Temperature prediction and regulation method based on earthen wall sunlight greenhouse

Abstract

The invention discloses a temperature prediction, regulation and control method and system based on a cob wall sunlight greenhouse, and relates to the field of agricultural facility environment control. Environmental data such as film transmission irradiation and indoor and outdoor temperature and humidity are collected through a multi-source sensor, and after preprocessing, an energy conservation differential equation set covering a multi-layer structure is established; a Crank-Nicolson format is adopted for discretization, an RK4 method and other numerical strategies are adopted for solving, and a temperature prediction curve is output in combination with parameter sample parallel computing and post-processing; and based on a prediction result and real-time data, multi-dimensional control instructions of film rolling, ventilation and the like are generated through model-free reinforcement learning, and compliance operation is guaranteed in cooperation with a safety supervision mechanism. The method integrates the advantages of a physical model and data driving, solves the problems of low greenhouse temperature regulation and control precision, response lag and the like in northwest cold and arid regions, has prediction robustness, regulation and control intelligence and engineering expandability, and is suitable for large-scale popularization.

Description

Technical Field

[0001] This invention relates to the field of agricultural facility environmental control technology, specifically a method for temperature prediction and control based on earthen wall solar greenhouses. Background Technology

[0002] As a distinctive agricultural facility in the cold and arid regions of northern China, solar greenhouses have become a key component in ensuring winter vegetable production and improving agricultural economic benefits in cold regions due to their core advantages of low cost, low energy consumption, and excellent heat preservation and storage performance. Among them, earthen-walled solar greenhouses utilize the high specific heat properties of thick earthen walls to achieve efficient heat storage and slow heat release. Even in severe winters without additional heating, they can maintain the lowest indoor temperature above 10°C in the early morning, effectively meeting the basic temperature requirements for crop growth. They are widely used in extreme climate regions such as the cold and arid northwest.

[0003] Currently, environmental control in earthen-walled solar greenhouses relies heavily on farmers' manual experience, lacking automated adjustment methods based on precise environmental perception and scientific theoretical models. This results in low temperature control accuracy and significant response lag. To address these issues, we propose a temperature prediction and control method for earthen-walled solar greenhouses. Summary of the Invention

[0004] One objective of this invention is to provide a method for temperature prediction and control based on earthen wall solar greenhouses, comprising the following steps: S1. Collect environmental data of the earthen wall greenhouse through multi-source sensors. The environmental data includes the transmitted radiation of the film, indoor and outdoor temperature and humidity, layer temperature of the wall, soil temperature and humidity, light and wind speed. S2. Preprocess the collected environmental data, including noise reduction, missing value compensation, time alignment and unit conversion, to form standardized observation data; S3. Establish a set of differential equations for energy conservation for the greenhouse covering layer, indoor air, crop canopy, layered earthen wall and ground layer. The set of equations includes solar radiation incident terms, convection heat transfer terms, transpiration latent heat terms, ventilation heat transfer terms, layered earthen wall heat conduction terms and actuator sensible heat terms. S4. The energy conservation differential equations are solved using a numerical solution strategy, and preliminary temperature prediction results are obtained by combining parameter samples with parallel calculations. S5. Post-process the preliminary temperature prediction results and output the final temperature prediction curve. S6. Based on temperature prediction curves and real-time sensor data, multi-dimensional control commands are generated through a model-free reinforcement learning agent and sent to the controller to execute corresponding actions; at the same time, a safety supervision mechanism is used to verify compliance and provide fallback for anomalies in the control process.

[0005] Preferably, in S3, the energy conservation differential equations are: In the formula: The heat capacity equation for air; To allow direct absorption of short-wave components from the air and canopy through the membrane; : Convection in the overlying layer; Long-wave radiation heat transfer can be written as (Nonlinear), in engineering, it is often linearized to ; : Convection heat transfer with the top layer of the wall; Ventilation and heat exchange ; The latent heat carried away by crop transpiration. (J / kg) represents the latent heat of vaporization of water. (kg / s) represents the amount of vaporization.

