A shed factory microclimate prediction model construction method based on big data analysis

Abstract

The application discloses a shed factory microclimate prediction model construction method based on big data analysis, aiming to solve the defects of single data dimension, conventional algorithm and weak generalization ability of the existing model. The method first synchronously collects conventional environment, implicit association, production activities and historical time sequence multi-source heterogeneous data, and after preprocessing such as cleaning, normalization and time sequence alignment, extracts core features by using an improved sparse local linear embedding algorithm, constructs an improved autonomous echo state network prediction model fused with a snow ablation optimizer, and outputs prediction results combined with a sliding window dynamic optimization and error feedback correction mechanism. The application has comprehensive data coverage, strong algorithm innovation, high prediction accuracy and excellent generalization ability, can be adapted to various shed factory scenes, provides reliable support for accurate environmental regulation, and has outstanding practicality.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of environmental control and big data analysis in greenhouse factories, and particularly to a method for constructing a microclimate prediction model for greenhouse factories based on big data analysis. Background Technology

[0002] The microclimate of greenhouse factories is a core factor influencing their internal production activities (such as crop growth, product storage, and equipment operation). Accurate prediction of microclimate parameter trends is crucial for optimizing control strategies, reducing energy consumption, and improving production efficiency. Existing greenhouse factory microclimate prediction technologies mainly suffer from the following shortcomings: Single data dimension: Most prediction models only collect conventional environmental parameters such as temperature, humidity and light, ignoring implicit influencing factors such as greenhouse structure stress, soil microbial metabolic intensity and surrounding dynamic shading, resulting in poor adaptability of prediction models to complex scenarios; Insufficient algorithm innovation: Conventional time series prediction algorithms such as LSTM, BP neural network, and ARIMA are commonly used, which have limited ability to capture the nonlinear and chaotic characteristics of microclimate parameters and are prone to overfitting and long-term prediction error accumulation. Poor technical coherence: The steps of data collection, preprocessing, feature extraction, and model building lack causal correspondence, and no adaptive technical means are designed for the special characteristics of implicit parameters, resulting in the model's prediction accuracy and robustness failing to meet the needs of practical applications. Weak generalization ability: Existing models are mostly designed for specific types of greenhouse factories (such as solar greenhouses) and do not take into account the differences in different greenhouse structures and production scenarios, making it difficult to quickly adapt to the microclimate prediction needs of various greenhouse factories.

[0003] To address the aforementioned technical shortcomings, there is an urgent need for a method to construct a microclimate prediction model for greenhouse factories that features comprehensive data dimensions, novel algorithms, and strong generalization capabilities. This would help overcome the deficiencies of existing technologies and improve the accuracy and practicality of microclimate prediction. Summary of the Invention

[0004] The purpose of this invention is to overcome the technical shortcomings of existing microclimate prediction models for greenhouse factories, such as limited data dimensions, conventional algorithms, and weak generalization ability. It provides a method for constructing microclimate prediction models for greenhouse factories based on big data analysis. This improves the accuracy, robustness, and scenario adaptability of microclimate parameter predictions, providing reliable data support for environmental control in greenhouse factories.

[0005] The technical solution of this invention is implemented as follows: A method for constructing a microclimate prediction model for greenhouse factories based on big data analysis includes the following steps: Step 1: Multi-source heterogeneous data acquisition, synchronously collecting routine environmental parameters, implicit correlation parameters, production activity parameters and historical time series parameters of the shed factory to form the original dataset; Step 2: Multi-dimensional data preprocessing, cleaning, normalizing and time-series alignment of the original dataset to obtain a clean dataset; Step 3: Based on the feature extraction of improved sparse local linear embedding, construct a delayed embedding space, adaptively select the neighborhood, and reduce the dimensionality of high-dimensional data through sparse reconstruction to obtain the core feature set; Step 4: Construct a prediction model based on the improved autonomous echo state network, integrate the snow ablation optimizer to optimize the model hyperparameters, adapt to the chaotic characteristics of microclimate parameters, and build a prediction model. Step 5: Model training and dynamic optimization. Divide the dataset and train the model. Update the training set periodically and fine-tune the parameters using a sliding window mechanism. Step 6: Prediction and Feedback Correction. Input real-time feature data to predict microclimate, correct model parameters based on prediction error feedback, and output the corrected prediction results.

