady-state grain conditions, it effectively reduces equipment fatigue and system operating energy efficiency through dynamic optimization allocation of power to the actuators, achieving a balance between grain storage safety and green preservation.

[0038] Example 2: This embodiment applies an intelligent temperature and humidity control system for grain storage to a 5,000-ton large-span flat warehouse for summer temperature control operations. Under summer conditions of 38°C external temperature and 75% relative humidity, the data sensing module collects data every 15 minutes. Through sensors deployed deep within the grain pile (6 meters), the system obtains the real-time temperature of the rice at 22.5°C and the moisture content of the grain at 14.1%. The system calculates the saturated vapor pressure at this temperature to be 2.72 kPa, and the actual vapor pressure to be 2.04 kPa, thus calculating the vapor pressure deficit (VPD) to be 0.68 kPa. By balancing the logarithmic ratio of absolute humidity to the humidity of the air above the grain, the basic water potential difference is calculated to be -12.5 J / kg. Using a neurodynamic observer based on a spatiotemporal attention mechanism, historical sampling period data from the past 72 hours were retrieved for time-series modeling. The multi-scale convolutional layer extracted the instantaneous change rate of the initial spatial vector as 0.15% / h. The cross-attention mechanism, by matching the 5 m / s wind speed interference transmitted in real time from the external weather station, calculated the infiltration interference weight of the external environment on the surface layer (0.5 meters) of the grain pile as 0.62. The system predicted that the temperature evolution baseline of the warehouse in the next 2 hours would be 24.2°C, and performed residual extraction accordingly, locking the net value of the total disturbance state as +1.2°C. This prediction-driven logic enables the system to provide early warning of environmental infiltration, solving the regulation oscillation caused by strong temperature and humidity coupling. The generated total disturbance state is used as the input for reinforcement learning. The system searches within the preset control parameter optimization range (weight interval from 0 to 1) and automatically determines the optimal decoupling strategy, constructing a 2×2 dynamic decoupling matrix. The cross-influence factor between temperature and humidity variables is set to 0.35. The calculated feedforward decoupling compensation components include a temperature compensation term of -1.5°C and a humidity compensation term of +5%RH. By superimposing these components onto the independent loop, the system achieves logical offsetting of control commands before deviations occur, avoiding large humidity fluctuations caused by cooling adjustments. The collaborative control module drives a hybrid expert architecture through a 6-dimensional physical feature vector. When the real-time water potential gradient exceeds the 0.5Pa / m warning line, the first gating network increases the activation weight of the expert sub-network responsible for grain water retention constraints to 0.78. At the same time, it obtains the real-time current (12.5A) and rated maximum current (25A) of the actuator, calculates the remaining adjustment space to be 50%, and the system finally generates a dynamic energy saturation index of 0.42. This index deeply integrates the remaining adjustment capacity of the equipment, effectively preventing command overshoot and energy waste. The signal generation module acquires the feature vectors of each actuator (including an industrial air conditioning unit with a rated power of 15kW and a variable frequency fan with a rated frequency of 50Hz). The multi-head attention mechanism calculates that the matching degree score of the air conditioning unit to the current temperature control requirements is 0.88, while the score of the power coupling unit (fan) is only 0.45 due to being in the low efficiency range. Based on the scores, the global control law command (temperature gain 1.6) is orthogonally projected and mapped to generate the power allocation weights of each device (air conditioning: fan = 0.75: 0.25). The system converts this weight into a 4-20mA current signal and sends it to the execution end, ensuring that heterogeneous devices do not interfere with each other when operating simultaneously, and achieving closed-loop control with optimal energy efficiency.

[0039] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent temperature and humidity control system for the interior of a grain warehouse, characterized in that, include: Data sensing module: Real-time collection of grain temperature, relative humidity, moisture content and environmental meteorological parameters at different spatial levels in the grain warehouse, and simultaneous calculation of vapor pressure deficit and basic water potential difference on the grain surface to establish an initial spatial vector; The disturbance decoupling module: learns from the initial spatial vector using a deep temporal prediction network to generate a total disturbance state; and uses reinforcement learning to generate feedforward decoupling compensation components based on the total disturbance state to construct a temperature control loop and a humidity control loop. Collaborative control module: Extracts the real-time water potential gradient and spatial air enthalpy value of the grain surface, and uses a multi-objective optimization neural network to perform nonlinear mapping on the real-time water potential gradient and spatial air enthalpy value in combination with the temperature control loop and humidity control loop to generate a dynamic energy state saturation index; calculates the feedback gain of the controlled process in real time based on the dynamic energy state saturation index to generate a global control law command. Signal generation module: Invokes the global control law instruction and the characteristic parameters of the actuator, uses the attention mechanism to dynamically decouple the temperature and humidity control vectors, transforms the control requirements into the power allocation weights of the actuator, and generates the drive signal.

