Greenhouse turtle breeding environment temperature intelligent regulation system based on photovoltaic energy

Abstract

The application discloses a greenhouse soft-shelled turtle breeding environment temperature intelligent regulation and control system based on photovoltaic energy and relates to the technical field of facility breeding intelligentization. The system comprises a pretreatment module, a collection module, an extraction module, an adjustment module and a prediction module. The pretreatment module collects real-time environment data, energy data, water body water quality data and thermal infrared soft-shelled turtle group activity video streams in a greenhouse and performs pretreatment. The extraction module extracts soft-shelled turtle behavior characteristic vectors representing feeding activities and group distribution based on the thermal infrared soft-shelled turtle group activity video streams after pretreatment. The adjustment module adjusts the thermal working characteristics of a spectral selective photovoltaic assembly at the top of the greenhouse based on the soft-shelled turtle behavior characteristic vectors and real-time environment data. The prediction module combines the soft-shelled turtle behavior characteristic vectors, the water body water quality data and the thermal working characteristic parameters of the spectral selective photovoltaic assembly. Through an online learning and abnormal diagnosis mechanism, the system can be continuously self-corrected, thereby ensuring the continuous effectiveness and operation reliability of regulation and control.

Description

Technical Field

[0001] This invention relates to the field of intelligent facility aquaculture technology, and in particular to an intelligent temperature control system for greenhouse turtle farming based on photovoltaic energy. Background Technology

[0002] Greenhouse-based factory farming of turtles is an important model for modernizing aquaculture, aiming to create a stable and controllable growth environment for turtles. Temperature is the most critical environmental factor affecting turtle metabolism and healthy growth. To reduce energy consumption and carbon emissions, utilizing photovoltaic modules covering the greenhouse roof to generate electricity and power temperature control equipment has become an important technological direction in the industry. Existing solutions typically monitor water temperature through sensors and use photovoltaic power to drive heat pumps and other equipment for temperature control based on preset thresholds or simple meteorological data, achieving preliminary energy substitution and automated management.

[0003] Existing solutions use water temperature as the sole control target and employ a passive control logic of monitoring and compensation. This approach fails to fully consider the complex coupled system comprised of photovoltaic module operation, greenhouse thermal environment, aquaculture water, and turtle behavior. The photothermal characteristics of photovoltaic modules directly affect the greenhouse thermal field, and changes in the thermal field can trigger changes in turtle behavior. Existing technologies cannot effectively sense and utilize the behavioral feedback information from turtles, making it difficult to predict the dynamic evolution of the system state. Furthermore, they cannot achieve dynamic and adaptive synergistic optimization among multiple objectives such as ensuring biological welfare, maintaining temperature stability, and optimizing energy efficiency, which can easily lead to control lag and low energy efficiency. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an intelligent temperature control system for greenhouse turtle farming based on photovoltaic energy to solve the problem of lagging control and low energy efficiency in the existing technology due to the inability to integrate feedback from turtle biological behavior, which leads to complex coupling of photovoltaic-thermal environment-water body-biological system.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides an intelligent temperature control system for greenhouse turtle farming based on photovoltaic energy, which includes a preprocessing module that collects real-time environmental data, energy data, water quality data and thermal infrared turtle group activity video streams in the greenhouse, and performs preprocessing.

[0008] The extraction module extracts behavioral feature vectors of turtles that characterize feeding activities and group distribution based on the preprocessed thermal infrared video stream of turtle population activity.

[0009] The adjustment module adjusts the thermal characteristics of the spectral selective photovoltaic modules at the top of the greenhouse based on the turtle behavior feature vector and real-time environmental data.

[0010] The prediction module inputs turtle behavior feature vectors, water quality data, and thermal characteristic parameters of spectrally selective photovoltaic modules into a high-fidelity digital twin model to predict the operating status of the environment and temperature intelligent control unit in future time periods.

[0011] The optimization module, based on the prediction results of the digital twin model, uses a deep reinforcement learning policy network for rolling optimization to generate the optimal cooperative control sequence;

[0012] The learning module parses and distributes the optimal collaborative control sequence as driving instructions for the underlying temperature control and execution devices, monitors the effect of the driving instructions, and feeds back the actual operation data to the digital twin model and deep reinforcement learning strategy network for online learning. When the temperature intelligent control unit deviates from its operation, it performs anomaly diagnosis and switches to degraded control mode.

[0013] As a preferred embodiment of the intelligent temperature control system for greenhouse turtle farming based on photovoltaic energy described in this invention, the pretreatment includes:

[0014] The collected environmental data, energy data, water quality data, and thermal infrared turtle activity video streams were standardized to obtain pre-processed environmental data, energy data, water quality data, and thermal infrared turtle activity video streams.

[0015] As a preferred embodiment of the intelligent temperature control system for greenhouse turtle farming based on photovoltaic energy described in this invention, the extraction of turtle behavioral feature vectors characterizing feeding activities and population distribution includes:

[0016] By using the preprocessed thermal infrared video stream of turtle group activity, the thermal image analysis branch is used to extract the thermal contour region of the turtle in a single frame image, and the temporal analysis branch is used to analyze the motion characteristics of the thermal contour region between consecutive frames to obtain the thermal motion trajectory sequence of the group.

[0017] Based on the population thermal motion trajectory sequence, the feeding activity index is calculated, and a population distribution thermal density map is generated.

[0018] The feeding activity index, the thermal density map of the population distribution, and the temporal change features extracted from the thermal motion trajectory sequence are fused and encoded to form a turtle behavior feature vector that characterizes feeding activity and population distribution.

[0019] As a preferred embodiment of the intelligent temperature control system for greenhouse turtle farming based on photovoltaic energy described in this invention, the adjustment of the thermal characteristics of the spectrally selective photovoltaic module at the top of the greenhouse includes:

[0020] The population distribution thermal density map, combined with real-time collected internal thermal environment parameters of the greenhouse, drives the fuzzy inference unit of behavior-thermal environment mapping to calculate the target transmittance adjustment command of the spectrally selective photovoltaic module in the thermal band.

[0021] Using the target transmittance adjustment command, the electrochromic drive unit of the spectral selective photovoltaic module deployed in the top section of the greenhouse is controlled. By changing the voltage applied to the functional layer, the transmittance of the spectral selective photovoltaic module in the heat band is adjusted to the preset target value.

[0022] By utilizing the transmittance of the spectrally selective photovoltaic module in the heat band, combined with the total solar irradiance and wind direction and speed in real-time environmental data, a dynamic thermal disturbance compensation model is used to generate power feedforward compensation commands for the active temperature control equipment inside the greenhouse.