[0006] Preferably, the implicit correlation parameters in step 1 include the structural stress of the shed, the metabolic intensity of soil microorganisms, and the surrounding dynamic shading data; the structural stress of the shed is collected by a strain gauge sensor, the metabolic intensity of soil microorganisms is indirectly characterized by CO2 release collected by a soil microorganism sensor, and the surrounding dynamic shading data is collected and identified by a visual sensor.

[0007] Preferably, in step 2, the data cleaning uses a density-based outlier detection algorithm to remove abnormal data, and missing data is filled using linear interpolation; the data normalization uses Z-score standardization for common environmental parameters and min-max normalization for implicit correlation parameters.

[0008] Preferably, in step 3, the neighborhood size of the improved sparse local linear embedding algorithm is dynamically adjusted according to the local density of the data. When the local density is high, the neighborhood radius is reduced, and when the local density is low, the neighborhood radius is expanded. The sparse reconstruction weights are solved by L1 regularization constraints.

[0009] Preferably, the reservoir of the improved autonomous echo state network in step 4 adopts a sparse random matrix with a sparsity of 0.85-0.95; the snow ablation optimizer is used to optimize the reservoir spectral radius, reservoir size and input weight hyperparameters without the need for manual setting of initial values.

[0010] Preferably, in step 4, a Lyapunov time threshold is introduced to quantify the chaotic characteristics of microclimate parameters and dynamically adjust the update frequency of the model reservoir.

[0011] Preferably, the dataset in step 5 is divided into a training set, a validation set, and a test set in a ratio of 7:2:1; the model training uses gradient descent to minimize the mean square error, with a convergence threshold ≤0.001; the sliding window size is set to 72 hours, and the training set is updated periodically.

[0012] Preferably, the prediction error in step 6 is characterized by mean absolute percentage error. When the mean absolute percentage error is greater than 5%, a feedback mechanism is triggered to adjust the neighborhood radius parameter in step 3 and the reserve pool update frequency in step 4.

[0013] Preferably, the conventional environmental parameters mentioned in step 1 include indoor temperature, humidity, light intensity, CO2 concentration, wind speed, air pressure, and outdoor temperature, humidity, wind direction, and precipitation intensity; the production activity parameters include the operating status of the indoor environmental control equipment and the crop growth cycle or product storage density.

[0014] Preferably, the greenhouse factory includes agricultural planting greenhouses, industrial storage greenhouses, and seedling greenhouses; the time delay τ in step 3 is adaptively adjusted according to the type of greenhouse factory, and the value range is 2-4 hours.

[0015] The embodiments of the present invention have the following advantages due to the adoption of the above technical solutions: Data dimension innovation: Introducing implicit parameters such as greenhouse structure stress and soil microbial metabolic intensity, combined with conventional parameters and production activity parameters, to comprehensively cover microclimate influencing factors and solve the shortcomings of existing technology data being singular; The algorithm is highly novel: it adopts an improved SLLE feature extraction algorithm and an improved AESN prediction algorithm, both of which are niche applications in the field of microclimate prediction for greenhouse factories. Furthermore, the algorithm improvements have solved the inherent defects of traditional technologies, demonstrating outstanding creativity. High prediction accuracy: With the support of multi-dimensional data, accurate feature extraction and adaptive model optimization, the prediction MAPE is ≤5%, which is 15%-30% higher than the existing LSTM and BP models, and the long-term prediction error accumulation is significantly reduced. Strong generalization ability: Through hyperparameter adaptive optimization and dynamic feedback correction, it can be adapted to various shed and factory scenarios such as agricultural planting sheds and industrial storage sheds, without the need to remodel for specific scenarios, making it highly practical.

