Intelligent ventilation control system and method for grain depot based on deep reinforcement learning

Abstract

This application discloses a grain depot intelligent ventilation control system and method based on deep reinforcement learning, belonging to the field of intelligent control technology. The system includes: a state construction module that acquires temperature and humidity data of the grain pile and environmental meteorological data, calculates the humidity change gradient to obtain the moisture hysteresis characteristics of the grain pile, and splices them to generate a state feature vector; an action generation module that generates an initial ventilation action through a deep reinforcement learning model; a safety verification module that calculates the real-time dew point temperature of the airflow convergence surface and infers and predicts the airflow mixing temperature; in response to the predicted airflow mixing temperature being lower than the sum of the real-time dew point temperature and the preset safety margin, the initial ventilation action is reduced in dimensionality or truncated to generate a safe ventilation action; and a control execution module that performs dynamic ventilation adjustment based on the safe ventilation action. This application solves the problems of traditional control hysteresis, neglect of the characteristics of moisture and heat coupling, and the tendency of pure AI models to cause condensation on the grain pile, thereby improving the accuracy and safety of ventilation control.

Description

Technical Field

[0001] This application relates to the field of intelligent control technology, and in particular to an intelligent ventilation control system and method for grain depots based on deep reinforcement learning. Background Technology

[0002] Grain is a living organism that breathes during storage, and as a porous medium, its internal heat transfer and moisture migration processes are quite complex. Proper ventilation is a key means of cooling grain piles, dissipating moisture, and preventing mold and pests.

[0003] In existing technologies, traditional grain depot ventilation control is mostly based on preset temperature and humidity threshold rules for start-up and shutdown. However, this rule-based control has a lag and cannot adapt to complex and ever-changing external weather conditions and nonlinear thermodynamic changes inside the grain pile. In recent years, some research has begun to introduce general artificial intelligence algorithms, such as deep reinforcement learning, to dynamically adjust ventilation strategies. Reinforcement learning can theoretically achieve optimization, but its direct application in the real physical scenario of grain depots still faces technical challenges.

[0004] First, general reinforcement learning relies on continuous trial and error to obtain rewards. However, in a grain storage environment, if the fans are turned on at inappropriate times, such as in high humidity and cold weather, condensation can easily occur due to the convergence of hot and cold air currents, leading to localized mold growth on the grain. Therefore, purely data-driven algorithms lack basic thermodynamic boundary knowledge and cannot guarantee physical safety.

[0005] Secondly, the water absorption and loss process of grains is not a simple reversible reaction, but rather involves a hysteresis loop effect. Under the same ambient temperature and humidity, the equilibrium moisture content of grains differs between the desorption and adsorption phases. Existing algorithms, which make decisions based on the current temperature and humidity, can easily lead to excessive water loss in the pursuit of rapid cooling.

[0006] Third, existing models mostly take cooling rate or minimum energy consumption as the single optimization objective, and lack the ability to coordinate the time series of multiple indicators such as cooling, water retention, condensation prevention and low energy consumption. Summary of the Invention

[0007] To address the aforementioned shortcomings, this application provides a grain depot intelligent ventilation control system and method based on deep reinforcement learning, which solves the problems of condensation safety hazards, excessive moisture loss, and short-sighted decision-making in existing pure data-driven reinforcement learning methods for grain depot ventilation control.

[0008] Firstly, this application provides a grain depot intelligent ventilation control system based on deep reinforcement learning, comprising:

[0009] The state construction module is used to acquire temperature and humidity data and environmental meteorological data of the grain pile, calculate the humidity change gradient based on the temperature and humidity data within a preset time window, and obtain the moisture lag feature of the grain pile; based on the moisture lag feature, the temperature and humidity data and the environmental meteorological data, feature splicing is performed to generate a state feature vector.

[0010] An action generation module is used to generate an initial ventilation action by performing forward propagation calculations using a pre-trained deep reinforcement learning model based on the state feature vector.

[0011] The safety verification module is used to calculate the real-time dew point temperature of the airflow junction based on the temperature and humidity data; to deduce the predicted airflow mixing temperature under the corresponding ventilation state based on the initial ventilation action; in response to the predicted airflow mixing temperature being lower than the sum of the real-time dew point temperature and a preset safety margin, to perform a dimensionality reduction or truncation operation on the initial ventilation action to generate a safe ventilation action; and in response to the predicted airflow mixing temperature being not lower than the sum of the real-time dew point temperature and the preset safety margin, to use the initial ventilation action as the safe ventilation action.

[0012] The control execution module is used to generate equipment control commands based on the safety ventilation action, and control the underlying ventilation equipment to perform dynamic ventilation adjustment.

[0013] Optionally, the step of calculating the humidity change gradient based on temperature and humidity data within a preset time window to obtain the moisture lag characteristics of the grain pile includes:

[0014] Extract the temperature and relative humidity data sequences of each measuring point in the grain pile within a preset time window, calculate the corresponding absolute humidity value for each time node, and generate an absolute humidity time series.

[0015] The absolute humidity time series is differentiated over time to obtain the absolute humidity change rate, which is used as the humidity change gradient.

[0016] The humidity change gradient is compared with preset negative hysteresis thresholds and positive hysteresis thresholds to determine whether the grain pile is in a state of water decomposition and adsorption, water adsorption, or water balance, thus generating a water hysteresis feature.

[0017] Optionally, the determination of whether the grain pile is in a state of water decomposition and adsorption, water adsorption, or water balance, and the generation of water retardation characteristics, includes:

[0018] In response to the humidity change gradient being less than the negative hysteresis threshold, a first feature code characterizing the grain pile being in a state of water decomposition and absorption is generated as a moisture hysteresis feature.

[0019] In response to the humidity change gradient being greater than the positive hysteresis threshold, a second feature code characterizing the grain pile being in a state of moisture adsorption is generated as a moisture hysteresis feature.

[0020] In response to the humidity change gradient being between the negative hysteresis threshold and the positive hysteresis threshold, a third feature code characterizing the grain pile being in a state of moisture balance is generated as a moisture hysteresis feature.

[0021] Optionally, the step of generating a state feature vector by concatenating features based on the moisture hysteresis feature, the temperature and humidity data, and the environmental meteorological data includes:

[0022] Extract the outdoor temperature and outdoor relative humidity data from the environmental meteorological data;

[0023] Using a preset extreme value normalization algorithm, the temperature and humidity data, the outdoor temperature data, and the outdoor relative humidity data are linearly mapped to obtain a continuous state feature vector with uniform dimensions.

[0024] The moisture hysteresis feature is converted into a one-heat coding format to generate a discrete state feature vector.

[0025] The continuous state feature vector and the discrete state feature vector are concatenated and spliced ​​along a preset feature dimension to output one-dimensional tensor data, thus obtaining the state feature vector.

[0026] Optionally, the environmental meteorological data includes at least outdoor temperature data and outdoor relative humidity data; the environmental meteorological data also includes at least one of environmental wind speed data, atmospheric pressure, and precipitation status indicators.

[0027] Optionally, based on the state feature vector, forward propagation calculation is performed using a pre-trained deep reinforcement learning model to generate an initial ventilation action, including:

[0028] The state feature vector is input into the policy network of a pre-trained deep reinforcement learning model;

[0029] The hidden layer of the policy network is used to perform multi-layer nonlinear mapping on the state feature vector to extract the hidden layer state feature tensor that represents the coupling relationship between the grain pile and the environmental meteorological thermodynamics.

[0030] Through the output layer of the policy network, the hidden layer state feature tensor is mapped and output as the target operating frequency of the variable frequency fan and the target opening of the electric damper.

[0031] The target operating frequency and the target opening degree are vector-combined to generate the initial ventilation action.

[0032] Optionally, calculating the real-time dew point temperature at the airflow junction based on the temperature and humidity data includes:

[0033] The physical contact interface between the surface of the grain pile and the air at the top of the warehouse, and the physical contact interface between the air outlet of the ventilation duct and the bottom of the grain pile, are defined as the airflow convergence surface.

[0034] Extract the temperature data and relative humidity data of the airflow convergence surface from the temperature and humidity data;

[0035] Substituting the temperature data and relative humidity data of the intersection surface into the Magnus empirical formula, the saturated water vapor pressure at the airflow intersection surface is calculated.

[0036] The real-time dew point temperature of the airflow confluence surface is obtained by performing an inverse logarithmic solution based on the saturated vapor pressure.

[0037] Optionally, the step of extrapolating the predicted airflow mixing temperature under the corresponding ventilation state based on the initial ventilation action includes:

[0038] The initial ventilation action is analyzed, the target operating frequency of the corresponding variable frequency fan and the target opening degree of the electric damper are extracted, and the current estimated air volume and estimated air velocity are calculated by combining the pre-stored fan performance curves.

[0039] Outdoor temperature data is extracted from the environmental meteorological data as the air inlet boundary temperature, and temperature data at each measuring point of the grain pile is extracted from the temperature and humidity data as the initial thermal field temperature.

[0040] Substitute the estimated inlet air volume, the estimated inlet air velocity, the inlet air boundary temperature, and the initial thermodynamic field temperature into the porous medium convection heat transfer equation.

[0041] The convection heat transfer equation of the porous medium is solved by time-domain integral iteration using a preset time step, and the thermodynamic temperature of the airflow junction at the next time step is output as the predicted airflow mixing temperature.

[0042] Optionally, in response to the predicted airflow mixing temperature being lower than the sum of the real-time dew point temperature and a preset safety margin, the initial ventilation action is reduced or truncated to generate a safe ventilation action, including:

[0043] The difference between the sum of the real-time dew point temperature and the preset safety margin and the predicted airflow mixing temperature is calculated, and the maximum allowable airflow for condensation prevention is calculated using a thermodynamic inversion algorithm.