[0006] Preferably, in S4, the numerical solution strategy includes discretizing the wall heat conduction equation using the Crank–Nicolson scheme and solving it using the Thomas algorithm, and calculating the air energy equation using the explicit RK4 method, achieving a coupled solution of the two through semi-implicit iteration; The heat conduction equation for the i-th layer of the wall is as follows: Using Crank–Nicolson time discretization: in This is a linear interlayer heat conduction operator. Merging all layers yields a tridiagonal linear system: Matrix (A) has a tridiagonal structure (coefficients are given by...). constitute); Air temperature equation writing: The air temperature is advanced using a fourth-order Runge-Kutta (RK4). If a sub-loop is used, RK4 uses the current or latest temperature reading within each air step. ; RK4 step (time) ): in This is the currently available temperature of the top layer of the wall.

[0007] Preferably, in step S5, the post-processing includes calculating the mean curve and confidence band, performing deviation correction through a short-term residual model, extracting decision indicators such as low temperature / high temperature risk probability and threshold crossing time, and outputting the final temperature prediction curve.

[0008] Preferably, the training and deployment of the model-free reinforcement learning agent in S6 includes: (1) State vector: The temperature prediction curve of the future time domain H is fused with the observation data at the current moment and then input into the network after standardization and time coding; (2) Action space: Actions can be described in a continuous or mixed manner; (3) Reward function: Multi-objective weighted reward to simultaneously consider temperature control, energy consumption, equipment wear and safety; (4) Offline training: In a digital twin simulation environment with energy conservation differential equations as the core, training is conducted using a variety of weather scenarios and parameter samples, and experience playback and important sampling are used to improve the learning effect; (5) Online deployment: Read real-time data and prediction curves according to the set period, generate control commands through policy network inference, and issue them after verification by the security supervision mechanism.

[0009] Preferably, the security monitoring mechanism includes: (1) Rule base: contains static temperature and humidity thresholds and dynamic rules based on prediction probability; (2) Real-time anomaly detection and automatic rollback: Detect equipment execution anomalies and decision anomalies, and execute local fallback, record and report or disable online learning according to priority; (3) Redundancy and fault tolerance mechanism: Redundant deployment of key sensors / actuators to support data error correction and cleaning when sensors fail, and backup solution switching when actuators fail; (4) Audit and alarm: Record the decision-making process, safety supervision operation and equipment execution status, and support the automatic generation of multi-level alarms and operation and maintenance orders.

[0010] The present invention has the following beneficial effects: 1. This invention combines physical differential equations based on energy conservation with offline / online parameter estimation, which ensures interpretability and can correct model parameters through historical and real-time data, thereby improving the prediction robustness under abnormal climate or out-of-sample conditions.

[0011] 2. Numerical solution for layered earthen walls: Implicit discretization and tridiagonal solution are adopted to make the one-dimensional unsteady heat conduction calculation of the wall stable and efficient. It is suitable for minute-level updates on edge devices and overcomes the strict limitations of explicit methods on step size.

[0012] 3. Multi-source, multi-scale sensing and online correction: The design covers a multi-depth sensing grid encompassing membranes, indoor spaces, wall layers, and soil, enhancing the reliability of control decisions.

[0013] 4. Collaborative control of model-free reinforcement learning and safety supervision: Offline training is conducted in a simulation environment driven by a physical model; during online deployment, prediction-based forward-looking scenario evaluation and rule constraints are adopted to achieve high performance while ensuring operational safety.

[0014] 5. Engineered Interfaces and Scalable Deployment: The engineering implementation of the edge-cloud layered architecture, MQTT / time-series database, and RESTful API ensures system scalability, ease of operation and maintenance, and facilitates promotion and application under different scales and equipment conditions. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0016] This invention discloses a method for temperature prediction and control based on earthen wall solar greenhouses, specifically including the following steps: S1. Collect environmental data of the earthen wall greenhouse through multi-source sensors. The environmental data includes the transmitted radiation of the film, indoor and outdoor temperature and humidity, layer temperature of the wall, soil temperature and humidity, light and wind speed.

[0017] The earthen-walled greenhouse has thick earthen walls on the back and sides, and is covered with a film on the front and top.