[0023] As a preferred embodiment of the intelligent temperature control system for greenhouse turtle farming based on photovoltaic energy described in this invention, the operating status of the intelligent temperature control unit for predicting future time periods includes:

[0024] By utilizing the thermal characteristic parameter sequence of historical spectrally selective photovoltaic modules, water quality data, and corresponding turtle behavior feature vector sequences representing feeding activities and population distribution, a photovoltaic thermal perturbation-turtle population thermal distribution response coupled subnetwork is constructed through an adversarial training mechanism, and a high-fidelity digital twin model is embedded in it.

[0025] Real-time behavioral feature vectors of turtles representing feeding activities and group distribution, real-time water quality data, and real-time thermal characteristic parameters of spectrally selective photovoltaic modules are input into a high-fidelity digital twin model that has an embedded photovoltaic thermal disturbance-turtle group thermal distribution response coupling subnetwork, driving the high-fidelity digital twin model to perform online state synchronization.

[0026] By utilizing a high-fidelity digital twin model that has completed online status synchronization and combining it with future weather forecast data, forward simulation calculations are performed to generate a predicted trajectory of the operating status of the intelligent environmental and temperature control unit for future periods.

[0027] As a preferred embodiment of the intelligent temperature control system for greenhouse turtle farming based on photovoltaic energy described in this invention, the optimal cooperative control sequence includes:

[0028] Based on the predicted trajectory of the operating status of the intelligent environmental and temperature control unit in the future time period, combined with the real-time turtle behavior feature vector representing feeding activities and group distribution and real-time water quality data, a dedicated state of the deep reinforcement learning strategy network is constructed.

[0029] By utilizing the dedicated state of a deep reinforcement learning policy network, a deep reinforcement learning policy network with a hierarchical constraint satisfaction mechanism is driven to select a subset of feasible actions in the action space that satisfy the lower limit constraint of photovoltaic power generation.

[0030] By utilizing the subset of feasible actions that satisfy the lower limit constraint of photovoltaic power generation, a deep reinforcement learning strategy network with a hierarchical constraint satisfaction mechanism is driven to screen out the subset of feasible actions that satisfy the hard constraints of aquaculture water temperature and water quality indicators.

[0031] By utilizing a subset of feasible actions that satisfy the lower limit constraint of photovoltaic power generation and the hard constraints of aquaculture water temperature and water quality, a deep reinforcement learning policy network with a hierarchical constraint satisfaction mechanism is driven to calculate the comprehensive evaluation value of each feasible action and select the action with the highest comprehensive evaluation value.

[0032] The action with the highest comprehensive evaluation value selected by the deep reinforcement learning strategy network that satisfies the hierarchical constraint conditions is extended into the optimal collaborative control sequence of the thermal characteristic target value sequence and the temperature setpoint sequence of the temperature intelligent control unit.

[0033] As a preferred embodiment of the intelligent temperature control system for greenhouse turtle farming based on photovoltaic energy described in this invention, the driving commands for the underlying temperature control and execution device include:

[0034] Extract the target value of the thermal characteristics of the spectral selective photovoltaic module and the temperature setpoint of the temperature intelligent control unit at the current decision step from the optimal cooperative control sequence, and generate atomic control instructions with high-precision execution timestamps.

[0035] By using atomic control instructions with high-precision execution timestamps, and through an instruction conflict arbitrator that predicts response delay, the timing of atomic control instructions with differences in device response speed or physical coupling conflicts is finely adjusted and rearranged to obtain a conflict-free time-stamped drive instruction queue.

[0036] The conflict-free, time-stamped drive command queue is synchronously sent to the underlying temperature control and execution devices corresponding to the electrochromic drive unit and the temperature intelligent control unit of the spectral selective photovoltaic module, according to the time stamp.

[0037] As a preferred embodiment of the intelligent temperature control system for greenhouse turtle farming based on photovoltaic energy described in this invention, the effects of the monitoring and driving commands include:

[0038] After the underlying temperature control and execution equipment executes the drive command, it collects the actual status feedback signal returned by the underlying temperature control and execution equipment, the real-time readings of the environmental sensors in the greenhouse, and the real-time data of the water quality sensors to generate actual operating data that reflects the immediate effect of the drive command.

[0039] As a preferred embodiment of the intelligent temperature control system for greenhouse turtle farming based on photovoltaic energy described in this invention, the online learning includes:

[0040] By utilizing real-time readings from environmental sensors, water quality data, and corresponding drive commands from actual operational data, the digital twin model is driven to perform online incremental parameter correction, dynamically adjusting the physical parameters and dynamic water quality parameters in the digital twin model that affect prediction accuracy.

[0041] By using the digital twin model after online incremental parameter correction, the state transition process corresponding to the actual operating data is re-evaluated, and an enhanced transition trajectory is generated.

[0042] The enhanced transfer trajectory is input into the online learning pipeline of the deep reinforcement learning policy network, triggering an asynchronous incremental update of the policy network parameters by a dual-value network based on a competitive architecture.

[0043] As a preferred embodiment of the intelligent temperature control system for greenhouse turtle farming based on photovoltaic energy described in this invention, the process of performing anomaly diagnosis and switching to a degraded control mode includes,

[0044] When the temperature intelligent control unit deviates from its operation, the actual status feedback signal of the actuator in the temperature intelligent control unit, the real-time reading of the environmental sensor in the greenhouse and the water quality sensor data are collected. The actual status feedback signal, the real-time reading of the environmental sensor and the water quality data are compared with the expected state of the optimal collaborative control sequence to generate a deviation analysis report on the deviation pattern and trend.

[0045] Input the deviation analysis report of deviation patterns and trends into the anomaly diagnosis engine of the knowledge graph to locate the root cause of the deviation in the operation of the temperature intelligent control unit.

[0046] Based on the root cause identified by the anomaly diagnosis engine, the index's pre-defined multi-level fault-tolerant rule base smoothly switches from the current optimal collaborative control strategy to the corresponding degraded control mode.

[0047] The beneficial effects of this invention are as follows: It combines the behavioral feature vectors of turtles obtained through thermal infrared imaging with the photothermal characteristic adjustment of spectrally selective photovoltaic modules, thereby driving a high-fidelity digital twin model for dynamic simulation and prediction. This model system can proactively simulate the dynamic evolution of the coupled system of photovoltaic-thermal environment-aquaculture water-turtle organisms, providing crucial basis for subsequent collaborative optimization decisions. Based on predictions, it utilizes a deep reinforcement learning strategy network for multi-objective, multi-constraint rolling optimization, automatically generating optimal collaborative control sequences that satisfy multiple objectives such as energy utilization, temperature stability, and biological welfare. Through online learning and anomaly diagnosis mechanisms, the system can continuously self-correct, ensuring the continuous effectiveness and operational reliability of regulation. Attached Figure Description

[0048] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a schematic diagram of an intelligent temperature control system for greenhouse turtle farming based on photovoltaic energy. Detailed Implementation

[0050] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0051] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0052] Secondly, the term "one embodiment" or "example" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the invention. The appearance of an embodiment in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that mutually excludes other embodiments.