[0044] Based on the maximum allowable air intake and the pre-stored fan performance curve, the maximum allowable operating frequency of the variable frequency fan and the maximum allowable opening of the electric damper are calculated to construct the safe ventilation action boundary.

[0045] In response to the initial ventilation action's target operating frequency being greater than the maximum permissible operating frequency, or the target opening being greater than the maximum permissible opening, a truncation operation is performed to truncate the exceeding target operating frequency to the maximum permissible operating frequency, and / or truncate the exceeding target opening to the maximum permissible opening; or a dimensionality reduction operation is performed to calculate a global reduction coefficient based on the safe ventilation action boundary, and to proportionally reduce the target operating frequency and target opening of the initial ventilation action based on the global reduction coefficient, until the mapped operating frequency and opening are both within the safe ventilation action boundary;

[0046] The operating frequency and opening degree obtained after the truncation operation or the dimensionality reduction operation are combined into a vector to generate a safe ventilation action.

[0047] Secondly, this application provides a method for intelligent ventilation control in grain depots based on deep reinforcement learning, including:

[0048] Acquire temperature and humidity data and environmental meteorological data of the grain pile, calculate the humidity change gradient based on the temperature and humidity data within a preset time window to obtain the moisture lag characteristic of the grain pile; perform feature splicing based on the moisture lag characteristic, the temperature and humidity data and the environmental meteorological data to generate a state feature vector.

[0049] Based on the state feature vector, forward propagation calculation is performed through a pre-trained deep reinforcement learning model to generate the initial ventilation action;

[0050] Calculate the real-time dew point temperature at the airflow junction based on the temperature and humidity data; deduce the predicted airflow mixing temperature under the corresponding ventilation state based on the initial ventilation action; in response to the predicted airflow mixing temperature being lower than the sum of the real-time dew point temperature and a preset safety margin, perform a dimensionality reduction or truncation operation on the initial ventilation action to generate a safe ventilation action; in response to the predicted airflow mixing temperature being not lower than the sum of the real-time dew point temperature and the preset safety margin, use the initial ventilation action as the safe ventilation action.

[0051] Based on the aforementioned safe ventilation action, a device control command is generated to control the underlying ventilation equipment to perform dynamic ventilation adjustment.

[0052] Compared with the prior art, the beneficial effects of the present invention are:

[0053] This application introduces moisture lag characteristics to identify the dynamic microstate of grain piles in adsorption, desorption, or equilibrium. Combined with a deep reinforcement learning model, the system can autonomously deduce the optimal strategy based on multidimensional temperature, humidity, and meteorological conditions. This solves the problems of traditional grain depot ventilation relying on fixed temperature and humidity thresholds, exhibiting significant control lag, and ignoring the moisture-heat coupling characteristics of grain. It achieves an upgrade from passive response to active predictive regulation, improving the accuracy and foresight of ventilation control.

[0054] By introducing physical rules and constraints, the dew point temperature at the airflow confluence surface is calculated in real time and the mixing temperature is extrapolated. Once a risk is predicted, the anti-condensation boundary is immediately determined through thermodynamic inversion, and the initial action is reduced in dimension or truncated to intervene. This solves the problem of the black box and uncontrollable risk of pure data-driven AI models, which are prone to outputting dangerous actions that may cause local condensation in grain piles. This ensures that the actions of the underlying equipment are within the safe and compliant boundary, preventing the grain from becoming moldy and heated due to condensation.

[0055] This application replaces the traditional fully open and fully closed extensive control by continuously and dynamically optimizing the operating frequency of the variable frequency fan and the opening degree of the electric damper. It effectively avoids the waste of electricity and loss of grain moisture caused by ineffective ventilation and excessive ventilation, and realizes energy saving and efficiency improvement in grain depot management. Attached Figure Description

[0056] Figure 1 A schematic diagram of a grain depot intelligent ventilation control system based on deep reinforcement learning provided in an embodiment of this application;

[0057] Figure 2 A flowchart illustrating the generation of moisture hysteresis features provided in this application embodiment;

[0058] Figure 3 A flowchart for generating safe ventilation actions is provided for embodiments of this application;

[0059] Figure 4 A flowchart of a grain depot intelligent ventilation control method based on deep reinforcement learning provided in an embodiment of this application. Detailed Implementation

[0060] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0061] See Figure 1 The diagram shown is a schematic of a grain depot intelligent ventilation control system based on deep reinforcement learning provided in an embodiment of this application, including:

[0062] The state construction module 10 is used to acquire temperature and humidity data and environmental meteorological data of the grain pile, calculate the humidity change gradient based on the temperature and humidity data within a preset time window, and obtain the moisture lag feature of the grain pile; and perform feature splicing based on the moisture lag feature, the temperature and humidity data and the environmental meteorological data to generate a state feature vector.

[0063] The action generation module 20 is used to generate an initial ventilation action by performing forward propagation calculations through a pre-trained deep reinforcement learning model based on the state feature vector.

[0064] The safety verification module 30 is used to calculate the real-time dew point temperature of the airflow junction based on the temperature and humidity data; to deduce the predicted airflow mixing temperature under the corresponding ventilation state based on the initial ventilation action; in response to the predicted airflow mixing temperature being lower than the sum of the real-time dew point temperature and a preset safety margin, to perform a dimensionality reduction or truncation operation on the initial ventilation action to generate a safe ventilation action; and in response to the predicted airflow mixing temperature being not lower than the sum of the real-time dew point temperature and the preset safety margin, to use the initial ventilation action as the safe ventilation action.

[0065] The control execution module 40 is used to generate equipment control commands based on the safety ventilation action, and control the underlying ventilation equipment to perform dynamic ventilation adjustment.

[0066] Existing grain depot ventilation control technologies are mostly based on simple threshold judgments using current static temperature and humidity values, ignoring the hysteresis loop effect of grain as a porous medium during moisture absorption and desiccation. This results in the system being unable to accurately identify the real micro-dynamic changes in the grain pile, which can easily lead to problems such as excessive moisture loss or ineffective ventilation when making decisions.

[0067] In this embodiment, the state construction module 10 is based on the principles of multi-source data fusion and porous media thermodynamics. It extracts the temperature and humidity change gradient within a historical time window to infer the moisture migration state inside the grain pile, i.e., adsorption or desorption. It then concatenates and splices this discrete micro-dynamic feature with continuous environmental meteorological data after unifying the dimensions and format conversion.

[0068] In practical implementation, the state construction module 10 first periodically acquires temperature and humidity data and environmental meteorological data of the grain pile through an industrial switch and a programmable logic controller, and stores the data in a time-series database. The temperature and humidity data are real-time temperature and relative humidity sequences collected by an array of temperature and humidity sensors deployed inside the grain silo and at different depths within the grain pile. The environmental meteorological data are external environmental parameters collected by a small weather station outside the grain depot, including at least outdoor temperature data and outdoor relative humidity data, and may also include at least one of environmental wind speed data, atmospheric pressure, and precipitation status indicators, used to characterize the air intake boundary conditions for ventilation.

[0069] Next, the temperature and relative humidity data sequences of each measuring point in the grain pile within a preset time window are extracted, and resampling and missing value interpolation alignment are performed according to a fixed time step. The size of the preset time window is set based on the physical characteristics of the grain pile as a porous medium, whose internal thermodynamic response has a significant delay, and can be set to 12 to 24 hours, including at least one complete day-night environmental temperature difference cycle. It should be noted that the size of the preset time window is not fixed. The system can dynamically adapt and select it according to the grain variety, the degree of change in the storage environment, and the control requirements. Specifically, in routine monitoring and stable control scenarios, for large-volume grain piles such as wheat and corn with significant thermodynamic response delays, the time window is preferentially set to 12 to 24 hours to filter out short-term environmental fluctuation noise and capture the long-term moisture migration trend under the complete diurnal temperature cycle. In meteorological change or high-frequency control scenarios, such as sudden cooling and heating, strong winds and rain, to prevent excessive lag in model decision-making, the time window can be switched to a shorter time window, such as 2 to 4 hours. At this time, a high-frequency sampling step size, such as 30 minutes, is used to quickly capture the transient moisture absorption / desorption state response of the surface grain pile under severe environmental stimuli. The fixed time step can be set according to the actual operation frequency of the bottom ventilation equipment and the data acquisition frequency, and can be set to 30 minutes to 1 hour.

[0070] In the specific resampling operation, the system divides the collected raw data with non-uniform timestamps into continuous time grids according to a set fixed time step, and calculates the arithmetic mean of multiple sets of data collected within the same time grid. The arithmetic mean is used as the representative temperature and humidity observation value of that time grid node.

[0071] In the missing value interpolation and alignment operation, for cases where there is no data in a certain time grid due to a brief failure of a local sensor or packet loss in network communication, the system can use a first-order linear interpolation algorithm. That is, by extracting the previous valid observation value before the timestamp of the missing data node and the next valid observation value after it, the system calculates the temperature and humidity estimates of the missing point according to the time interval ratio and fills them into the corresponding grid. This generates a continuous temperature data sequence and a relative humidity data sequence on the time axis with data alignment between different measurement points.

[0072] See Figure 2 The flowchart shown is a process for generating moisture hysteresis features according to an embodiment of this application, including steps S201 to S203, wherein:

[0073] S201, extract the temperature data sequence and relative humidity data sequence of each measuring point in the grain pile within the preset time window, calculate the corresponding absolute humidity value for each time node, and generate an absolute humidity time series;

[0074] S202, Perform time derivative on the absolute humidity time series to obtain the absolute humidity change rate, which is used as the humidity change gradient;

[0075] S203, compare the humidity change gradient with the preset negative hysteresis threshold and positive hysteresis threshold to determine whether the grain pile is in a state of water decomposition and adsorption, water adsorption or water balance, and generate a water hysteresis feature.