[0018] The multi-source sensors include air temperature and humidity sensors and light sensors arranged inside the greenhouse, layered temperature sensors and soil temperature and humidity sensors arranged along the depth direction of the soil wall, and solar radiation sensors, temperature and humidity sensors and wind speed sensors arranged outside the greenhouse.

[0019] Air temperature and humidity monitoring: Taking a plot of land with dimensions of 60*180m as an example, there are 18 data collection nodes at locations with widths of 10, 30, and 50 meters, and lengths of 20, 50, 80, 110, 140, and 170 meters. The air temperature and humidity are collected to serve as state variables for the air energy equation and as inputs for evapotranspiration estimation.

[0020] Layered temperature monitoring of the wall: 6 probes at a depth of 1.5 m in the rear wall, one point every 0.3 m; the same for the side walls; to obtain the temperature profile inside the earthen wall to characterize the time lag behavior of heat storage and release, and to support the solution of the wall heat conduction equation.

[0021] Soil temperature and humidity monitoring: Sensors were deployed near the crop root zone at depths of 0.1 m, 0.3 m, and 0.5 m, at the same locations as the temperature and humidity sensors. Monitoring soil moisture content and temperature is used for modeling irrigation water and heat input, assessing the contribution of soil evaporation to the energy balance, and making irrigation control decisions.

[0022] Light monitoring: Probes are placed at the top, front, and edges of the greenhouse's central axis; an external meteorological radiation station is placed in an unobstructed location as a reference. The incident shortwave and transmitted light flux through the film are measured to calculate the heat input through light and provide direct feedback for shading film rolling control.

[0023] S2. Preprocess the collected environmental data, including noise reduction, missing value compensation, time alignment and unit conversion, to form standardized observation data.

[0024] S2.1 Data Acquisition and Preprocessing Retrieves the most recent observations (indoor multi-point temperature and humidity, wall layer temperature, soil moisture, light intensity, actuator status) from the time-series database and aligns them with timestamps to form an observation vector. ; Check for missing / outlier points and fill them in using interpolation or nearby redundant sensors; generate an initial state estimate. (See below) and parameter priors Data is retrieved via edge gateway → MQTT → time series database → prediction service; data cleaning is performed through threshold verification, sliding median filtering (3-5 sampling points in a window), and mutation detection.

[0025] S2.2 Handling of Parameter Uncertainty and Sample Preparation Draw (N) parameter samples from the parameter posterior or estimation error. Run parallel solutions for each sample. Use offline MCMC, historical least squares fitting residual estimation, or online UKF covariance output.

[0026] S3. Establish a set of differential equations for energy conservation for the greenhouse covering layer, indoor air, crop canopy, layered earthen wall and ground layer. The set of equations includes solar radiation incident terms, convection heat transfer terms, transpiration latent heat terms, ventilation heat transfer terms, layered earthen wall heat conduction terms and actuator sensible heat terms. (1) Indoor air energy equation: In the formula: The heat capacity equation for air; To allow direct absorption of short-wave components from the air and canopy through the membrane; : Convection in the overlying layer; Long-wave radiation heat transfer can be written as (Nonlinear), in engineering, it is often linearized to ; : Convection heat transfer with the top layer of the wall; Ventilation and heat exchange ; The latent heat carried away by crop transpiration. (J / kg) represents the latent heat of vaporization of water. (kg / s) represents the amount of vaporization.

[0027] (2) Heat conduction equation of wall layer For the i-th layer of the wall The equation is: In the formula: : is the equivalent heat capacity of the i-th layer (volume heat capacity × volume); : is the thermal conductivity operator between two layers or at the interface. For layer thickness; For the outermost layer (i=1), the boundary flux term with air / film needs to be added, for example, by... Included as ; For the bottom layer (i=n), the bottom boundary can be a Dirichlet constant for ground temperature. ).

[0028] (3) Crop canopy transpiration, humidity (air moisture content) mass balance Transpiration affects energy, which is expressed by the law of conservation of water vapor mass: in Derived from the porosity / illuminance model; this equation is coupled with the energy equation: latent heat term It appears in the air energy equation.

[0029] S4. The energy conservation differential equations are solved using a numerical solution strategy, and preliminary temperature prediction results are obtained by combining parameter samples with parallel calculations.