[0053] Reference Figure 1 This is one embodiment of the present invention, which provides an intelligent temperature control system for greenhouse turtle farming based on photovoltaic energy, comprising the following steps:

[0054] The preprocessing module collects real-time environmental data, energy data, water quality data, and thermal infrared video streams of turtle activity within the greenhouse, and performs preprocessing.

[0055] The collected environmental data, energy data, water quality data, and thermal infrared turtle activity video streams were standardized to obtain pre-processed environmental data, energy data, water quality data, and thermal infrared turtle activity video streams.

[0056] Furthermore, the collected environmental data, energy data, water quality data, and thermal infrared turtle activity video stream were standardized. The environmental data was processed by removing noise, compensating for sensor zero-point drift, and scaling according to a unified dimension. The energy data was standardized by normalizing photovoltaic power generation, grid feed power, and load power to standard operating conditions and aligning the time series. The water quality data was filtered for outliers and temperature-compensated for dissolved oxygen and pH parameters based on water temperature. The thermal infrared turtle activity video stream was processed by time-series jitter removal and non-uniformity correction to eliminate detector noise and unify frame rate and resolution. At the same time, the timestamps of each data stream were synchronized and aligned based on high-precision time synchronization to obtain the preprocessed environmental data, energy data, water quality data, and thermal infrared turtle activity video stream.

[0057] The extraction module extracts behavioral feature vectors representing feeding activities and group distribution of turtles based on the preprocessed thermal infrared turtle activity video stream.

[0058] By using the preprocessed thermal infrared video stream of turtle group activity, the thermal contour region of the turtles in a single frame image is extracted using the thermal image analysis branch, and the motion characteristics of the thermal contour region between consecutive frames are analyzed using the temporal analysis branch to obtain the thermal motion trajectory sequence of the group.

[0059] Furthermore, for each frame of the preprocessed thermal infrared turtle activity video stream, an adaptive threshold segmentation algorithm is executed using the thermal image analysis branch. Regions with temperatures higher than the water background temperature and that are connected are marked as candidate hot spots. Morphological closing operations are used to fill the internal holes of the candidate hot spots and remove noise points with areas smaller than a set threshold, thereby extracting the independent thermal contour regions of each turtle in each frame. Simultaneously, a multi-target tracking algorithm is used to process the marked thermal contour regions in consecutive frames through the temporal analysis branch. A cost matrix is ​​constructed based on the Euclidean distance of the centroid positions of the thermal contour regions between adjacent frames and the shape similarity. The Hungarian algorithm is used to solve for the optimal matching association, establishing an identity correspondence between the thermal contour regions of the same turtle in different frames. The centroid coordinate changes of each turtle in the continuous time series are recorded, ultimately forming a group thermal motion trajectory sequence containing the centroid trajectory of each turtle over time.

[0060] Based on the population's thermal motion trajectory sequence, the feeding activity index is calculated, and a population distribution thermal density map is generated.

[0061] Furthermore, based on the group thermal motion trajectory sequence, for each frame time point, the displacement vector magnitude of each turtle within that time window is calculated. Individuals with displacement vector magnitudes greater than the movement distance threshold are identified as valid moving individuals. The displacement magnitudes of valid moving individuals are accumulated and then divided by the product of the total number of individuals and the time interval to obtain the feeding activity index, which characterizes the intensity of group movement. Simultaneously, the centroid coordinates of each turtle at a specific moment in the group thermal motion trajectory sequence are used as input sample points for Gaussian kernel density estimation. A fixed Gaussian kernel bandwidth is set, and the probability density within the grid area divided in the breeding pond plane is integrally calculated to generate a two-dimensional matrix-form group distribution thermal density map reflecting the spatial aggregation degree of the turtle group.

[0062] The expression for the food intake index is:

[0063] ;

[0064] in, For a moment The group's feeding activity The total number of turtle individuals detected in the current frame. For individuals In the time window displacement vector magnitude within, For individuals exist The thermal profile area at any given time. The normalized baseline for the thermal profile area of ​​a turtle individual is used. For time window, For indicator functions, In the first At the [time]th moment The change in position of the turtle. For individual turtle indexes, This represents the threshold for movement distance.

[0065] The feeding activity index, the thermal density map of the population distribution, and the temporal change features extracted from the thermal motion trajectory sequence are fused and encoded to form a turtle behavior feature vector that characterizes feeding activity and population distribution.

[0066] Furthermore, the feeding activity index is directly incorporated into the feature vector as a scalar feature value. A two-dimensional discrete Fourier transform is performed on the thermal density map of the population distribution to extract the first few low-frequency coefficients as spatial distribution feature vectors. The displacement amplitude histogram distribution characteristics and trajectory direction angle variance characteristics of all turtle individuals within a unit time window are statistically analyzed from the thermal motion trajectory sequence of the population as temporal variation characteristics. Principal component analysis is used to reduce the dimensionality of the above-mentioned spliced ​​high-dimensional feature vector, remove the linear correlation and redundant information between features, and generate a fixed-dimensional, numerical feature vector representing the feeding activity and population distribution of turtles.

[0067] The adjustment module adjusts the thermal characteristics of the spectral selective photovoltaic modules at the top of the greenhouse based on the turtle behavior feature vector and real-time environmental data.

[0068] The population distribution thermal density map, combined with real-time collected internal thermal environment parameters of the greenhouse, drives the fuzzy inference unit of behavior-thermal environment mapping to calculate the target transmittance adjustment command of the spectrally selective photovoltaic module in the thermal band.

[0069] Furthermore, the normalized density integral values ​​of the feeding and resting areas are extracted from the thermal density map of the population distribution as behavioral input variables. At the same time, the real-time collected air temperature and aquaculture water temperature inside the greenhouse are read, and the deviations from the set values ​​are calculated as thermal environment input variables. The behavioral input variables and thermal environment input variables are substituted into the preset fuzzy membership function to obtain the corresponding membership values. Based on several predefined inference rules of IF behavioral state and THEN target transmittance in the fuzzy rule base, the activation intensity of each rule is calculated. The output fuzzy set of all activation rules is aggregated by the product method, and the numerical value of the target transmittance adjustment command of the spectral selective photovoltaic module in the thermal band is obtained by defuzzification calculation using the centroid method.