[0076] For the aligned data sequence, calculate the corresponding absolute humidity value for each time node to generate an absolute humidity time series. Specifically, this can be calculated using the Magnos empirical formula combined with the ideal gas law. First, calculate the saturated vapor pressure E1 (in hPa) at the current temperature T. The formula is as follows:

[0077]

[0078] Where e represents the natural constant, and then the actual water vapor pressure E2 is calculated based on the current relative humidity RH, using the following formula:

[0079]

[0080] Finally, the absolute humidity value AH is calculated using the following formula:

[0081]

[0082] Subsequently, a time derivative operation is performed on the absolute humidity time series. Specifically, a first-order backward difference algorithm can be used, that is, the difference between the absolute humidity value at the current time and the absolute humidity value at the previous time is calculated and divided by the time step to obtain the absolute humidity change rate, which is used as the current humidity change gradient.

[0083] Furthermore, the humidity change gradient is compared with preset negative and positive hysteresis thresholds to determine the physical state of moisture migration in the grain pile. The preset negative and positive hysteresis thresholds can be determined based on the specific characteristics of the stored grain, such as wheat, corn, or rice, and the corresponding isothermal adsorption and desorption curve data. These thresholds characterize the physical boundary values ​​for maintaining dynamic moisture balance under normal, minor environmental fluctuations. They are typically set as the minimum absolute humidity change rate critical value that can filter out normal measurement noise from temperature and humidity sensors and minor environmental fluctuations, ensuring that the adsorption or desorption state identified by the system represents the actual physical process of moisture migration, rather than equipment reading drift. For example, the negative and positive hysteresis thresholds can be set to -0.05 g / (m³·h) and 0.05 g / (m³·h), respectively.

[0084] As an optional implementation, the step of determining whether the grain pile is in a state of water decomposition and adsorption, water adsorption, or water balance, and generating a water retardation characteristic, includes:

[0085] In response to the humidity change gradient being less than the negative hysteresis threshold, a first feature code characterizing the grain pile being in a state of water decomposition and absorption is generated as a moisture hysteresis feature.

[0086] In response to the humidity change gradient being greater than the positive hysteresis threshold, a second feature code characterizing the grain pile being in a state of moisture adsorption is generated as a moisture hysteresis feature.

[0087] In response to the humidity change gradient being between the negative hysteresis threshold and the positive hysteresis threshold, a third feature code characterizing the grain pile being in a state of moisture balance is generated as a moisture hysteresis feature.

[0088] Specifically, if the current humidity gradient is less than the negative hysteresis threshold, it indicates that moisture inside the grain pile is evaporating into the air. At this point, a first feature code representing the grain pile's state of water decomposition and adsorption is generated. This first feature code is set as a unique thermal vector with the first bit set to 1 and the remaining bits set to 0, serving as the moisture hysteresis feature. This moisture hysteresis feature is a dynamic indicator reflecting whether the grain is in a state of water decomposition and adsorption, water adsorption, or moisture equilibrium under the current temperature and humidity due to different historical moisture evolution paths. It is used to correct the static estimation error of the equilibrium moisture content of grain in traditional models.

[0089] If the current humidity gradient is greater than the positive hysteresis threshold, it indicates that the grain pile is absorbing moisture from the air. In this case, a second feature code representing the grain pile's moisture adsorption state is generated. This second feature code is set as a unique thermal vector where the second bit is 1 and the rest are 0, serving as the moisture hysteresis feature. If the current humidity gradient is between the negative and positive hysteresis thresholds, including cases where it equals the thresholds, it indicates that the moisture exchange within the grain pile is weak and has reached dynamic equilibrium. In this case, a third feature code representing the grain pile's moisture equilibrium state is generated. This third feature code is set as a unique thermal vector where the third bit is 1 and the rest are 0, serving as the moisture hysteresis feature.

[0090] Among them, because the physical state of moisture migration in the grain pile within a specific time window is divided into three mutually exclusive and discrete micro-dynamic modes of water decomposition and absorption, water adsorption and water balance, the total number of bits of the feature encoding one-hot vector is set to three.

[0091] As an optional implementation, the step of generating a state feature vector by concatenating features based on the moisture hysteresis feature, the temperature and humidity data, and the environmental meteorological data includes:

[0092] Extract the outdoor temperature and outdoor relative humidity data from the environmental meteorological data;

[0093] Using a preset extreme value normalization algorithm, the temperature and humidity data, the outdoor temperature data, and the outdoor relative humidity data are linearly mapped to obtain a continuous state feature vector with uniform dimensions.

[0094] The moisture hysteresis feature is converted into a one-heat coding format to generate a discrete state feature vector.

[0095] The continuous state feature vector and the discrete state feature vector are concatenated and spliced ​​along a preset feature dimension to output one-dimensional tensor data, thus obtaining the state feature vector.

[0096] In practice, outdoor temperature and relative humidity data are extracted from environmental meteorological data and linearly mapped using a pre-defined extreme value normalization algorithm. Specifically, global physical upper and lower limits for each variable are set based on historical meteorological and grain condition extreme values ​​for the region where the grain depot is located. Historical meteorological and grain condition extreme values ​​refer to the highest and lowest temperatures and humidity recorded in official meteorological records over many years for the location of the grain depot, as well as the highest extreme heating temperature or extreme temperature and humidity data measured inside the grain pile during the actual storage period over the years. The variables refer to multi-point temperature and relative humidity inside the warehouse, outdoor temperature, and outdoor relative humidity.

[0097] When setting the global physical upper and lower limits, the system first retrieves the aforementioned extreme data from the meteorological database and the grain depot's historical operation and maintenance database. Then, it adds a reasonable safety margin to the historical highest extreme value as the global physical upper limit, and subtracts a safety margin from the historical lowest extreme value as the global physical lower limit. This constructs an absolute boundary that completely encompasses all possible operating conditions. For example, the global physical lower limit for outdoor temperature is set to -30℃, and the upper limit is set to 50℃. The global physical lower and upper limits for relative humidity are set to 0 and 100%, respectively.

[0098] For any continuously collected data, the following formula is used for scaling to eliminate the influence of deviations caused by differences in units and absolute numerical magnitudes between different physical quantities such as temperature and humidity:

[0099]

[0100] Where X represents the actual collected data, X min X is the set global physical lower bound. max X is the set global physical limit. norm This is the normalized data.

[0101] The above formula is used to compress and transform continuous physical values ​​such as temperature and humidity data with specific units into a unified dimensionless interval between 0 and 1. This results in a continuous state feature vector with unified dimensions, composed of normalized multi-point temperature inside the warehouse, multi-point relative humidity inside the warehouse, outdoor temperature, and outdoor relative humidity.

[0102] Finally, the continuous state feature vector and the discrete state feature vector representing the water hysteresis feature are concatenated and spliced ​​on the feature channel to output a one-dimensional tensor data that integrates environmental meteorology and micro water dynamics, which can then be input into the state feature vector of the deep reinforcement learning model.

[0103] For example, a preset time window is set to the past two hours, and the sampling time step is one hour. The state construction module 10 obtains from the database that the temperature of a core measuring point at the previous moment was 25℃ and the relative humidity was 65%. At the current moment, the temperature drops to 24℃ and the relative humidity rises to 68%. Based on the aforementioned Magnolia empirical formula and related meteorological equations, the saturated water vapor pressure at the previous moment is calculated to be approximately 31.69 hPa, the actual water vapor pressure is approximately 20.6 hPa, and the absolute humidity is approximately 14.96 g / m³. Similarly, the absolute humidity at the current moment is calculated to be approximately 14.81 g / m³. The humidity change gradient is calculated using backward difference as -0.15 g / (m³·h).

[0104] Assuming the system's preset negative hysteresis threshold is -0.05 and the positive hysteresis threshold is 0.05, since -0.15 is less than -0.05, the state construction module 10 determines that the grain pile is currently in a state of water decomposition and absorption, generating a first feature code with feature values ​​of 1, 0, and 0 as the moisture hysteresis feature. Simultaneously, the state construction module 10 obtains the current outdoor temperature of 15℃ and the outdoor relative humidity of 55%. Using an extreme value normalization algorithm, with the temperature upper and lower limits of -20℃ to 50℃ and the relative humidity upper and lower limits of 0 to 100% as benchmarks, it maps the indoor temperature and humidity and the outdoor temperature and humidity to 0.628, 0.680, 0.500, and 0.550 respectively, forming a continuous state feature vector.

[0105] Finally, the discrete features 1, 0, 0 are concatenated and concatenated with the four values ​​of the continuous features to output the final state feature vector, which includes the values ​​0.628, 0.680, 0.500, 0.550, 1, 0, 0 in sequence, and serves as the input base for subsequent model decisions.

[0106] Thus, by introducing an absolute humidity difference derivative mechanism within a time window and combining it with thermodynamic physical threshold division, this application elevates the system from static temperature and humidity monitoring to capturing the dynamic moisture absorption and desorption micro-processes of grain. Furthermore, by efficiently cascading this physical mechanism with unique thermal encoding and normalized continuous data, it enhances the representation quality of the deep reinforcement learning model in the state space and its ability to predict potential moisture loss risks.