[0016] The above overview is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features of the invention will become readily apparent from the accompanying drawings and the following detailed description. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the technical solution of the present invention; Figure 2 This is a schematic diagram of the improved AESN prediction model structure of the present invention; Figure 3 This is a multi-source heterogeneous data acquisition dimension diagram of the present invention; Figure 4 This is a diagram illustrating the data preprocessing steps of the present invention; Figure 5 This is the prediction-feedback correction closed-loop diagram of the present invention. Detailed Implementation

[0019] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0020] It is important to note that terms such as "first," "second," "symmetric," and "array" are used only to distinguish between descriptive and positional descriptions and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, features specified with terms such as "first" or "symmetric" may explicitly or implicitly include one or more of that feature; similarly, when the quantity of certain features is not limited by words such as "two" or "three," it should be noted that such features also explicitly or implicitly include one or more features. In this invention, unless otherwise explicitly specified and limited, terms such as "installation," "connection," and "fixation" should be interpreted broadly; for example, they can refer to a fixed connection, a detachable connection, or an integral molding; they can refer to a mechanical connection, a direct connection, a welding connection, or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the accompanying drawings and specific circumstances.

[0021] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0022] like Figure 1-5As shown, this invention provides a method for constructing a microclimate prediction model for greenhouse factories based on big data analysis, comprising the following steps: Step 1: Multi-source heterogeneous data acquisition Multiple types of sensors and data acquisition devices were deployed inside and around the factory to simultaneously collect four types of microclimate-related data, forming a raw dataset. (1) Conventional environmental parameters: temperature, humidity, light intensity, CO2 concentration, wind speed, air pressure inside the greenhouse; temperature, humidity, wind direction, and precipitation intensity outside the greenhouse; (2) Implicit correlation parameters: structural stress of the shed (collected by strain gauge sensors to reflect the impact of shed deformation on ventilation and lighting), soil microbial metabolic intensity (indirectly characterized by CO2 release collected by soil microbial sensors to reflect implicit changes in soil temperature and humidity), and surrounding dynamic shading data (collected by visual sensors to identify shading factors such as temporary structures and tree swaying). (3) Production activity parameters: operating status of equipment in the greenhouse (fans, shade nets, humidifiers, etc.), crop growth cycle (agricultural greenhouse) / product storage density (industrial greenhouse); (4) Historical time series parameters: Historical monitoring data of the above three types of parameters in the past 3-5 years, organized at the hourly level.

[0023] Collecting implicit correlation parameters and production activity parameters is to address the problem that existing technologies have limited data dimensions and cannot capture the complex influencing factors of microclimates, thus providing a comprehensive data foundation for subsequent accurate modeling; historical time series parameters, on the other hand, provide support for the model to learn the temporal change patterns.

[0024] Step 2: Multi-dimensional data preprocessing To address the heterogeneity and noise characteristics of the original dataset, a three-step preprocessing process is performed to obtain a clean dataset: (1) Data cleaning: Density-based outlier detection (LOF) algorithm is used to remove abnormal sensor data (such as vibration interference data of stress sensor and sudden noise of microbial sensor), and missing data is filled by linear interpolation (missing rate ≤5%). (2) Data normalization: Differential normalization is performed on parameters of different dimensions. Conventional environmental parameters are normalized using Z-score, while implicitly correlated parameters (such as stress data) are normalized using min-max (to adapt to their narrow value range and small fluctuation). (3) Data time alignment: Using hours as the time granularity, align various parameters according to timestamps to construct a time series dataset to ensure the time consistency of subsequent feature extraction and model training.

[0025] Noise and missing values ​​in the raw data can seriously affect the accuracy of the model, and the different parameters have large differences in dimensions. Therefore, it is necessary to eliminate interference and unify the dimensions through preprocessing. Temporal alignment is to match the temporal variation characteristics of microclimate parameters and lay the foundation for subsequent temporal feature extraction.

[0026] Step 3: Feature extraction based on improved sparse local linear embedding To address the redundancy issue in preprocessed high-dimensional datasets, an improved Sparse Local Linear Embedding (SLLE) algorithm is employed to extract key features. The specific process is as follows: (1) Constructing a delayed embedding space: Construct a high-dimensional feature space for the time series dataset according to the time delay τ (τ=2-4 hours, adaptively adjusted according to the type of shed factory) to capture the temporal correlation of parameters; (2) Adaptive neighborhood selection: The neighborhood size is dynamically adjusted according to the local density of the data (the neighborhood radius shrinks when the local density is high and expands when the density is low), which solves the problem of inaccurate feature extraction caused by the fixed neighborhood size in traditional SLLE; (3) Sparse reconstruction and dimensionality reduction: The sparse reconstruction weights are solved by L1 regularization constraints. Only a few key neighborhood points are retained to participate in the reconstruction. Redundant features are eliminated, and high-dimensional data is mapped to low-dimensional feature space to obtain the core feature set.