2. The intelligent temperature and humidity control system for grain storage according to claim 1, characterized in that, The specific process of simultaneously calculating the vapor pressure deficit and baseline water potential difference on the grain surface includes: The grain temperature is converted into the saturated vapor pressure that can be generated when the air reaches a moisture saturation state at the corresponding temperature. The actual vapor pressure of the current environment is calculated based on the relative humidity. The saturated vapor pressure is subtracted from the actual vapor pressure to generate a vapor pressure deficit value. Extract the grain moisture content and calculate the equilibrium absolute humidity in the grain pile pores based on the current grain temperature. Calculate the basic water potential difference using the logarithmic ratio of the equilibrium absolute humidity to the humidity of the air on the grain surface. Normalize and merge the vapor pressure deficit, basic water potential difference, and original temperature and humidity data to generate an initial spatial vector.

3. The intelligent temperature and humidity control system for grain storage according to claim 1, characterized in that, The deep temporal prediction network is a neurodynamic observer based on a cross-attention mechanism, including multi-scale convolutional layers, long short-term memory layers, cross-attention mechanism layers, nonlinear mapping layers, and residual extraction layers; the process of generating the total perturbation state includes: using multi-scale convolution to extract the instantaneous rate of change of the initial spatial vector, and using long short-term memory units to accumulate gradients of water content and water potential difference within the historical sampling period to generate water migration trend features; By matching environmental parameters with the internal state of the grain through a cross-attention mechanism, the infiltration interference weights of the external environment on different layers of the grain pile are calculated. The moisture migration trend characteristics and infiltration interference weights are input into a nonlinear mapping layer to generate an evolution benchmark. The residual is extracted from the instantaneous rate of change using the evolution benchmark to generate the total disturbance state.

4. The intelligent temperature and humidity control system for grain storage according to claim 1, characterized in that, The specific process for constructing the temperature control loop and humidity control loop is as follows: Using the total disturbance state as the environmental state input for reinforcement learning, a search is performed within the preset control parameter optimization range to determine the optimal decoupling strategy that matches the coupling characteristics of the current grain condition; a dynamic decoupling matrix is ​​constructed based on the optimal decoupling strategy, and a feedforward decoupling compensation component is generated in combination with the total disturbance state; the feedforward decoupling compensation component is deconstructed into a temperature compensation term and a humidity compensation term to establish a temperature control loop and a humidity control loop.

5. The intelligent temperature and humidity control system for grain storage according to claim 1, characterized in that, The specific process of generating the dynamic energy state saturation index includes extracting the real-time water potential gradient and spatial air enthalpy value of the grain surface, simultaneously acquiring the current adjustment deviation and response rate of the temperature control loop and the humidity control loop, performing feature-level fusion with the real-time water potential gradient and spatial air enthalpy value, calculating the real-time urgency of the current physical feature vector converging to the dynamic equilibrium benchmark determined by the temperature control loop and the humidity control loop through a multi-objective optimization neural network, and performing nonlinear mapping in combination with the remaining adjustment space of the temperature control loop and the humidity control loop to generate the dynamic energy state saturation index.

6. The intelligent temperature and humidity control system for grain storage according to claim 5, characterized in that, The multi-objective optimization neural network is a deep neural network based on a multi-gated hybrid expert architecture. The process of generating the dynamic energy state saturation index includes: concatenating the real-time water potential gradient, spatial air enthalpy, and the current adjustment deviation and response rate of the temperature control loop and humidity control loop to construct the physical feature vector; setting up multiple parallel expert sub-networks to share the physical feature vector, extracting grain water retention constraint features and temperature and humidity regulation efficiency features respectively, setting up independent gating networks for convergence urgency prediction and dynamic energy state calculation tasks respectively, dynamically calculating the activation weights of each expert sub-network based on the physical feature vector; fusing the weighted features of the expert layer, and performing nonlinear mapping in combination with the remaining adjustment space of the temperature control loop and humidity control loop to generate the dynamic energy state saturation index.

7. The intelligent temperature and humidity control system for grain storage according to claim 5, characterized in that, The specific process of generating the global control law instruction includes calculating the feedback gain matrix of the controlled process in the current energy state based on the dynamic energy state saturation index using a nonlinear mapping function. The feedback gain matrix includes a temperature gain term and a humidity gain term. Obtain the real-time output limits of the temperature control loop and humidity control loop, and perform adaptive scaling on the feedback gain matrix in combination with the real-time urgency; multiply the calculated feedback gain with the current temperature and humidity adjustment deviation execution matrix, and superimpose the feedforward decoupling compensation component to generate a global control law instruction.

8. The intelligent temperature and humidity control system for grain storage according to claim 7, characterized in that, The process of generating drive signals includes obtaining the physical coordinates, rated power, and response frequency of each actuator, and generating feature vectors for each actuator through linear mapping; the actuators include a temperature control unit, a humidity control unit, and a power coupling unit; using the global control law instruction as the query vector, and the feature vectors of each actuator as the key vector and value vector, a multi-head attention mechanism is used to calculate the matching degree score of each actuator for the current temperature control requirements and humidity control requirements; Based on the matching score, the global control law instruction is subjected to orthogonal projection mapping, decoupled into temperature control vector and humidity control vector pointing to specific actuator nodes; the decoupled actuator control vectors are normalized to generate the power allocation weights of each actuator, and converted into drive signals.

Images