[0107] Existing grain depot ventilation control systems mostly use simple PID control or logic control based on static temperature and humidity thresholds. However, grain piles are porous media with large hysteresis, nonlinearity and strong coupling characteristics. Static rules cannot predict the long-term impact of ventilation actions on the moisture gradient inside the grain pile, often leading to excessive water loss of grain in order to cool down, or wasting electricity due to the fans running at full speed.

[0108] Existing multi-source data-driven grain depot ventilation control technologies often rely on preset static temperature and humidity thresholds for simple feedback control, which faces problems such as poor adaptability to nonlinear meteorological conditions, decision-making lag, and difficulty in simultaneously achieving cooling, moisture retention, and low energy consumption. In this embodiment, the action generation module 20 generates initial ventilation actions based on the state feature vector and performs forward propagation calculations using a pre-trained deep reinforcement learning model. Its core principle lies in abstracting the dynamic ventilation adjustment process of the grain depot into a Markov decision process. The intelligent agent, through interaction with the thermodynamic environment of the grain pile, aims to maximize long-term cumulative rewards. Utilizing the powerful high-dimensional feature extraction and nonlinear mapping capabilities of deep neural networks, it directly outputs continuous control actions from cascaded features including microscopic moisture lag and macroscopic meteorological conditions, achieving multi-objective joint optimization control.

[0109] As an optional implementation, based on the state feature vector, a forward propagation calculation is performed using a pre-trained deep reinforcement learning model to generate an initial ventilation action, including:

[0110] The state feature vector is input into the policy network of a pre-trained deep reinforcement learning model;

[0111] The hidden layer of the policy network is used to perform multi-layer nonlinear mapping on the state feature vector to extract the hidden layer state feature tensor that represents the coupling relationship between the grain pile and the environmental meteorological thermodynamics.

[0112] Through the output layer of the policy network, the hidden layer state feature tensor is mapped and output as the target operating frequency of the variable frequency fan and the target opening of the electric damper.

[0113] The target operating frequency and the target opening degree are vector-combined to generate the initial ventilation action.

[0114] In its implementation, the action generation module 20 is divided into an offline model building and training phase and an online action generation and inference phase. Since deep reinforcement learning does not rely on pre-collected static labeled datasets for supervised partitioning, but rather dynamically generates Markov decision process data through agent-environment interaction, the system first collects historical temperature and humidity sensor data from the grain depot over the past five years, including data from all four seasons (spring, summer, autumn, and winter), continuous meteorological data collected from external weather stations, and actual energy consumption records of fans and dampers. Furthermore, it establishes data filtering and classification standards, classifying the five-year historical data based on different grain varieties such as wheat and corn, which have different porosities and specific heat capacities, as well as different storage seasons and operating conditions. Abnormal outliers are then removed, thereby constructing multiple scenario-based training subsets that match real, varied scenarios. Based on the state construction module 10, the effective raw data after the above classification and screening are resampled and aligned, linearly interpolated for missing values, differentially derived for absolute humidity, and spliced ​​with extreme value normalization and one-hot encoding to construct a Markov state feature dataset, which is used to provide a real initial state space and meteorological boundary driving for the construction and training of the subsequent numerical simulation environment.

[0115] In the model and environment construction phase, a numerical simulation environment reflecting the physical characteristics of a real grain silo is first constructed based on the partial differential equations of convection heat transfer and moisture migration in porous media. This simulation environment receives simulated equipment actions and deduces the three-dimensional field distribution of temperature and humidity in the grain pile for the next time step. Based on this, an Actor-Critic dual-network architecture based on the Proximal Policy Optimization (PPO) algorithm can be established. The Actor policy network is used to generate actions, and the Critic value network is used to evaluate the long-term benefits of the current state. Taking the Actor policy network as an example, its network structure can be set as follows: the number of neurons in the input layer matches the dimensions of the concatenated state feature vector, such as including normalized data from 40 grain pile temperature measurement points, normalized data from 40 humidity measurement points, 4 meteorological normalized data, and a 3D moisture hysteresis unique thermal encoding, totaling 87 input nodes. Next, three fully connected hidden layers are cascaded sequentially. The first hidden layer contains 256 neurons with an internal weight matrix of dimension 256×87. It calculates the weighted sum of features using matrix multiplication and adds a bias term, then filters out negative signals using the ReLU activation function, achieving a nonlinear mapping from low-dimensional physical features to a high-dimensional feature space. The second hidden layer contains 128 neurons with a weight matrix of dimension 128×256, and also uses the ReLU activation function for deep feature cross-fertilization and abstraction. The third hidden layer contains 64 neurons with a weight matrix of dimension 64×128, used to extract high-order decision feature tensors representing the coupling relationship between hydrothermal dynamics. Finally, an output layer is connected, containing two neurons corresponding to the control actions of the fan operating frequency and the electric damper opening, respectively. To confine the output actions within defined mathematical boundaries, the output layer uses the Tanh hyperbolic tangent activation function to smoothly truncate the output values ​​of the two-dimensional continuous actions within a continuous floating-point range from -1 to 1.

[0116] The Critic value network, trained synchronously with the Actor policy network, employs the exact same hidden layer structure, but its output layer is reduced to a single linear neuron used to predict the global long-term reward value of the current state. The control time step for the environmental simulation is set to 1 hour. At the beginning of each training round, the system randomly extracts a segment of historical time-series data from the Markov state feature dataset. Using the real meteorological data and grain condition status in this data segment, the system initializes the meteorological boundary conditions and initial grain temperature and humidity field of the virtual simulation environment to simulate the operating conditions of the grain depot under different seasons and regions.

[0117] At each time step, the current 87-dimensional state vector is input into the Actor policy network. The network's forward propagation outputs the mean and standard deviation vectors of the action distribution, thus constructing a multidimensional Gaussian distribution. The system samples from this Gaussian distribution, adding action noise conforming to the standard deviation to the output mean, generating a two-dimensional action vector that is actually executed. Subsequently, the value of this two-dimensional action vector is substituted into the probability density function of the Gaussian distribution to calculate its probability value and take the logarithm, thereby synchronously generating the action log probability corresponding to the action. Finally, this two-dimensional action vector is applied to the virtual simulation environment.

[0118] After receiving the two-dimensional action vector, the simulation environment solves the internal thermodynamic physical equations, deduces and outputs the new state vector for the next hour, and calculates an instant reward value to evaluate the quality of the ventilation action and feeds it back to the Actor policy network.

[0119] To guide the model to learn an optimal multi-objective strategy that balances rapid cooling, safe water retention, and low-carbon energy conservation, a comprehensive reward function was designed. This comprehensive reward function uses a linear weighted summation method to quantify and integrate various indicators. The weight coefficients of each indicator can be determined based on the physical priority of grain depot safety production. Cooling and water retention, as core safety indicators, are given dominant weights, while energy conservation, as an economic indicator, is given secondary weights.

[0120] Specifically, the target cooling rate is set as a non-linear positive reward to encourage effective cooling of the system, and is assigned a first weighting coefficient, for example, 0.5. Simultaneously, a negative penalty is added due to the grain moisture content deviating from the safe threshold to prevent excessive water loss, and is assigned a second weighting coefficient, for example, 0.4. Finally, an energy consumption penalty is subtracted by multiplying the cube of the fan operating frequency by an electricity price coefficient. This electricity price coefficient is a mapping constant pre-stored in the system database that converts physical energy consumption into economic costs. Its value is set according to the real-time industrial electricity price standard of the grain depot's location. It is used to map the abstract energy loss of equipment operating frequency into an economic cost penalty under a unified dimension, thereby guiding the agent to perform high-intensity ventilation during low-electricity-price periods and reduce energy consumption during off-peak or peak-peak electricity-price periods. By introducing this coefficient, the system can constrain the power consumption of the underlying equipment, and is assigned a third weighting coefficient, for example, 0.1. The above first, second, and third weighting coefficients constitute a normalized weight allocation with a sum of 1.

[0121] The system stores a quintuple of data—current state, action taken, reward obtained, next state, and log probability of the action—into a replay buffer. Once the replay buffer has collected interaction data for 2048 time steps, data sampling stops, and the system enters the network parameter update sub-stage. The system utilizes the Generalized Advantage Estimation (GAE) algorithm, combined with a discount factor of 0.99 and a GAE smoothing parameter of 0.95, to calculate the advantage function value for each action. This value measures the relative advantage of the long-term expected return of the specific ventilation action performed, compared to the average expected return under the current grain condition.

[0122] Subsequently, with a mini-batch size of 256, the 2048 time-step quintuple data collected in the replay buffer are treated as a complete dataset. The Adam optimizer is then used to repeatedly perform forward and backward propagation on this dataset, completing 10 rounds of iterative optimization. After completing the parameter updates for the 10 rounds, the replay buffer is cleared, and the agent re-interacts with the virtual simulation environment based on the updated Actor policy network parameters, entering the next data sampling phase.

[0123] During backpropagation updates, the initial learning rate of the Actor policy network is set to 0.0003, and the initial learning rate of the Critic value network is set to 0.001. An importance sampling ratio is introduced, clipping it to a range of 0.8 to 1.2, i.e., the Clip parameter of the Proximal Policy Optimization (PPO) algorithm is set to 0.2. The iteration process continues until the system monitors that the average cumulative reward value over 50 consecutive training epochs converges to a small fluctuation within an extreme range, such as ±3% to ±5%. At this point, training stops, and all fixed weight matrices and bias parameters in the Actor policy network are extracted and saved as a pre-trained deep reinforcement learning model file.