[0027] The preprocessed dataset has high dimensionality and contains redundant information. Conventional dimensionality reduction algorithms are prone to losing key features of implicitly correlated parameters. Therefore, an improved SLLE algorithm is adopted, which can preserve the local nonlinear features and temporal correlation of parameters, and achieve efficient dimensionality reduction, providing high-quality input for model construction.

[0028] Step 4: Constructing a prediction model based on an improved autonomous echo state network Using the core feature set as input, an improved autonomous echo state network (AESN) prediction model incorporating the snow ablation optimizer (SAO) is constructed. The specific design is as follows: (1) Model structure design: including input layer, reservoir, and output layer. The dimension of the input layer matches the dimension of the core feature set. The reservoir adopts a sparse random matrix (sparseness 0.85-0.95). The output layer adopts a linear activation function. (2) Hyperparameter adaptive optimization: The SAO algorithm is used to optimize hyperparameters such as the reservoir radius, reservoir size, and input weight. No initial values ​​need to be set manually, which solves the problem of cumbersome hyperparameter tuning and easy getting trapped in local optima in traditional AESN. (3) Chaotic characteristics adaptation: Lyapunov time threshold is introduced to quantify the chaotic characteristics of microclimate parameters, dynamically adjust the update frequency of the model's reservoir, and improve the model's ability to fit chaotic time series data.

[0029] The core feature set has nonlinear, temporal, and chaotic characteristics, which are difficult for conventional models to fit accurately. The improved AESN model can accurately capture the changing patterns of microclimate parameters through hyperparameter adaptive optimization and adaptation to chaotic characteristics, while solving the problem of long-term error accumulation in traditional prediction algorithms.

[0030] Step 5: Model Training and Dynamic Optimization (1) Data set partitioning: The core feature set is divided into training set, validation set and test set in a ratio of 7:2:1. The training set is used for model parameter learning, the validation set is used for hyperparameter fine-tuning, and the test set is used for model performance evaluation. (2) Model training: The gradient descent method is used to minimize the mean square error (MSE) between the model prediction and the actual value, and the model is trained iteratively until the error converges (convergence threshold ≤ 0.001). (3) Dynamic optimization: Introduce a sliding window mechanism (window size is 72 hours), regularly update the training set (add the latest monitoring data and remove expired data), and readjust the model parameters to ensure that the model adapts to the dynamic changes of the microclimate.

[0031] Model training aims to teach the model the mapping relationship between core features and microclimate parameters, while dynamic optimization addresses the dynamic changes in microclimate parameters due to external environmental and production activities, preventing a decline in model accuracy after long-term operation and ensuring the continuity and reliability of predictions.

[0032] Step 6: Prediction and Feedback Correction (1) Microclimate prediction: Input the real-time collected and preprocessed feature data into the optimized prediction model, and output the prediction results of microclimate parameters (temperature, humidity, light intensity, etc.) for the shed factory in the next 24-72 hours. (2) Error feedback: Calculate the absolute percentage error (MAPE) between the predicted value and the real-time monitored value. When MAPE > 5%, trigger the feedback mechanism to adjust the neighborhood radius parameter in step 3 and the reserve pool update frequency in step 4. (3) Result output: The corrected prediction results are output in a preset format to provide control signal input for the environmental control equipment of the shed factory.

[0033] Prediction is the ultimate goal of model building, while error feedback is to correct the deviation of model parameters in a timely manner, solve the problem of increased prediction error caused by dynamic interference in practical applications, and form a complete closed loop of "collection-preprocessing-extraction-modeling-prediction-correction".