[0030] S4.1 Differentiated Numerical Discretization and Solution Implicit wall solution: Crank–Nicolson + Thomas algorithm Discretization of wall heat conduction: Heat conduction equation for the i-th layer of the wall: Using Crank–Nicolson time discretization: in This is a linear interlayer heat conduction operator. Merging all layers yields a tridiagonal linear system: Matrix (A) has a tridiagonal structure (coefficients are given by...). constitute).

[0031] Thomas' algorithm (O(n)) Using the Thomas algorithm on the tridiagonal A allows for efficient solution on edge devices. Boundary terms (coupled with air convection / radiation, ground boundaries) are included in the right-hand term B.

[0032] Explicit solution for air: RK4 Air temperature equation writing: The air temperature is advanced using a fourth-order Runge-Kutta (RK4). If a sub-loop is used, RK4 uses the current or latest temperature reading within each air step. RK4 step (time) ): in The currently available top wall temperature 3. Coupled Iteration For each wall step (step length) ), with known Solve for the current value of the wall CN to obtain .

[0033] Within the wall step, there are several air substeps. Update each substep using RK4 Using linear interpolation or constant As a coupling term.

[0034] S5. Post-process the preliminary temperature prediction results and output the final temperature prediction curve. 1. Parallelization of parameter samples To obtain the confidence band, parameter samples are solved in parallel: Parameter sample source: Offline: MCMC posterior sampling (if Bayesian calibration has been performed); Online: UKF-based covariance pairs Wait for random sampling; Parameter sample set .

[0035] Parallel solution process Perform a complete numerical prediction in parallel for each parameter sample j (within S4 steps) to obtain the trajectory. .

[0036] Aggregation and Output For time (t_k), calculate the sample mean. with sample standard deviation And quantiles (e.g., 2.5% / 97.5%) yield a 95% confidence band.

[0037] Calculate risk indicators, such as: temperature below a threshold in the next 3 hours. probability 2. Output format and data flow The final output (example JSON) includes: Time series: mean and confidence band at each time point; Events: Threshold crossover time, minimum / maximum predicted temperature; Risk indicators: probability of low temperature, probability of exceeding temperature; Recommended control actions (based on cost function and risk threshold).

[0038] Data flow: Sensor → Edge gateway → Time series database → Prediction service (pull) → Parallel solution → Aggregation → Write back to time series database / push control module → Security monitoring and inspection → Gateway distribution.

[0039] S6. Based on real-time sensor data and temperature prediction curves, multi-dimensional control commands are generated through a model-free reinforcement learning agent and sent to the controller to execute corresponding actions; at the same time, a safety supervision mechanism is used to verify compliance and provide fallback for anomalies in the control process.

[0040] S6.1 Status, Actions, and Rewards 1. State vector: The observations and predictions at the current time t are concatenated as the input for reinforcement learning: This is the temperature curve output by the prediction module in the time domain H. The observation vectors are standardized (mean / variance) and time-of-day encoded before being input into the network.

[0041] 2. Motion space: Actions can be described in continuous or mixed manner: Continuous actions: The network output is scaled to the actual physical range using tanh / sigmoid.

[0042] Discrete / Hybrid Actions: Discretize each actuator into several levels using a DQN-type algorithm.

[0043] 3. Reward function: Design a multi-objective weighted reward system to simultaneously consider temperature control, energy consumption, equipment wear, and safety: From left to right: temperature squared error, instantaneous energy consumption caused by action, penalty for action change to avoid frequent switching, and penalty for serious safety violations (such as exceeding limits / freezing damage, 1{·} is the indicator function). 4. Offline training Simulation environment construction: A digital twin environment based on the energy conservation differential equations, realizing air-wall-soil-transpiration coupling (using the numerical kernel of the reusable prediction module). Training Process (Key Points): The agent is trained under a large number of weather scenarios (sunny / cloudy / freezing rain / strong winds, etc.) and parameter samples. Experience replay and important sampling are used to improve learning in rare scenarios. Prediction curves are used as observation inputs to enable the agent to learn feedforward decision-making (e.g., closing the blankets or turning on the ventilation in advance).