[0124] During the action generation and inference phase, the edge computing gateway at the grain depot site loads the offline-trained model weight file into memory upon system startup. In real-time operation, the state construction module 10 inputs an 87-dimensional state feature vector, collected on-site and processed with normalization and one-hot encoding, into the Actor policy network of the pre-trained deep reinforcement learning model every control cycle, such as one hour. The data undergoes forward propagation computation within the network, sequentially passing through the three hidden layers for linear multiplication-addition and ReLU activation processing, performing multi-layer nonlinear mapping on the state feature vector to extract the hidden state feature tensor representing the coupling relationship between the grain pile and the environmental meteorological thermodynamics. Subsequently, through the output layer of the Actor policy network, the hidden state feature tensor is mapped and output, with the Tanh function outputting two deterministic floating-point values ​​between -1 and 1. Solving and physical mapping are then performed to convert these values ​​into the target operating frequency of the variable frequency fan and the target opening degree of the electric damper. Specifically, the following formula is used to restore the abstract mathematical actions to the actual engineering instructions of the underlying equipment:

[0125]

[0126] For example, assuming the output one-dimensional tensor is [0.3, -0.2], the pre-stored fan frequency range is 0 to 50Hz, and the damper opening range is 0 to 100%, then the target operating frequency is calculated as (0.3+1) / 2×50=32.5Hz, and the target opening is calculated as (-0.2+1) / 2×100%=40%. The system then combines 32.5Hz and 40% into a vector to generate the initial ventilation action command and performs the next step of thermodynamic boundary verification.

[0127] Thus, this embodiment of the application introduces a deep reinforcement learning algorithm based on the Actor-Critic architecture and combines it with a thermodynamic physical twin environment of porous media for model pre-training. It fuses continuous meteorological features with discrete moisture hysteresis features and inputs them into the Actor policy network. By utilizing the multi-level nonlinear mapping and feature dimensionality reduction mechanism of deep neural networks, it deduces and generates continuous combination actions of fan operating frequency and damper opening. This improves the model's feature extraction capability and the accuracy of autonomous decision-making for the implicit thermodynamic nonlinear coupling relationship in the multi-source heterogeneous state space. It optimizes the global dynamic ventilation adjustment effect of grain depots under the constraints of multiple indicators such as rapid cooling, safe water retention and low carbon energy saving when facing complex and variable meteorological conditions.

[0128] After the state construction module 10 and action generation module 20 complete the perception and intelligent decision-making of the grain condition, the initial ventilation action generated by the deep reinforcement learning model, although optimal in terms of long-term reward expectation, may lack absolute physical boundary common sense in transient thermodynamics, posing a risk of condensation on the grain pile due to the output action. To solve the above problem, this embodiment introduces a safety verification module 30.

[0129] When general artificial intelligence algorithms are applied to real-world physical scenarios in grain depots, inappropriate actions, such as turning on a large-volume fan in high-humidity, cold weather, leading to the convergence of hot and cold air currents, can easily cause condensation, resulting in localized heating and mold growth in the grain. The safety verification module 30 establishes a pre- and interception mechanism based on physical laws. Before the equipment actually performs its actions, it uses thermodynamic formulas to calculate the dew point and modifies the initial ventilation actions to ensure compliance through forward temperature prediction and reverse boundary condition calculation.

[0130] In this way, by introducing physical rules and making thermodynamic predictions and interventions before the actions are issued, this application ensures that the actions of the underlying equipment are within the boundaries of safety and compliance, optimizes the reliability of pure AI decision-making, and enhances the safety defense line of grain depot anti-condensation management.

[0131] As an optional implementation, the calculation of the real-time dew point temperature at the airflow junction based on the temperature and humidity data includes:

[0132] The physical contact interface between the surface of the grain pile and the air at the top of the warehouse, and the physical contact interface between the air outlet of the ventilation duct and the bottom of the grain pile, are defined as the airflow convergence surface.

[0133] Extract the temperature data and relative humidity data of the airflow convergence surface from the temperature and humidity data;

[0134] Substituting the temperature data and relative humidity data of the intersection surface into the Magnus empirical formula, the saturated water vapor pressure at the airflow intersection surface is calculated.

[0135] The real-time dew point temperature of the airflow confluence surface is obtained by performing an inverse logarithmic solution based on the saturated vapor pressure.

[0136] In practical implementation, based on the thermodynamic state equation of moist air, when the water vapor content in the air remains constant while the temperature drops to saturation, the water vapor will condense and precipitate as moisture. The airflow convergence surface refers to the two-dimensional spatial boundary within the grain silo where air masses of different temperatures and humidity levels undergo intense mixing and physical contact. Specifically, it can be defined as the physical contact interface between the surface of the grain pile and the air at the top of the silo, and the physical contact interface between the air outlet of the bottom ventilation duct and the bottom layer of the grain pile.

[0137] The system extracts the convergence surface temperature data T at the intersection of the two airflows from the temperature and humidity data obtained from the state construction module 10. c relative humidity data (RH) at the junction c Subsequently, the temperature data T at the interface was... c Substituting the values ​​into the Magnus empirical formula, the saturated vapor pressure E3 at this interface is calculated as follows:

[0138]

[0139] Where e represents the natural constant; The value represents the saturated vapor pressure reference value at a temperature of 0℃, in hectopascals (hPa); a and b are empirical coefficients in the Magnos empirical formula, where a is a dimensionless empirical coefficient and b is a temperature empirical coefficient, in degrees Celsius (℃).

[0140] Next, the actual water vapor pressure E4 is calculated based on the relative humidity data at the interface, using the following formula:

[0141]

[0142] Finally, by performing an inverse logarithmic calculation based on the actual vapor pressure E4, the real-time dew point temperature T at the airflow junction is obtained. d The formula is:

[0143]

[0144] For example, let's set the constant parameters a in the Magnus empirical formula to 17.27, b to 237.3℃, and E0 to 6.11hPa. Assuming the temperature of the current extraction of the surface airflow at the grain pile is 25℃ and the relative humidity is 65%, we can substitute these parameters into the above formula to calculate the actual water vapor pressure as 30.6hPa. After inverse logarithmic calculation, we can obtain the real-time dew point temperature of the interface as 18℃.

[0145] Furthermore, it is necessary to predict in advance the actual thermodynamic impact of the ventilation strategy generated by the model on the microenvironment of the grain warehouse within a future time step, assuming that the physical equipment does not operate.

[0146] As an optional implementation, the step of extrapolating the predicted airflow mixing temperature under the corresponding ventilation state based on the initial ventilation action includes:

[0147] The initial ventilation action is analyzed, the target operating frequency of the corresponding variable frequency fan and the target opening degree of the electric damper are extracted, and the current estimated air volume and estimated air velocity are calculated by combining the pre-stored fan performance curves.

[0148] Outdoor temperature data is extracted from the environmental meteorological data as the air inlet boundary temperature, and temperature data at each measuring point of the grain pile is extracted from the temperature and humidity data as the initial thermal field temperature.

[0149] Substitute the estimated inlet air volume, the estimated inlet air velocity, the inlet air boundary temperature, and the initial thermodynamic field temperature into the porous medium convection heat transfer equation.

[0150] The convection heat transfer equation of the porous medium is solved by time-domain integral iteration using a preset time step, and the thermodynamic temperature of the airflow junction at the next time step is output as the predicted airflow mixing temperature.

[0151] In practical implementation, the safety verification module 30 utilizes computational fluid dynamics and porous media heat transfer to mathematically solve the energy exchange process between the external cold source introduced by the fan and the initial heat source inside the grain pile within a specific space. Specifically, the system first analyzes the initial ventilation action output by the action generation module 20, extracting the corresponding target operating frequency of the variable frequency fan and the target opening degree of the electric damper. Based on the fan performance curves (pressure-flow curves) and the affine mapping matrix of the rotational speed pre-stored in the system, the estimated air volume and estimated air velocity under the current frequency and opening degree combination are calculated through table lookup interpolation.

[0152] The pressure-flow curve refers to the relationship between air volume and air pressure of the fan at the factory rated speed; the affine mapping matrix of the speed refers to a mathematical transformation matrix constructed based on the fan similarity law. It dynamically transforms the pre-stored rated pressure-flow curve into the actual operating curve at the current target frequency by using the ratio of the target operating frequency to the rated frequency as a first-order scaling factor on the air volume dimension and the square of the ratio as a second-order scaling factor on the air pressure dimension.

[0153] It should be noted that, considering the physical limitations of actual equipment operation, the above calculations based on the fan similarity law have applicable conditions and correction rules. Specifically, when the target operating frequency of the fan is within the normal operating range, such as 25Hz to 50Hz, the air volume and frequency are calculated according to the first-order scaling factor mentioned above. When the target operating frequency is lower, such as below 25Hz, due to decreased motor efficiency and increased internal system losses, the actual output air volume will be lower than the theoretical ratio. In this case, the system will introduce a preset low-frequency correction coefficient to appropriately reduce the calculated air volume. The specific value of the low-frequency correction coefficient can be obtained in advance through fan factory calibration or on-site wind measurement experiments and stored in a data table. For example, when the frequency is set to 20Hz, the low-frequency correction coefficient obtained from the table is 0.9, meaning the actual air volume is 90% of the theoretical calculated value.

[0154] Furthermore, since the closing of the electric damper directly increases the airflow resistance of the duct, the actual airflow decreases non-linearly with increasing resistance. After calculating the theoretical airflow based on the fan frequency, the system also looks up the corresponding flow reduction coefficient in a pre-stored one-dimensional damper opening and flow reduction coefficient lookup table based on the current target opening of the electric damper. The flow reduction coefficient in this lookup table represents the proportion of airflow reduction caused by duct resistance; it is an empirical parameter pre-generated by measuring and fitting the actual airflow reduction rate under different damper openings in a standard grain silo environment. Finally, multiplying the theoretical airflow by this flow reduction coefficient yields the final, actual dynamic airflow.