[0034] In this embodiment, the present invention operates as follows: First, various data acquisition devices deployed inside and around the greenhouse were activated, including 10 conventional environmental sensors, 4 strain gauge sensors, 2 soil microbial sensors, 2 visual sensors, and an external weather station. The devices were used to collect four types of parameters at a frequency of 10 minutes per time: temperature and humidity inside the greenhouse, CO2 concentration, greenhouse structural stress, soil microbial CO2 release, and surrounding dynamic shading. The hourly historical monitoring data of the flower greenhouse for the past 3 years were also retrieved to form the original dataset. Subsequently, the original dataset was preprocessed. Abnormal data such as vibration interference from stress sensors and sudden noise from microbial sensors were removed using the LOF algorithm with a neighborhood parameter k=10. Approximately 2.3% of the missing data were filled using linear interpolation. Z-score standardization was performed on the conventional environmental parameters, and min-max normalization (mapped to the [0,1] interval) was performed on the implicit correlation parameters. Then, the time series alignment of all parameters was completed at a 1-hour granularity to obtain a clean time series dataset. Next, feature extraction was performed based on the improved SLLE algorithm. A high-dimensional delayed embedding space was constructed by setting a time delay τ=3 hours. The neighborhood radius (0.1-0.3) was dynamically adjusted according to the local density of the data. The sparse reconstruction weight was solved by the L1 regularization constraint with regularization parameter λ=0.01, and the 28-dimensional high-dimensional data was reduced to an 8-dimensional core feature set (including key features such as temperature, humidity, light, and stress). Next, an improved AESN prediction model incorporating the Snow Ablation Optimizer (SAO) was loaded. This model's input layer dimension was adapted to an 8-dimensional core feature set, the reservoir size was set to 200 (sparseness 0.9), and the output layer dimension was 3 (corresponding to temperature, humidity, and light intensity prediction). Hyperparameters were optimized using the SAO algorithm (50 iterations, 30 population size), and a 48-hour Lyapunov time threshold was introduced to adapt to the chaotic characteristics of the microclimate. During the model's operation, the core feature set was divided into training, validation, and test sets in a 7:2:1 ratio. The model was trained using gradient descent with a learning rate of 0.005 and 1000 iterations until the error converged (convergence value 0.0008). A 72-hour sliding window mechanism was used to update the training set and fine-tune the model parameters every 24 hours. Finally, the feature data is collected and preprocessed in real time and input into the model to output the microclimate prediction results for the next 48 hours in the greenhouse. At the same time, the MAPE of the predicted value and the real-time monitoring value is calculated. When the MAPE is greater than 5%, feedback correction is triggered to adjust the SLLE neighborhood radius and the AESN reservoir update frequency. The corrected prediction results are output in a preset format to provide control signals for the ventilation, shading, humidification and other control equipment in the greenhouse, so as to realize the accurate prediction and adaptive control of the microclimate in the flower planting greenhouse.

[0035] The following are several other specific embodiments of the application of this invention: Example 1: Industrial storage shed (for storing precision electronic components) (a) Implementation preparation This embodiment focuses on a sealed precision electronic component storage shed (80m long, 30m wide, and 6m high). The core requirement of this scenario is to accurately predict the temperature, humidity, and static electricity concentration inside the shed (key microclimate parameters that affect the storage life of components) to avoid temperature and humidity fluctuations and static electricity damage to components.

[0036] Data acquisition equipment deployment: 15 conventional environmental sensors (temperature, humidity, static electricity concentration, CO2 concentration, wind speed, and air pressure) are deployed inside the shed; 6 strain gauge sensors (installed on the load-bearing beams and key stress points of the storage racks to reflect the impact of rack deformation on ventilation); 3 soil microbial sensors (buried 15cm deep in non-storage areas in the corners of the shed to indirectly characterize the latent changes in humidity at the bottom of the shed); and 4 visual sensors (installed on the roof and around the perimeter of the shed to identify dynamic interference factors such as vehicle obstruction and construction dust from the surrounding area); one weather station is deployed outside the shed (collecting temperature, humidity, wind direction, precipitation intensity, wind speed, and dust concentration); equipment operation status data (operation status of air conditioners, dehumidifiers, static eliminators, and ventilation fans) are collected through the warehouse IoT system; historical data uses hourly monitoring data of the shed for the past 4 years (including records of changes in component storage density).