[0044] Safety training mechanism: Implement safety supervision (rule engine) simulation in the training environment: reject or replace high-risk actions with baseline actions in the environment, and use this information as a negative reward / instruction for the agent to learn "avoid risks". Start training in mild scenarios and gradually increase the proportion of extreme scenarios.

[0045] 5. Online deployment and fine-tuning Inference cycle: Inference frequency, e.g., every minute: Read the latest sensor and prediction curves, and perform forward computation actions in the policy network. After being checked by the Supervisor rules, the executor is issued.

[0046] Online security policy: Before issuing an action, the Supervisor performs a short-term (H') prediction simulation in parallel locally (or on a lightweight model copy): If a security threshold violation is predicted to occur within the next H', then: the original action is rejected, replaced with the baseline solution, and reported.

[0047] Online learning / fine-tuning: Safe fine-tuning is achieved using a small learning rate, batch replay, and policy freeze strategies; during online updates, a model backup and rollback mechanism is maintained, and the Supervisor can roll back to the previous stable policy version when performance degradation is detected.

[0048] Pseudocode (example): # Offline training pseudocode env = GreenhouseSim(param_sample=True) policy, critic = init_networks() for epoch in range(EPOCHS): trajectories = [] for episode in range(N_EPISODES_PER_EPOCH): state = env.reset() done = False while not done: action = policy.sample(state) next_state, reward, done, info = env.step(action) store_transition(state, action, reward, next_state, done) state = next_state # Use collected tracks for policy & criticism updates (PPO update) ppo_update(policy, critic, replay_buffer) # Regularly evaluate & save the best model # Online Reasoning (Production) while True: obs = read_sensors() # Real-time observation pred_curve = fetch_prediction() # Prediction curve for future H s = preprocess(obs, pred_curve) a_candidate = policy.predict(s) # Forward inference # Supervisor Inspection safe, a_safe = supervisor.check(a_candidate, s) if not safe: a_to_exec = fallback_action(a_safe) # Restricted action or PID output log_event("action_blocked", a_candidate, reason) else: a_to_exec = a_candidate send_to_actuators(a_to_exec) log_metrics(obs, pred_curve, a_to_exec) sleep(update_interval) 6. Safety supervision and verification The safety oversight layer reviews and verifies actions and statuses before, during, and after policy implementation. Key components and workflows are as follows: (1) Rule base (static + dynamic): Maintain static thresholds (minimum / maximum temperature and humidity) and dynamic rules based on prediction probability (e.g., if the predicted low temperature probability in the next half hour is expected to be below 10°C, then the heat preservation strategy will be implemented first).

[0049] (2) Real-time anomaly detection and automatic rollback: Real-time detection of equipment anomalies and decision anomalies. When equipment execution anomalies (execution failure) or output errors are recorded. When an anomaly is detected, it is executed according to priority: local fallback → record and report → if serious, online learning is stopped and manual intervention is required.

[0050] (3) Redundancy and fault tolerance mechanism: Redundancy is implemented for key sensors / actuators (multiple sensor deployment, dual power supply for equipment). When some sensors fail (such as power failure or abnormal data generation), data correction and cleaning are automatically performed to ensure the accuracy of data sources. When the execution equipment fails, the backup power supply is switched or an alarm is issued to relevant personnel as appropriate.

[0051] (4) Audit and alarm: Record each decision and its cause, security supervision coverage operation records, equipment execution receipts and environmental changes after execution; alarm rules can be dynamically set, and multi-level alarms (local SMS, email and telephone alarms) and maintenance orders can be automatically generated for different scenarios that trigger alarm rules.