[0070] The expression for the target transmittance adjustment command is:

[0071] ;

[0072] in, For target transmittance instructions, For the total number of rules, For rule indexing, For the first The activation strength of a fuzzy inference rule. For the first The center value of the output fuzzy set corresponding to each rule.

[0073] Using the target transmittance adjustment command, the electrochromic drive unit of the spectral selective photovoltaic module deployed in the top section of the greenhouse is controlled. By changing the voltage applied to the functional layer, the transmittance of the spectral selective photovoltaic module in the heat band is adjusted to the preset target value.

[0074] Furthermore, the value of the target transmittance adjustment command is compared with the current actual transmittance measurement value of the spectral selective photovoltaic module. The proportional-integral controller algorithm is used to determine the voltage adjustment amount that needs to be applied. This voltage adjustment amount is converted into a corresponding digital-to-analog conversion output signal and sent to the electrochromic drive unit of the spectral selective photovoltaic module deployed in the top of the greenhouse. The electrochromic drive unit outputs a DC voltage of corresponding amplitude and polarity to the functional layer, driving the electrochromic material in the functional layer to undergo a redox reaction to change its optical absorption characteristics, thereby adjusting the transmittance of the spectral selective photovoltaic module in the heat band to the preset target value.

[0075] By utilizing the transmittance of the spectrally selective photovoltaic module in the heat band, combined with the total solar irradiance and wind direction and speed in real-time environmental data, a dynamic thermal disturbance compensation model is used to generate power feedforward compensation commands for the active temperature control equipment inside the greenhouse.

[0076] Furthermore, the dynamic thermal disturbance compensation model is an empirical model based on the greenhouse heat balance equation. Its inputs are total solar irradiance, real-time transmittance of spectrally selective photovoltaic modules in the heat band, wind direction and wind speed, instantaneous changes in the internal heat load of the greenhouse caused by changes in the above external factors, and the instantaneous power compensation required to maintain the target temperature is converted according to the energy efficiency ratio of active temperature control equipment such as heat pumps. This power compensation is used as the power feedforward compensation command.

[0077] The prediction module inputs turtle behavior feature vectors, water quality data, and thermal characteristic parameters of spectrally selective photovoltaic modules into a high-fidelity digital twin model to predict the operating status of the environment and temperature intelligent control unit in future periods.

[0078] By utilizing the thermal characteristic parameter sequence of historical spectrally selective photovoltaic modules, water quality data, and corresponding turtle behavior feature vector sequences representing feeding activities and population distribution, a photovoltaic thermal perturbation-turtle population thermal distribution response coupled subnetwork is constructed through an adversarial training mechanism, and a high-fidelity digital twin model is embedded.

[0079] Furthermore, a historical dataset is prepared, consisting of a sequence of thermal characteristic parameters of historical spectrally selective photovoltaic modules, a corresponding sequence of water quality data, and a sequence of turtle behavioral feature vectors representing feeding activities and group distribution. A generator network and a discriminator network are constructed. The generator network takes a sequence of historical thermal characteristic parameters and a sequence of water quality data as input and generates a sequence of turtle behavioral feature vectors representing feeding activities and group distribution for a future time series. The discriminator network is used to determine whether the paired input sequences, i.e., a sequence of thermal characteristic parameters and water quality data and its corresponding sequence of turtle behavioral feature vectors representing feeding activities and group distribution, are real sequences from the historical dataset or fake sequences generated by the generator network. The generator network and the discriminator network are trained alternately through an adversarial training mechanism until the discriminator network can no longer effectively distinguish between real sequences and generated sequences. The trained generator network is then used as a photovoltaic thermal perturbation-turtle group thermal distribution response coupling subnetwork, and its output nodes are connected to the input interface representing the turtle group thermal distribution state variables in the high-fidelity digital twin model to complete the embedding.

[0080] Real-time behavioral feature vectors of turtles representing feeding activities and group distribution, real-time water quality data, and real-time thermal characteristic parameters of spectrally selective photovoltaic modules are input into a high-fidelity digital twin model that has an embedded photovoltaic thermal disturbance-turtle group thermal distribution response coupling subnetwork, driving the high-fidelity digital twin model to perform online state synchronization.

[0081] Furthermore, real-time behavioral characteristic vectors of turtles representing feeding activities and group distribution, real-time water quality data, and real-time thermal performance parameters of spectrally selective photovoltaic modules are input as a set of observations into a high-fidelity digital twin model embedded with a photovoltaic thermal perturbation-turtle group thermal distribution response coupling subnetwork. The high-fidelity digital twin model uses an extended Kalman filter algorithm to fuse these observations with the previous state estimate within the model. The predicted value output by the photovoltaic thermal perturbation-turtle group thermal distribution response coupling subnetwork serves as a priori estimate of the turtle group thermal distribution state variable and participates in the filtering update, thereby correcting all state variables within the model and ensuring that the virtual state of the high-fidelity digital twin model is consistent with the physical state of the real greenhouse, thus achieving online state synchronization.

[0082] By utilizing a high-fidelity digital twin model that has completed online status synchronization and combining it with future weather forecast data, forward simulation calculations are performed to generate a predicted trajectory of the operating status of the intelligent environmental and temperature control unit for future periods.

[0083] Furthermore, future weather forecast data includes time series of total solar irradiance, ambient temperature, humidity, wind direction, and wind speed over a future period. Performing forward simulation calculations refers to using the current state of the synchronized high-fidelity digital twin model as a starting point, and taking the future weather forecast data as dynamic input, iteratively running the high-fidelity digital twin model step by step according to discrete time steps. At each time step, the high-fidelity digital twin model calculates the greenhouse internal environment state, water quality state, spectrally selective photovoltaic module state, and the operating state of the temperature intelligent control unit for the next time step based on all current internal states, external weather inputs, and built-in control logic. Iterative calculations across multiple consecutive time steps generate a complete time series—the prediction trajectory—extending from the current moment to a specified future moment, concerning the greenhouse environmental parameters, water quality parameters, photovoltaic module state, and the operating states of the temperature intelligent control unit setpoint and energy consumption.

[0084] The predicted trajectory expression is:

[0085]

[0086] in, For the future The predicted trajectory at each time step, For a high-fidelity digital twin model that has completed online status synchronization, For the current moment The synchronization state vector, For control input sequences in future time periods, This represents a sequence of external disturbances for future periods. To predict the length of the time domain.