[0155] When performing table lookup interpolation, the system pre-stores a one-dimensional lookup table of damper opening and flow reduction coefficient. First, based on the standard operating pressure set by the grain depot, the system directly reads the baseline airflow corresponding to the current frequency from the actual operating curve after the aforementioned dynamic affine transformation. Then, using the target opening of the electric damper as an index, the system looks up the corresponding flow reduction coefficient in the lookup table. If the target opening falls between two known opening records in the lookup table, the system can use a simple one-dimensional linear interpolation method, i.e., calculate the accurate flow reduction coefficient by proportionally reducing the opening difference. Next, multiplying the baseline airflow by the flow reduction coefficient yields the estimated airflow under the current frequency and opening combination. Finally, dividing the estimated airflow by the system's preset effective ventilation cross-sectional area of ​​the grain depot yields the corresponding estimated airflow velocity.

[0156] Subsequently, outdoor temperature data is extracted from environmental meteorological data as the inlet air boundary temperature, and temperature data at various measuring points on the grain pile are extracted from temperature and humidity data, such as the overall average temperature of 22℃ and the airflow convergence surface temperature of 25℃, as the initial thermal field temperature. The system substitutes the above-mentioned estimated inlet air volume, estimated inlet air velocity, inlet air boundary temperature, and initial thermal field temperature into the set porous media convection heat transfer equation. Specifically, the porous media convection heat transfer equation refers to the partial differential equation of local or non-local thermal equilibrium caused by the temperature difference between the fluid, i.e., air, and the porous solid skeleton, i.e., grain grains.

[0157] Considering that during forced ventilation in a grain silo, the airflow is mainly driven by the fan along a single air intake direction, such as from bottom to top, traversing the grain pile. The internal lateral heat exchange is much smaller than the longitudinal convection of the main airflow. Therefore, to meet the low computational cost requirement for real-time system simulation, this application simplifies the three-dimensional grain pile space along the main air intake direction into a one-dimensional physical path. For example, the porous medium convection heat transfer equation can adopt a one-dimensional airflow energy conservation formula, specifically in the form:

[0158]

[0159] Among them, T aThe predicted airflow mixing temperature is represented by t; the ventilation time is represented by x; the spatial distance node along the airflow direction is represented by v; and the estimated airflow velocity is represented by T. s The initial thermal field temperature of the grain pile is represented by ε; the preset porosity is represented by ρ. a and c a h represents air density and specific heat capacity at constant pressure, respectively. v This represents the volumetric convective heat transfer coefficient.

[0160] It should be noted that the estimated inlet air velocity v in the porous media convection heat transfer equation is a velocity dynamically calculated by the system based on the estimated inlet air volume divided by the effective ventilation cross-sectional area of ​​the grain silo. This achieves an equivalent substitution of the estimated inlet air volume into the porous media convection heat transfer equation. Furthermore, considering that the numerical integration of the partial differential equation depends on the boundary conditions, the system substitutes the inlet boundary temperature as the first-type boundary condition of the equation at the physical interface of the air inlet, i.e., at the spatial distance node x=0, and combines it with the initial thermodynamic field temperature T as the initial value of time. s These factors together constitute the premise that makes the equation have a unique solution.

[0161] The core thermodynamic parameters of the aforementioned porous media can be determined based on empirical values ​​of the target stored grain variety and standard atmospheric conditions. Considering the significant differences in physical properties among different grain varieties, to ensure the accuracy of temperature extrapolation, the selection of the aforementioned core parameters is specifically based on the following: the preset porosity ε is obtained by consulting a preset grain property parameter table, based on the actual grain variety currently stored in the grain warehouse, such as wheat or corn. This grain property parameter table is a built-in mapping table of common grain varieties and their standard porosity empirical values. The air density ρ... a and specific heat capacity at constant pressure c a The standard physical constants are then taken based on meteorological conditions. The volumetric convective heat transfer coefficient h... v The calculation is based on the current estimated inlet air velocity and the corresponding empirical formula for convective heat transfer for the grain type. This empirical formula is typically represented as a nonlinear correlation of the flow velocity. For example, the specific form of the empirical formula for convective heat transfer can be:

[0162]

[0163] Where v represents the estimated intake air velocity; C h The coefficient of variation represents the heat transfer correlation coefficient, which is mainly determined by the surface roughness and particle shape factor of the corresponding grain type; m represents the flow index, which is usually related to the Reynolds number; the C... h Both m and are empirical constants obtained through previous wind tunnel experiments or literature, corresponding to specific grain varieties such as wheat and corn.

[0164] For example, taking a flat warehouse for storing wheat as an example, the constant value of the preset porosity ε can be set to 0.4; within the range of normal pressure and conventional ventilation temperature, the air density ρ a The value is taken as 1.2 kg / m³, and the specific heat capacity at constant pressure is c. a The value is taken as 1005 J / (kg·K); the volumetric convective heat transfer coefficient h v The estimated inlet air velocity can be used to calculate the heat transfer using the aforementioned empirical formula for convective heat transfer. For example, for wheat with a specific moisture content, a heat transfer correlation coefficient C can be set. h The value is 1520, and the flow index m is 0.59. The system substitutes the above-mentioned quantified thermophysical parameters into the porous medium convection heat transfer equation to complete the subsequent differential iterative calculation.

[0165] Using a preset time step, such as a control cycle of 1 hour and an internal integration step of 1 minute, the system first divides the three-dimensional physical space of the grain pile into multiple equally spaced discrete grid nodes along the direction of the incoming airflow, and transforms the spatial derivative terms in the porous medium convection heat transfer equation into an algebraic difference scheme of the temperature of adjacent grid nodes.

[0166] If the explicit difference method is used, the system directly substitutes the known temperature of each grid in the previous minute into the difference equation and independently calculates the new temperature of each grid in the current minute. If the implicit difference method is used, the system keeps the unknown temperature term on the same side of the equation, constructs a system of multivariate linear algebraic equations that includes the unknown temperature of all grid nodes in the current minute, and obtains the result through matrix inversion to ensure numerical stability under large step size.

[0167] The system uses the inlet air boundary temperature as the boundary condition and the initial thermodynamic field temperature as the initial value. It continuously iterates on the time axis according to the above steps, and finally outputs the thermodynamic temperature when the cold air flows through and arrives at the airflow junction at the end of the next control cycle, i.e., 1 hour later. This temperature is used as the predicted airflow mixing temperature.

[0168] For example, by performing an integral solution of the porous medium convection heat transfer equation 60 times, i.e. once per minute, the system predicts that after 1 hour of blowing 10°C cold air at a flow rate of 5000 cubic meters per hour, the temperature at the surface of the grain pile where the airflow converges will drop from 25°C to 16°C.

[0169] In this way, by introducing the fan performance curve and the integral iteration of convective heat transfer, this application solves the problem of significant lag in traditional ventilation control and improves the foresight of ventilation regulation.

[0170] Furthermore, when a dangerous action is detected during the rehearsal, it is necessary to use minimal computational cost and reasonable engineering logic to pull the dangerous action back into the safe zone, while preserving the original optimization intent of deep reinforcement learning.

[0171] See Figure 3 The flowchart shown is a process for generating a safe ventilation action according to an embodiment of this application, including steps S301 to S304, wherein:

[0172] S301, calculate the difference between the sum of the real-time dew point temperature and the preset safety margin and the predicted airflow mixing temperature, and calculate the maximum allowable air intake for condensation prevention using a thermodynamic inversion algorithm;

[0173] S302, Based on the maximum allowable air intake and the pre-stored fan performance curve, calculate the maximum allowable operating frequency of the variable frequency fan and the maximum allowable opening of the electric damper, and construct the safe ventilation action boundary;

[0174] S303, in response to the target operating frequency of the initial ventilation action being greater than the maximum permissible operating frequency, or the target opening being greater than the maximum permissible opening, a truncation operation is performed to truncate the excess target operating frequency to the maximum permissible operating frequency, and / or truncate the excess target opening to the maximum permissible opening; or a dimensionality reduction operation is performed to calculate a global reduction coefficient based on the safety ventilation action boundary, and to proportionally reduce the target operating frequency and target opening of the initial ventilation action based on the global reduction coefficient until the mapped operating frequency and opening are both within the safety ventilation action boundary;

[0175] S304, the operating frequency and opening degree obtained after the truncation operation or the dimensionality reduction operation are combined by vector to generate a safe ventilation action.

[0176] The preset safety margin refers to the temperature buffer margin set to cope with sensor measurement errors and local airflow unevenness. It can be determined comprehensively based on the characteristics of the grain variety, the current moisture content of the grain, and the spatial location of the airflow convergence surface. Specifically, when setting according to grain variety and moisture content, for grains with low porosity and easy moisture absorption and heat generation, such as high-moisture corn or soybeans, the safety margin should be set higher, for example, 1.5℃ to 2.0℃; for grains with high porosity and stable moisture content, such as dry wheat, the safety margin can be appropriately relaxed, for example, set to 0.5℃ to 1.0℃. When setting according to the type of airflow convergence surface, for the convergence surface between the grain pile surface and the silo roof space, due to the greater influence of external temperature and silo roof radiation and the easy disturbance of airflow, the safety margin needs to be increased by a compensation value on the baseline, such as adding an extra 0.5℃; for the convergence surface of the bottom ventilation duct, since the airflow distribution is relatively uniform, the baseline safety margin can be used, for example, set to 1℃. The safe ventilation action boundary refers to the set of maximum operating parameter extreme values ​​that the bottom ventilation equipment is allowed to reach without causing condensation.