[0052] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for temperature prediction and control based on earthen wall solar greenhouses, characterized in that, Includes the following steps: S1. Collect environmental data of the earthen wall greenhouse through multi-source sensors. The environmental data includes the transmitted radiation of the film, indoor and outdoor temperature and humidity, layer temperature of the wall, soil temperature and humidity, light and wind speed. S2. Preprocess the collected environmental data, including noise reduction, missing value compensation, time alignment and unit conversion, to form standardized observation data; S3. Establish a set of differential equations for energy conservation for the greenhouse covering layer, indoor air, crop canopy, layered earthen wall and ground layer. The set of equations includes solar radiation incident terms, convection heat transfer terms, transpiration latent heat terms, ventilation heat transfer terms, layered earthen wall heat conduction terms and actuator sensible heat terms. S4. The energy conservation differential equations are solved using a numerical solution strategy, and preliminary temperature prediction results are obtained by combining parameter samples with parallel calculations. S5. Post-process the preliminary temperature prediction results and output the final temperature prediction curve. S6. Based on temperature prediction curves and real-time sensor data, multi-dimensional control commands are generated through a model-free reinforcement learning agent and sent to the controller to execute corresponding actions; at the same time, a safety supervision mechanism is used to verify compliance and provide fallback for anomalies in the control process.

2. The temperature prediction and control method for a greenhouse based on earthen walls as described in claim 1, characterized in that, In S3, the energy conservation differential equations are as follows: In the formula: The heat capacity equation for air; To allow direct absorption of short-wave components from the air and canopy through the membrane; : Convection in the overlying layer; Long-wave radiation heat transfer can be written as (Nonlinear), in engineering, it is often linearized to ; : Convection heat transfer with the top layer of the wall; Ventilation and heat exchange ; The latent heat carried away by crop transpiration. (J / kg) represents the latent heat of vaporization of water. (kg / s) represents the amount of vaporization.

3. The temperature prediction and control method for a greenhouse based on earthen walls as described in claim 1, characterized in that, In S4, the numerical solution strategy includes discretizing the wall heat conduction equation using the Crank–Nicolson scheme and solving it using the Thomas algorithm, and calculating the air energy equation using the explicit RK4 method, achieving a coupled solution of the two through semi-implicit iteration; The heat conduction equation for the i-th layer of the wall is as follows: Using Crank–Nicolson time discretization: in This is a linear interlayer heat conduction operator. Merging all layers yields a tridiagonal linear system: Matrix (A) has a tridiagonal structure (coefficients are given by...). constitute); Air temperature equation writing: The air temperature is advanced using a fourth-order Runge-Kutta (RK4). If a sub-loop is used, RK4 uses the current or latest temperature reading within each air step. ; RK4 step (time) ): in This is the currently available temperature of the top layer of the wall.

4. The temperature prediction and control method for a greenhouse based on earthen walls as described in claim 1, characterized in that, In step S5, post-processing includes calculating the mean curve and confidence band, performing bias correction through a short-term residual model, extracting decision indicators such as low-temperature / high-temperature risk probability and threshold crossing time, and outputting the final temperature prediction curve.

5. The temperature prediction and control method for a greenhouse based on earthen walls as described in claim 1, characterized in that, The training and deployment of the model-free reinforcement learning agent in S6 includes: (1) State vector: The temperature prediction curve of the future time domain H is fused with the observation data at the current moment and then input into the network after standardization and time coding; (2) Action space: Actions can be described in a continuous or mixed manner; (3) Reward function: Multi-objective weighted reward to simultaneously consider temperature control, energy consumption, equipment wear and safety; Offline training: In a digital twin simulation environment centered on the energy conservation differential equations, training is conducted using various weather scenarios and parameter samples, and experience playback and important sampling are employed to improve the learning effect; (5) Online deployment: Read real-time data and prediction curves according to the set period, generate control commands through policy network inference, and issue them after verification by the security supervision mechanism.

6. The temperature prediction and control method for a greenhouse based on earthen walls as described in claim 1, characterized in that, The security monitoring mechanism includes: (1) Rule base: contains static temperature and humidity thresholds and dynamic rules based on prediction probability; (2) Real-time anomaly detection and automatic rollback: Detect equipment execution anomalies and decision anomalies, and execute local fallback, record and report or disable online learning according to priority; (3) Redundancy and fault tolerance mechanism: Redundant deployment of key sensors / actuators to support data error correction and cleaning when sensors fail, and backup solution switching when actuators fail; (4) Audit and alarm: Record the decision-making process, safety supervision operation and equipment execution status, and support the automatic generation of multi-level alarms and operation and maintenance orders.

Images