[0087] The optimization module, based on the prediction results of the digital twin model, uses a deep reinforcement learning policy network for rolling optimization to generate the optimal cooperative control sequence.

[0088] Based on the predicted trajectory of the operating status of the intelligent environmental and temperature control unit in the future time period, combined with real-time turtle behavior feature vectors representing feeding activities and group distribution and real-time water quality data, a dedicated state of the deep reinforcement learning strategy network is constructed.

[0089] Furthermore, from the predicted trajectory of the operating status of the intelligent environmental and temperature control unit in the future, key state variable values ​​at multiple discrete time points are extracted, including predicted values ​​of aquaculture water temperature, air temperature, photovoltaic power generation, and thermal characteristic parameters of spectrally selective photovoltaic modules. These predicted value sequences arranged in chronological order are then concatenated with real-time turtle behavior feature vectors representing feeding activities and group distribution, and various indicator values ​​from real-time water quality data to form an extended high-dimensional hybrid feature vector. This high-dimensional hybrid feature vector is then input into a multilayer perceptron encoder, where dimensionality reduction and feature extraction are performed through fully connected layers and nonlinear activation functions. Finally, a low-dimensional, fixed-length real-valued vector is output, which is the dedicated state of the constructed deep reinforcement learning policy network.

[0090] By utilizing the dedicated states of a deep reinforcement learning policy network, a deep reinforcement learning policy network with a hierarchical constraint satisfaction mechanism is driven to select a subset of feasible actions in the action space that satisfy the lower limit constraint of photovoltaic power generation.

[0091] Furthermore, the action space of the deep reinforcement learning policy network is defined as a combination of adjustment instructions for the target value of the thermal characteristics of the spectrally selective photovoltaic module and the temperature setpoint of the intelligent temperature control unit. The deep reinforcement learning policy network with a hierarchical constraint satisfaction mechanism first evaluates all possible action combinations under a specific state. Then, by querying the built-in simplified performance evaluation function or using rapid simulation with a high-fidelity digital twin model, it predicts the corresponding future photovoltaic power generation curve. Subsequently, the predicted power generation curve is compared with a preset lower limit constraint on photovoltaic power generation. Any action combination that might cause the power generation to fall below this lower limit constraint at any future time is directly eliminated. All remaining action combinations constitute a subset of feasible actions that satisfy the lower limit constraint on photovoltaic power generation.

[0092] By utilizing the subset of actionable actions that satisfy the lower limit constraint of photovoltaic power generation, a deep reinforcement learning strategy network with a hierarchical constraint satisfaction mechanism is driven to select the subset of actionable actions that satisfy the hard constraints of aquaculture water temperature and water quality indicators.

[0093] Furthermore, based on the initial screening, the deep reinforcement learning strategy network with a hierarchical constraint satisfaction mechanism continues to evaluate each action combination in the subset of possible actions, predicting its dynamic impact on future aquaculture water temperature and key water quality indicators such as pH and ammonia nitrogen concentration. Prediction is also achieved by querying a simplified evaluation function or calling a high-fidelity digital twin model's rapid simulation module. The predicted water temperature and water quality indicator change trajectories are compared with preset hard constraints such as the allowable fluctuation range of aquaculture water temperature and the safety thresholds for water quality indicators. Any action combinations that might cause the water temperature to exceed the allowable range or the water quality indicators to exceed the safety thresholds are further eliminated. The remaining action combinations constitute the subset of possible actions that simultaneously satisfy the lower limit constraint of photovoltaic power generation and the hard constraints of aquaculture water temperature and water quality.

[0094] By utilizing a subset of actionable actions that satisfy the lower limit constraint of photovoltaic power generation and the hard constraints of aquaculture water temperature and water quality, a deep reinforcement learning policy network with a hierarchical constraint satisfaction mechanism is driven to calculate the comprehensive evaluation value of each action and select the action with the highest comprehensive evaluation value.

[0095] Furthermore, the deep reinforcement learning strategy network with a hierarchical constraint satisfaction mechanism incorporates a value evaluation subnetwork. This subnetwork, for each action in the subset of possible actions, comprehensively considers multiple optimization objectives, including the predicted total photovoltaic power generation, the total energy consumption of the intelligent temperature control unit, the stability of water temperature, the quality of water indicators, and the comfort of turtle behavior. These objectives are quantified and weighted by preset weights to calculate the comprehensive evaluation value of each possible action. The comprehensive evaluation value reflects the overall performance of the action plan in terms of energy efficiency, environmental stability, and biological welfare, while satisfying all hard constraints. The deep reinforcement learning strategy network with a hierarchical constraint satisfaction mechanism selects the action with the highest comprehensive evaluation value from the subset of possible actions as the optimal action for the current decision step.

[0096] The expression for the comprehensive evaluation value is:

[0097] ;

[0098] in, In the state Next action The overall evaluation value, In the state Next action The original value of the action For smooth truncation function, For the standard deviation estimate of the penalty term, To integrate the value standard deviation output by the value network, This is a predicted trajectory for the operating status of the intelligent environmental and temperature control unit in the future. Candidate actions are selected from the subset of feasible actions to meet the lower limit constraints of photovoltaic power generation and the hard constraints of aquaculture water temperature. This is a dynamic penalty coefficient. This is the pessimistic coefficient.

[0099] The action with the highest comprehensive evaluation value selected by the deep reinforcement learning strategy network that satisfies the hierarchical constraint conditions is extended into the optimal collaborative control sequence of the thermal characteristic target value sequence and the temperature setpoint sequence of the temperature intelligent control unit.

[0100] Furthermore, the optimal action is the adjustment amount of the target thermal characteristics of the spectrally selective photovoltaic module and the temperature setpoint adjustment amount of the intelligent temperature control unit at a single decision time step. To generate a control command sequence for the entire future optimization cycle, starting from the current optimal action, and combining it with continuous predictions of future states using a high-fidelity digital twin model, the current single-step optimal decision is extended into a continuous sequence of thermal characteristic target value changes covering multiple future time steps, along with a matching sequence of temperature setpoint changes for the intelligent temperature control unit. This sequence ensures the temporal continuity and coordination of the control commands, constituting the optimal coordinated control sequence.

[0101] The learning module parses and distributes the optimal collaborative control sequence as driving instructions for the underlying temperature control and execution devices, monitors the effect of the driving instructions, and feeds back the actual operation data to the digital twin model and deep reinforcement learning strategy network for online learning. When the temperature intelligent control unit deviates from its operation, it performs anomaly diagnosis and switches to degraded control mode.

[0102] Extract the target value of the thermal characteristics of the spectral selective photovoltaic module and the temperature setpoint of the temperature intelligent control unit at the current decision step from the optimal cooperative control sequence, and generate atomic control instructions with high-precision execution timestamps.