[0177] In practical implementation, the system first compares the predicted airflow mixing temperature with the real-time dew point temperature plus a preset safety margin. The sum of the real-time dew point temperature and the preset safety margin is the anti-condensation safety threshold set by the system. If the predicted airflow mixing temperature is significantly lower than the anti-condensation safety threshold, it is determined that the initial action has a significant risk of condensation. The system immediately initiates the intervention procedure, first calculating the difference between the predicted airflow mixing temperature and the real-time dew point temperature plus the preset safety margin, and then using a thermodynamic inversion algorithm. Specifically, the thermodynamic inversion algorithm refers to the algorithm based on the aforementioned porous medium convection heat transfer equation and time-domain integral iterative solution logic. It can use the binary search method in numerical analysis, taking the anti-condensation safety threshold as the target final value, and continuously taking values ​​by binary search within the allowable physical airflow range of the fan and substituting them into the equation for forward pre-simulation trial and error, until the absolute value of the difference between the predicted airflow mixing temperature and the target temperature is less than the preset accuracy. At this point, the input pre-simulation airflow is the maximum allowable airflow.

[0178] Next, the system performs a reverse lookup based on the maximum allowable air intake and the fan performance curve. The specific reverse lookup calculation process is as follows: First, the maximum allowable operating frequency is calculated. The system sets the electric damper to a minimum resistance state at 100% full opening, at which point the flow reduction factor is 1. The maximum allowable air intake is directly used as the actual required air volume after dynamic affine transformation. Based on the first-order scaling relationship between air volume and frequency in the aforementioned fan similarity law, a reverse division is performed to obtain the maximum allowable operating frequency of the corresponding variable frequency fan. Second, the maximum allowable opening is calculated. The system extracts the target operating frequency of the variable frequency fan originally set in the initial ventilation action, reads the reference air intake corresponding to this frequency, divides the maximum allowable air intake by this reference air intake, and obtains the target flow reduction factor. Then, a reverse one-dimensional linear interpolation lookup is performed in the pre-stored damper opening and flow reduction factor lookup table to obtain the corresponding maximum allowable opening of the electric damper. This constructs a safe ventilation action boundary that includes the upper limit extreme value.

[0179] After determining the boundaries, the system intervenes in the initial ventilation action and provides two configurable implementation sub-strategies. Further, these two implementation sub-strategies are adaptively divided based on the severity of the condensation risk caused by the initial ventilation action. Specifically, if the predicted airflow mixing temperature is lower than the anti-condensation safety threshold, the system further determines the relationship between the predicted airflow mixing temperature and the real-time dew point temperature. If the predicted airflow mixing temperature is not greater than the real-time dew point temperature, it is determined that the predicted state has exceeded the physical phase change threshold, belonging to an emergency high-risk scenario for condensation. At this time, the system triggers a truncation operation. In response to the initial target operating frequency being greater than the maximum permissible operating frequency and the initial target opening being greater than the maximum permissible opening, the system directly cuts off the excess portion, assigning the target operating frequency to be truncated to the maximum permissible operating frequency and the target opening to be truncated to the maximum permissible opening, forcing the underlying equipment to return to the safe ventilation action boundary.

[0180] If the predicted airflow mixing temperature is greater than the real-time dew point temperature, i.e., between the real-time dew point temperature and the anti-condensation safety threshold, the predicted state is determined to have only intruded into the temperature buffer zone provided by the safety margin, belonging to a minor risk scenario where no substantial condensation has occurred. In this case, the system triggers a dimensionality reduction operation to calculate a global reduction coefficient, preserving the original AI intent regarding the ratio of airflow to air velocity. Specifically, the frequency reduction ratio and the opening reduction ratio are calculated separately, and the smaller of the two is taken as the global reduction coefficient. The maximum allowable operating frequency of the variable frequency fan, calculated using the aforementioned reverse lookup, is divided by the target operating frequency of the variable frequency fan in the initial ventilation action to obtain the frequency reduction ratio; similarly, the maximum allowable opening of the electric damper, calculated using the aforementioned reverse lookup, is divided by the target opening of the electric damper in the initial ventilation action to obtain the opening reduction ratio. Subsequently, the system compares the frequency reduction ratio and the opening reduction ratio numerically, and takes the smaller of the two as the global reduction coefficient.

[0181] Based on the global reduction coefficient, the initial action is proportionally scaled down and mapped. This involves multiplying the target operating frequency and the target opening by the global reduction coefficient to calculate the mapped operating frequency, continuing until both are within the safe ventilation action boundary. Finally, the operating frequency and opening obtained after the above truncation or dimensionality reduction operations are combined into a one-dimensional vector to generate the final safe ventilation action. Conversely, if a pre-simulation at another time shows that the predicted airflow mixing temperature is not lower than the anti-condensation safety threshold, the initial ventilation action is directly used as the safe ventilation action.

[0182] In this way, this application optimizes the robustness and application safety of the intelligent grain storage control system by thermodynamically inverting the anti-condensation boundary and performing truncation or dimensionality reduction operations.

[0183] After safety verification, the safety ventilation action command is a software-level digital floating-point signal. These continuously changing control variables need to be smoothly, stably, and without distortion converted into electrical drive signals for the underlying physical devices. In this embodiment, the control execution module 40 generates equipment control commands based on the safety ventilation action, and controls the underlying ventilation equipment to perform dynamic ventilation adjustment. Specifically, based on the industrial fieldbus communication and electrical closed-loop servo tracking principle, the physical quantity values ​​output by the algorithm are mapped into industrial standard protocol messages, which are then sent to the field programmable logic controller or directly to the frequency converter and servo actuator through the edge computing gateway. The pulse width modulation technology and position closed-loop loop of the underlying hardware are used to achieve continuous and precise drive of the mechanical parts.

[0184] The equipment control commands refer to hexadecimal electrical communication messages or standard analog signals that are formed by encoding and converting mathematical action values ​​at the software level. These signals are directly parsed by the underlying industrial drivers and are used to drive the operation of on-site variable frequency fans and damper actuators. The dynamic ventilation adjustment refers to breaking away from the traditional discrete operating states of fully open and fully closed for grain depot fans and dampers. This allows the underlying equipment to perform stepless, smooth proportional adjustments and maintain dynamic states within its physical range based on continuous commands issued in real time from the upper level.

[0185] In a specific implementation, the system first receives the safety ventilation action vector output by the safety verification module 30, and extracts the variable frequency fan operating frequency and the electric damper opening degree. Subsequently, the protocol conversion middleware built into the edge computing gateway maps these decimal floating-point numbers to the internal register values ​​of the target device. For example, the frequency value is multiplied by the computing resolution required by the underlying device and rounded, and the opening percentage is mapped to the absolute position pulse count of the actuator or a 4 to 20mA analog output digital quantity.

[0186] Next, the system assembles the converted values ​​into a device control command message according to a preset industrial communication protocol specification, such as Modbus RTU or Profinet. To ensure a successful establishment of a physical communication link between the edge computing gateway and the underlying ventilation equipment, the system pre-configures underlying communication parameters matching the target device. For example, when communicating via an RS-485 serial bus using the Modbus RTU protocol, the preset standard communication configuration parameters are a baud rate of 9600bps, 8 data bits, 1 stop bit, and no parity check. The device control command message includes: a device address pointing to the target ventilation equipment, a function code identifying the operation to write to the register, a data area for loading the encoded hexadecimal values ​​of the operating frequency and opening degree, and a cyclic redundancy check (CRC) code to ensure communication reliability.

[0187] Then, the command is sent in real time to the underlying frequency converter and electric damper servo controller via field industrial Ethernet or serial communication bus. After the underlying equipment receives the control command and verifies that it is correct, the microprocessor inside the frequency converter dynamically adjusts the pulse width modulation signal of the power switching device to change the frequency of the motor input power supply; at the same time, the electric damper actuator drives the internal servo or stepper motor to rotate, and drives the damper to deflect to the specified opening degree in the data area.

[0188] For example, the safety ventilation action generated after dimensionality reduction, namely the variable frequency fan operating frequency of 24.375Hz and the electric damper opening of 30%, is converted into the integer 2438 by the edge computing gateway according to the resolution requirement of 0.01Hz, and encoded as the hexadecimal number 0x0986. At the same time, the 30% opening is mapped to the absolute value of the position setting of the corresponding damper actuator. The system packages these data into Modbus command messages and sends them down instantly through the bus. After receiving the message, the underlying fan inverter controls the fan impeller to smoothly and without impact adjust from the current speed according to the internally set 10-second acceleration or deceleration ramp time, and finally stabilizes at the physical speed corresponding to 24.375Hz. Simultaneously, the electric damper actuator closes the air duct baffle to the opening position of 30% of the total stroke. The combined action of the two forms a precise dynamic air intake.

[0189] In this way, this application replaces the traditional all-on and all-off coarse control by converting high-dimensional safety optimization actions into smooth underlying electrical drive signals and performing them in real time. This reduces the start-up and shutdown impact and wear of underlying mechanical equipment, effectively avoids the waste of electricity and loss of grain moisture caused by ineffective ventilation and excessive ventilation, and improves the energy efficiency and microclimate control accuracy of grain depot management.

[0190] See Figure 4 The flowchart shown is a process for a grain depot intelligent ventilation control method based on deep reinforcement learning provided in an embodiment of this application, including steps S401 to S404, wherein:

[0191] S401, acquire temperature and humidity data and environmental meteorological data of the grain pile, calculate the humidity change gradient based on the temperature and humidity data within a preset time window, and obtain the moisture lag characteristic of the grain pile; based on the moisture lag characteristic, the temperature and humidity data and the environmental meteorological data, perform feature splicing to generate a state feature vector.