[0103] Furthermore, the analysis process decomposes the target value of thermal characteristics into specific voltage regulation commands or position adjustment commands for the electrochromic layer of the spectrally selective photovoltaic module, and decomposes the temperature setpoint into specific operation commands such as heat pump start / stop, power setting, or valve opening. Each basic operation command that cannot be further divided after decomposition is an atomic control command. An atomic control command is bound to a high-precision execution timestamp generated based on a network time protocol or a high-precision clock source. The timestamp precisely specifies the exact moment when the command is expected to be executed, thus generating an atomic control command with a high-precision execution timestamp.

[0104] By using atomic control instructions with high-precision execution timestamps, and through an instruction conflict arbitrator that predicts response delays, the timing of atomic control instructions with differences in device response speed or physical coupling conflicts is fine-tuned and rearranged to obtain a conflict-free time-stamped drive instruction queue.

[0105] Furthermore, the instruction conflict arbitrator for response delay prediction incorporates a knowledge base that records typical response delays, action setup times, and interrelationships among various underlying devices such as electrochromic components of spectrally selective photovoltaic modules, heat pump compressors, circulating water pumps, and electric valves. The arbitrator analyzes all atomic control instructions with high-precision execution timestamps. If multiple instructions have timestamps that are too close, potentially leading to insufficient parallel response from devices or physical coupling of instructions that could cause adverse interference, the arbitrator makes minor, compensatory adjustments to the timestamps of relevant instructions within the allowable decision execution time window, based on the response delay and physical coupling model. After timing rearrangement, the execution times of all instructions are rationally distributed, ensuring physical independence and executability between instructions, forming a conflict-free, time-stamped driven instruction queue.

[0106] The conflict-free, time-stamped drive command queue is synchronously sent to the underlying temperature control and execution devices corresponding to the electrochromic drive unit and the temperature intelligent control unit of the spectral selective photovoltaic module, according to the time stamp.

[0107] Furthermore, the conflict-free time-stamped drive command distribution and execution coordinator, based on the time stamp attached to each command in the drive command queue, sends the corresponding specific control command to the target underlying devices such as the electrochromic system of the spectral selective photovoltaic module, heat pump controller, water pump inverter, and electric valve controller the instant the high-precision clock inside the coordinator reaches the specified time stamp, via fieldbus, industrial Ethernet, or wireless communication protocol. This synchronized time stamp distribution ensures that the absolute or relative time at which different devices receive the control command strictly conforms to the preset coordination timing requirements.

[0108] After the underlying temperature control and execution equipment executes the drive command, it collects the actual status feedback signal returned by the underlying temperature control and execution equipment, the real-time readings of the environmental sensors in the greenhouse, and the real-time data of the water quality sensors to generate actual operating data that reflects the immediate effect of the drive command.

[0109] Furthermore, the actual status feedback signals include real-time voltage and transmittance readings of the electrochromic layer of the spectrally selective photovoltaic module, heat pump operating power and status, water pump speed, valve opening, etc. Real-time readings from environmental sensors include greenhouse air temperature, humidity, water temperature at different depths, and illuminance. Real-time data from water quality sensors include dissolved oxygen, pH value, and ammonia nitrogen concentration. All of these data, collected within a preset observation window after the command is issued, are correlated and packaged with the issued drive command itself to constitute the actual operating data. This data completely records the instantaneous correspondence between the control command input and the equipment and environmental status output.

[0110] By utilizing real-time readings from environmental sensors, water quality data, and corresponding drive commands from actual operational data, the digital twin model is driven to perform online incremental parameter correction, dynamically adjusting the physical parameters and dynamic water quality parameters in the digital twin model that affect prediction accuracy.

[0111] Furthermore, after receiving actual operational data, the digital twin model uses drive commands as input to perform a forward simulation within the model, obtaining predicted environmental and water quality values ​​at the corresponding time points. These predicted values ​​are then compared with real-time readings from environmental sensors and water quality data collected during actual operation to calculate the prediction error. Based on this error, parameter estimation algorithms such as recursive least squares, extended Kalman filtering, or Bayesian update are used to estimate and adjust pre-identified key physical parameters and dynamic water quality parameters with uncertainties in the digital twin model online. These parameters may include the heat transfer coefficient of the greenhouse enclosure structure, the heat capacity of the water body, the heat absorption coefficient of the photovoltaic modules, and the rate constant of water quality change kinetics. Through online incremental parameter correction, the internal parameters of the digital twin model are continuously fine-tuned, making its simulation output closer to the dynamic characteristics of a real greenhouse.

[0112] By using a digital twin model with online incremental parameter correction, the state transition process corresponding to the actual operating data is re-evaluated, and an enhanced transition trajectory is generated.

[0113] Furthermore, starting with the state of the digital twin model and the actual driving commands issued just before the actual data acquisition, a new simulation of the state evolution process from command issuance to data acquisition completion was performed using a digital twin model whose parameters were updated after online incremental parameter correction. This simulation was based on a more accurate corrected model, and the complete state transition path obtained from the previous state to the new state after the given command was applied is more consistent and has greater learning value than the path based on the uncorrected model or simply actual observation data. This path is labeled as the augmented transition trajectory. The augmented transition trajectory contains a complete four-tuple of information: state, action, reward, and next state, making it a high-quality training sample for reinforcement learning.

[0114] The enhanced transfer trajectory is input into the online learning pipeline of the deep reinforcement learning policy network, triggering an asynchronous incremental update of the policy network parameters by a dual-value network based on a competitive architecture.

[0115] Furthermore, the online learning pipeline of the deep reinforcement learning policy network continuously receives generated augmented transition trajectories. The dual-value network, based on a competitive architecture, consists of two structurally identical but independently updated value networks used to evaluate state-action values. During each incremental update, a batch of augmented transition trajectories is sampled from the experience replay pool, and the target Q-value is obtained using the one with the lower estimate from the dual-value network to mitigate overestimation. Loss is calculated using temporal difference error, and the parameters of the policy and value networks in the deep reinforcement learning policy network are updated asynchronously using gradient descent. This asynchronous incremental update method enables the deep reinforcement learning policy network to continuously learn from real-world operational experience, constantly optimizing its decision-making strategy and adapting to long-term changes in greenhouse environments, equipment performance, and turtle behavior.

[0116] When the temperature intelligent control unit deviates from its operation, the actual status feedback signal of the actuator in the temperature intelligent control unit, the real-time reading of the environmental sensor in the greenhouse, and the water quality sensor data are collected. The actual status feedback signal, the real-time reading of the environmental sensor, and the water quality data are compared with the expected state of the optimal collaborative control sequence to generate a deviation analysis report on the deviation pattern and trend.

[0117] Furthermore, when the actual control effect of the intelligent temperature control unit, such as water temperature, is detected to deviate continuously or significantly from the expected target value corresponding to the optimal collaborative control sequence, the anomaly diagnosis process is immediately triggered. The collected data includes the real-time operating status, power, and flow feedback of heat pumps, water pumps, and valve actuators, as well as greenhouse air temperature, humidity, water temperature at various points, and key water quality parameters. These actually collected data sequences are compared point-by-point with the expected state sequence predicted by the high-fidelity digital twin model corresponding to the optimal collaborative control sequence. The magnitude, direction, duration, trend, and spatial distribution pattern of the deviation are analyzed, such as whether it is a global temperature drift or a local temperature anomaly, whether it is accompanied by a sudden change in specific water quality parameters, and whether it is related to anomalies in the state of specific actuators. This generates a structured deviation pattern and trend deviation analysis report.

[0118] The deviation analysis report of deviation patterns and trends is input into the anomaly diagnosis engine of the knowledge graph to locate the root cause of the deviation in the operation of the temperature intelligent control unit.

[0119] Furthermore, the anomaly diagnosis engine of the knowledge graph pre-configures a structured knowledge graph. Its nodes represent various entities and concepts within the greenhouse, such as equipment, sensors, environmental parameters, turtle behavior, physical relationships, and failure modes, while edges represent the relationships between entities. After the deviation analysis report is parsed, it is matched and reasoned against the knowledge graph. The engine traverses the knowledge graph to find the most likely fault chain or combination of causes that can simultaneously explain all the anomaly patterns in the deviation analysis report. For example, combining multiple deviation facts such as low water temperature, normal heat pump power feedback but abnormally low outlet water temperature, and low water circulation velocity in a certain area, the knowledge graph may infer that the root cause is a partial blockage of the circulating water pump or a localized leak in the pipes, rather than simply insufficient heat pump heating capacity, thus pinpointing the root cause of the deviation in the operation of the intelligent temperature control unit.

[0120] Based on the root cause identified by the anomaly diagnosis engine, the index's pre-defined multi-level fault-tolerant rule base smoothly switches from the current optimal collaborative control strategy to the corresponding degraded control mode.

[0121] Furthermore, the pre-built multi-level fault-tolerant rule base is a set of rules, where each rule is associated with one or a group of diagnosed root causes and specifies the corresponding degraded control strategy. Degraded control strategies may include switching to backup equipment, adjusting control objectives, adopting a more conservative but robust control algorithm, or notifying manual intervention. For example, if the root cause is identified as a decrease in the efficiency of a circulating water pump, the fault-tolerant rule may instruct the startup of the backup circulating water pump and adjust the valve openings of relevant pipelines, while partially or completely transferring the control authority of the deep reinforcement learning strategy network to a pre-set, PID-based conservative temperature control loop. The switching process is smooth; that is, when executing the switch, the current system state is considered to ensure that control commands do not abruptly change, avoiding secondary impacts on the greenhouse environment and turtles, thus smoothly switching from the current optimal cooperative control strategy to the corresponding degraded control mode.

[0122] In summary, this invention combines the behavioral feature vectors of turtles obtained from thermal infrared imaging with the photothermal characteristics regulation of spectrally selective photovoltaic modules, and uses this to drive a high-fidelity digital twin model for dynamic simulation and prediction. This model system can proactively simulate the dynamic evolution of the coupled system of photovoltaic-thermal environment-aquaculture water-turtle organisms, providing crucial information for subsequent collaborative optimization decisions. Based on predictions, a deep reinforcement learning strategy network is used for multi-objective, multi-constraint rolling optimization, automatically generating optimal collaborative control sequences that satisfy multiple objectives such as energy utilization, temperature stability, and biological welfare. Through online learning and anomaly diagnosis mechanisms, the system can continuously self-correct, ensuring the continuous effectiveness and operational reliability of regulation.

[0123] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A smart temperature control system for greenhouse turtle farming based on photovoltaic energy, characterized in that: This includes a preprocessing module that collects real-time environmental data, energy data, water quality data, and thermal infrared video streams of turtle population activity within the greenhouse, and performs preprocessing on these data. The extraction module extracts behavioral feature vectors of turtles that characterize feeding activities and group distribution based on the preprocessed thermal infrared video stream of turtle population activity. The adjustment module adjusts the thermal characteristics of the spectral selective photovoltaic modules at the top of the greenhouse based on the turtle behavior feature vector and real-time environmental data. The prediction module inputs turtle behavior feature vectors, water quality data, and thermal characteristic parameters of spectrally selective photovoltaic modules into a high-fidelity digital twin model to predict the operating status of the environment and temperature intelligent control unit in future time periods. The optimization module, based on the prediction results of the digital twin model, uses a deep reinforcement learning policy network for rolling optimization to generate the optimal cooperative control sequence; The learning module parses and distributes the optimal collaborative control sequence as driving instructions for the underlying temperature control and execution devices, monitors the effect of the driving instructions, and feeds back the actual operation data to the digital twin model and deep reinforcement learning strategy network for online learning. When the temperature intelligent control unit deviates from its operation, it performs anomaly diagnosis and switches to degraded control mode.

2. The intelligent temperature control system for greenhouse turtle farming based on photovoltaic energy as described in claim 1, characterized in that: The preprocessing includes, The collected environmental data, energy data, water quality data, and thermal infrared turtle activity video streams were standardized to obtain pre-processed environmental data, energy data, water quality data, and thermal infrared turtle activity video streams.

3. The intelligent temperature control system for greenhouse turtle farming based on photovoltaic energy as described in claim 2, characterized in that: The extracted turtle behavioral feature vectors representing feeding activities and group distribution include... By using the preprocessed thermal infrared video stream of turtle group activity, the thermal image analysis branch is used to extract the thermal contour region of the turtle in a single frame image, and the temporal analysis branch is used to analyze the motion characteristics of the thermal contour region between consecutive frames to obtain the thermal motion trajectory sequence of the group. Based on the population thermal motion trajectory sequence, the feeding activity index is calculated, and a population distribution thermal density map is generated. The feeding activity index, the thermal density map of the population distribution, and the temporal change features extracted from the thermal motion trajectory sequence are fused and encoded to form a turtle behavior feature vector that characterizes feeding activity and population distribution.