[0192] S402, Based on the state feature vector, forward propagation calculation is performed through a pre-trained deep reinforcement learning model to generate the initial ventilation action;

[0193] S403, calculate the real-time dew point temperature of the airflow junction based on the temperature and humidity data; deduce the predicted airflow mixing temperature under the corresponding ventilation state based on the initial ventilation action; in response to the predicted airflow mixing temperature being lower than the sum of the real-time dew point temperature and a preset safety margin, perform a dimensionality reduction or truncation operation on the initial ventilation action to generate a safe ventilation action; in response to the predicted airflow mixing temperature being not lower than the sum of the real-time dew point temperature and the preset safety margin, use the initial ventilation action as the safe ventilation action;

[0194] S404, Based on the aforementioned safe ventilation action, generate equipment control commands to control the underlying ventilation equipment to perform dynamic ventilation adjustment.

[0195] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A grain depot intelligent ventilation control system based on deep reinforcement learning, characterized in that, include: The state construction module is used to acquire temperature and humidity data and environmental meteorological data of the grain pile, calculate the humidity change gradient based on the temperature and humidity data within a preset time window, and obtain the moisture lag characteristics of the grain pile. Based on the moisture hysteresis feature, the temperature and humidity data, and the environmental meteorological data, feature concatenation is performed to generate a state feature vector; Based on temperature and humidity data within a preset time window, the humidity change gradient is calculated to obtain the moisture lag characteristics of the grain pile, including: Extract the temperature and relative humidity data sequences of each measuring point in the grain pile within a preset time window, calculate the corresponding absolute humidity value for each time node, and generate an absolute humidity time series. The absolute humidity time series is differentiated over time to obtain the absolute humidity change rate, which is used as the humidity change gradient. The humidity change gradient is compared with the preset negative hysteresis threshold and positive hysteresis threshold to determine whether the grain pile is in a state of water decomposition and adsorption, water adsorption or water balance, and a water hysteresis feature is generated. To determine whether a grain pile is in a state of water decomposition and adsorption, water adsorption, or water balance, characteristics of water retardation are generated, including: In response to the humidity change gradient being less than the negative hysteresis threshold, a first feature code characterizing the grain pile being in a state of water decomposition and absorption is generated as a moisture hysteresis feature. In response to the humidity change gradient being greater than the positive hysteresis threshold, a second feature code characterizing the grain pile being in a state of moisture adsorption is generated as a moisture hysteresis feature. In response to the humidity change gradient being between the negative hysteresis threshold and the positive hysteresis threshold, a third feature code characterizing the grain pile being in a moisture balance state is generated as a moisture hysteresis feature. An action generation module is used to generate an initial ventilation action by performing forward propagation calculations using a pre-trained deep reinforcement learning model based on the state feature vector. The safety verification module is used to define the physical contact interface between the surface of the grain pile and the air at the top of the storage room, and the physical contact interface between the ventilation duct outlet and the bottom of the grain pile, as the airflow convergence surface; calculate the real-time dew point temperature of the airflow convergence surface based on the temperature and humidity data; deduce the predicted airflow mixing temperature under the corresponding ventilation state based on the initial ventilation action; in response to the predicted airflow mixing temperature being lower than the sum of the real-time dew point temperature and a preset safety margin, perform a dimensionality reduction or truncation operation on the initial ventilation action to generate a safe ventilation action; in response to the predicted airflow mixing temperature being not lower than the sum of the real-time dew point temperature and the preset safety margin, use the initial ventilation action as the safe ventilation action; The initial ventilation action is reduced in dimension or truncated to generate a safe ventilation action, including: The difference between the sum of the real-time dew point temperature and the preset safety margin and the predicted airflow mixing temperature is calculated, and the maximum allowable airflow for condensation prevention is calculated using a thermodynamic inversion algorithm. Based on the maximum allowable air intake and the pre-stored fan performance curve, the maximum allowable operating frequency of the variable frequency fan and the maximum allowable opening of the electric damper are calculated to construct the safe ventilation action boundary. In response to the initial ventilation action's target operating frequency being greater than the maximum permissible operating frequency, or the target opening being greater than the maximum permissible opening, a truncation operation is performed to truncate the exceeding target operating frequency to the maximum permissible operating frequency, and / or truncate the exceeding target opening to the maximum permissible opening; or a dimensionality reduction operation is performed to calculate a global reduction coefficient based on the safe ventilation action boundary, and to proportionally reduce the target operating frequency and target opening of the initial ventilation action based on the global reduction coefficient, until the mapped operating frequency and opening are both within the safe ventilation action boundary; The operating frequency and opening degree obtained after the truncation operation or the dimensionality reduction operation are combined by vector to generate a safe ventilation action. The control execution module is used to generate equipment control commands based on the safety ventilation action, and control the underlying ventilation equipment to perform dynamic ventilation adjustment.

2. The intelligent ventilation control system for grain depots based on deep reinforcement learning according to claim 1, characterized in that, The step of generating a state feature vector by concatenating features based on the moisture hysteresis feature, the temperature and humidity data, and the environmental meteorological data includes: Extract the outdoor temperature and outdoor relative humidity data from the environmental meteorological data; Using a preset extreme value normalization algorithm, the temperature and humidity data, the outdoor temperature data, and the outdoor relative humidity data are linearly mapped to obtain a continuous state feature vector with uniform dimensions. The moisture hysteresis feature is converted into a one-heat coding format to generate a discrete state feature vector. The continuous state feature vector and the discrete state feature vector are concatenated and spliced ​​along a preset feature dimension to output one-dimensional tensor data, thus obtaining the state feature vector.

3. The intelligent ventilation control system for grain depots based on deep reinforcement learning according to claim 1 or 2, characterized in that, The environmental meteorological data includes at least outdoor temperature data and outdoor relative humidity data; the environmental meteorological data also includes at least one of environmental wind speed data, atmospheric pressure, and precipitation status indicators.

4. The intelligent ventilation control system for grain depots based on deep reinforcement learning according to claim 1, characterized in that, Based on the state feature vector, a forward propagation calculation is performed using a pre-trained deep reinforcement learning model to generate an initial ventilation action, including: The state feature vector is input into the policy network of a pre-trained deep reinforcement learning model; The hidden layer of the policy network is used to perform multi-layer nonlinear mapping on the state feature vector to extract the hidden layer state feature tensor that represents the coupling relationship between the grain pile and the environmental meteorological thermodynamics. Through the output layer of the policy network, the hidden layer state feature tensor is mapped and output as the target operating frequency of the variable frequency fan and the target opening of the electric damper. The target operating frequency and the target opening degree are vector-combined to generate the initial ventilation action.

5. The intelligent ventilation control system for grain depots based on deep reinforcement learning according to claim 1, characterized in that, The calculation of the real-time dew point temperature at the airflow confluence surface based on the temperature and humidity data includes: Extract the temperature data and relative humidity data of the airflow convergence surface from the temperature and humidity data; Substituting the temperature data and relative humidity data of the intersection surface into the Magnus empirical formula, the saturated water vapor pressure at the airflow intersection surface is calculated. The real-time dew point temperature of the airflow confluence surface is obtained by performing an inverse logarithmic solution based on the saturated vapor pressure.

6. The intelligent ventilation control system for grain depots based on deep reinforcement learning according to claim 4, characterized in that, The method of extrapolating the predicted airflow mixing temperature under the corresponding ventilation state based on the initial ventilation action includes: The initial ventilation action is analyzed, the target operating frequency of the corresponding variable frequency fan and the target opening degree of the electric damper are extracted, and the current estimated air volume and estimated air velocity are calculated by combining the pre-stored fan performance curves. Outdoor temperature data is extracted from the environmental meteorological data as the air inlet boundary temperature, and temperature data at each measuring point of the grain pile is extracted from the temperature and humidity data as the initial thermal field temperature. Substitute the estimated inlet air volume, the estimated inlet air velocity, the inlet air boundary temperature, and the initial thermodynamic field temperature into the porous medium convection heat transfer equation. The convection heat transfer equation of the porous medium is solved by time-domain integral iteration using a preset time step, and the thermodynamic temperature of the airflow junction at the next time step is output as the predicted airflow mixing temperature.

7. A method for intelligent ventilation control of grain depots based on deep reinforcement learning, implemented based on the intelligent ventilation control system for grain depots based on deep reinforcement learning as described in any one of claims 1-6, characterized in that, The method includes: Acquire temperature and humidity data and environmental meteorological data of the grain pile, calculate the humidity change gradient based on the temperature and humidity data within a preset time window to obtain the moisture lag characteristic of the grain pile; perform feature splicing based on the moisture lag characteristic, the temperature and humidity data and the environmental meteorological data to generate a state feature vector. Based on the state feature vector, forward propagation calculation is performed through a pre-trained deep reinforcement learning model to generate the initial ventilation action; Calculate the real-time dew point temperature at the airflow junction based on the temperature and humidity data; deduce the predicted airflow mixing temperature under the corresponding ventilation state based on the initial ventilation action; in response to the predicted airflow mixing temperature being lower than the sum of the real-time dew point temperature and a preset safety margin, perform a dimensionality reduction or truncation operation on the initial ventilation action to generate a safe ventilation action; in response to the predicted airflow mixing temperature being not lower than the sum of the real-time dew point temperature and the preset safety margin, use the initial ventilation action as the safe ventilation action. Based on the aforementioned safe ventilation action, a device control command is generated to control the underlying ventilation equipment to perform dynamic ventilation adjustment.

Images