4. The intelligent temperature control system for greenhouse turtle farming based on photovoltaic energy as described in claim 3, characterized in that: The thermal characteristics of the photovoltaic module at the top of the greenhouse that adjusts the spectral selectivity include: The population distribution thermal density map, combined with real-time collected internal thermal environment parameters of the greenhouse, drives the fuzzy inference unit of behavior-thermal environment mapping to calculate the target transmittance adjustment command of the spectrally selective photovoltaic module in the thermal band. Using the target transmittance adjustment command, the electrochromic drive unit of the spectral selective photovoltaic module deployed in the top section of the greenhouse is controlled. By changing the voltage applied to the functional layer, the transmittance of the spectral selective photovoltaic module in the heat band is adjusted to the preset target value. By utilizing the transmittance of the spectrally selective photovoltaic module in the heat band, combined with the total solar irradiance and wind direction and speed in real-time environmental data, a dynamic thermal disturbance compensation model is used to generate power feedforward compensation commands for the active temperature control equipment inside the greenhouse.

5. The intelligent temperature control system for greenhouse turtle farming based on photovoltaic energy as described in claim 4, characterized in that: The operating status of the intelligent environmental and temperature control unit for predicting future time periods includes: By utilizing the thermal characteristic parameter sequence of historical spectrally selective photovoltaic modules, water quality data, and corresponding turtle behavior feature vector sequences representing feeding activities and population distribution, a photovoltaic thermal perturbation-turtle population thermal distribution response coupled subnetwork is constructed through an adversarial training mechanism, and a high-fidelity digital twin model is embedded in it. Real-time behavioral feature vectors of turtles representing feeding activities and group distribution, real-time water quality data, and real-time thermal characteristic parameters of spectrally selective photovoltaic modules are input into a high-fidelity digital twin model that has an embedded photovoltaic thermal disturbance-turtle group thermal distribution response coupling subnetwork, driving the high-fidelity digital twin model to perform online state synchronization. By utilizing a high-fidelity digital twin model that has completed online status synchronization and combining it with future weather forecast data, forward simulation calculations are performed to generate a predicted trajectory of the operating status of the intelligent environmental and temperature control unit for future periods.

6. The intelligent temperature control system for greenhouse turtle farming based on photovoltaic energy as described in claim 5, characterized in that: The optimal cooperative control sequence includes: Based on the predicted trajectory of the operating status of the intelligent environmental and temperature control unit in the future time period, combined with the real-time turtle behavior feature vector representing feeding activities and group distribution and real-time water quality data, a dedicated state of the deep reinforcement learning strategy network is constructed. By utilizing the dedicated state of a deep reinforcement learning policy network, a deep reinforcement learning policy network with a hierarchical constraint satisfaction mechanism is driven to select a subset of feasible actions in the action space that satisfy the lower limit constraint of photovoltaic power generation. By utilizing the subset of feasible actions that satisfy the lower limit constraint of photovoltaic power generation, a deep reinforcement learning strategy network with a hierarchical constraint satisfaction mechanism is driven to screen out the subset of feasible actions that satisfy the hard constraints of aquaculture water temperature and water quality indicators. By utilizing a subset of feasible actions that satisfy the lower limit constraint of photovoltaic power generation and the hard constraints of aquaculture water temperature and water quality, a deep reinforcement learning policy network with a hierarchical constraint satisfaction mechanism is driven to calculate the comprehensive evaluation value of each feasible action and select the action with the highest comprehensive evaluation value. The action with the highest comprehensive evaluation value selected by the deep reinforcement learning strategy network that satisfies the hierarchical constraint conditions is extended into the optimal collaborative control sequence of the thermal characteristic target value sequence and the temperature setpoint sequence of the temperature intelligent control unit.

7. The intelligent temperature control system for greenhouse turtle farming based on photovoltaic energy as described in claim 6, characterized in that: The driving instructions for the underlying temperature control and execution device include, Extract the target value of the thermal characteristics of the spectral selective photovoltaic module and the temperature setpoint of the temperature intelligent control unit at the current decision step from the optimal cooperative control sequence, and generate atomic control instructions with high-precision execution timestamps. By using atomic control instructions with high-precision execution timestamps, and through an instruction conflict arbitrator that predicts response delay, the timing of atomic control instructions with differences in device response speed or physical coupling conflicts is finely adjusted and rearranged to obtain a conflict-free time-stamped drive instruction queue. The conflict-free, time-stamped drive command queue is synchronously sent to the underlying temperature control and execution devices corresponding to the electrochromic drive unit and the temperature intelligent control unit of the spectral selective photovoltaic module, according to the time stamp.

8. The intelligent temperature control system for greenhouse turtle farming based on photovoltaic energy as described in claim 7, characterized in that: The effects of the monitoring drive commands include: After the underlying temperature control and execution equipment executes the drive command, it collects the actual status feedback signal returned by the underlying temperature control and execution equipment, the real-time readings of the environmental sensors in the greenhouse, and the real-time data of the water quality sensors to generate actual operating data that reflects the immediate effect of the drive command.

9. The intelligent temperature control system for greenhouse turtle farming based on photovoltaic energy as described in claim 8, characterized in that: The online learning includes, By utilizing real-time readings from environmental sensors, water quality data, and corresponding drive commands from actual operational data, the digital twin model is driven to perform online incremental parameter correction, dynamically adjusting the physical parameters and dynamic water quality parameters in the digital twin model that affect prediction accuracy. By using the digital twin model after online incremental parameter correction, the state transition process corresponding to the actual operating data is re-evaluated, and an enhanced transition trajectory is generated. The enhanced transfer trajectory is input into the online learning pipeline of the deep reinforcement learning policy network, triggering an asynchronous incremental update of the policy network parameters by a dual-value network based on a competitive architecture.

10. The intelligent temperature control system for greenhouse turtle farming based on photovoltaic energy as described in claim 9, characterized in that: The process of performing abnormal diagnosis and switching to degraded control mode includes, When the temperature intelligent control unit deviates from its operation, the actual status feedback signal of the actuator in the temperature intelligent control unit, the real-time reading of the environmental sensor in the greenhouse and the water quality sensor data are collected. The actual status feedback signal, the real-time reading of the environmental sensor and the water quality data are compared with the expected state of the optimal collaborative control sequence to generate a deviation analysis report on the deviation pattern and trend. Input the deviation analysis report of deviation patterns and trends into the anomaly diagnosis engine of the knowledge graph to locate the root cause of the deviation in the operation of the temperature intelligent control unit. Based on the root cause identified by the anomaly diagnosis engine, the index's pre-defined multi-level fault-tolerant rule base smoothly switches from the current optimal collaborative control strategy to the corresponding degraded control mode.

Images