[0037] (II) Implementation Steps 1. Multi-source heterogeneous data acquisition: Four types of parameters are collected simultaneously at a frequency of once every 5 minutes (to meet the high-frequency monitoring requirements of microclimate for precision component storage), for 45 consecutive days, forming the raw dataset. Among them, the conventional environmental parameters include the addition of static electricity concentration inside the shed; the implicit correlation parameters include the stress value of the shed structure ranging from 0-8MPa (the stress range of the storage shed is higher than that of the planting shed due to its greater load-bearing capacity), and the soil microbial metabolic intensity is characterized by CO2 release (mg / (kg·h)); production activity parameters include the operating status of the control equipment inside the shed and the component storage density (recorded by shelf area, unit: pieces / ㎡); historical time series parameters include the correlation data of the above three types of parameters and component loss rate for the past 4 years.

[0038] 2. Multi-dimensional data preprocessing: The LOF algorithm (neighborhood parameter k=12, adapted to the noise characteristics of high-frequency acquisition data) is used to remove outlier data (such as abnormal values ​​of strain gauge sensors caused by vibration during shelf handling, and sudden interference data of electrostatic sensors), and linear interpolation is used to fill in missing data (missing rate of about 1.8%); conventional environmental parameters (temperature, humidity, electrostatic concentration, etc.) are standardized using Z-score, and implicit correlation parameters (stress data, microbial metabolic data) are normalized using min-max (mapped to the [0,1] interval); time series alignment is performed with a granularity of 1 hour (consistent with the granularity of historical data) to obtain a clean time series dataset.

[0039] 3. Feature Extraction: Based on the relatively smooth microclimate fluctuations of industrial storage sheds, a time delay τ = 4 hours is set (the value range is limited to 2-4 hours according to claim 10) to construct a high-dimensional delayed embedding space; the neighborhood radius (0.15-0.35) is dynamically adjusted according to the local density of the data, and the sparse reconstruction weight is solved by L1 regularization constraint (regularization parameter λ = 0.02), reducing the high-dimensional data from 32 dimensions (with the addition of electrostatic concentration and dust-related parameters) to 10 dimensions, and obtaining the core feature set (including key features such as temperature, humidity, electrostatic concentration, shed structure stress, microbial metabolic intensity, storage rack load-bearing capacity, and external dust concentration).

[0040] 4. Model Construction: The AESN model was improved to adapt the input layer dimension to a 10-dimensional core feature set, and the reservoir size was set to 250 (sparseness 0.92, sparsity 0.85-0.95 as specified in claim 5). The output layer dimension was 3 (predicting temperature, humidity, and static electricity concentration). The SAO algorithm was used to optimize the hyperparameters (60 iterations, population size 35), and the Lyapunov time threshold was introduced and set to 60 hours (the chaotic characteristics of the microclimate in industrial greenhouses are weaker than those in planting greenhouses, so the threshold was appropriately increased). The reservoir update frequency was dynamically adjusted.

[0041] 5. Model Training and Dynamic Optimization: The core feature set is divided into training set, validation set, and test set in a 7:2:1 ratio (as specified in claim 7). The model is trained using gradient descent (learning rate 0.004, iterations 1200), and the error converges to 0.0007 (convergence threshold ≤ 0.001, in accordance with claim 7). The sliding window size is set to 72 hours (as specified in claim 7), and the training set is updated every 12 hours (high-frequency updates adapt to the dynamic storage requirements of the warehouse) and the model parameters are fine-tuned.

[0042] 6. Prediction and Feedback Correction: Input real-time monitoring data to predict the microclimate parameters inside the shed for the next 72 hours (adapting to the long-term storage prediction requirements of the warehouse). The initial predicted MAPE is 3.8%. When the predicted humidity MAPE is 5.6% (higher than the 5% threshold, as defined in claim 8), the feedback mechanism is triggered, the SLLE neighborhood radius is adjusted to 0.3, the ASN storage pool update frequency is adjusted to 10 hours / time, and the corrected MAPE is reduced to 4.2%.

[0043] In this embodiment, the invention works as follows: First, all data acquisition devices inside and around the greenhouse are activated, and parameters such as temperature and humidity, static electricity concentration, greenhouse structural stress, soil microbial metabolic intensity, and dust outside the greenhouse are collected synchronously at a frequency of 5 minutes / time. Historical monitoring data and component storage records from the past 4 years are retrieved to form the original dataset. Subsequently, outlier data was removed using the LOF algorithm with k=12, missing data was filled in using linear interpolation, Z-score standardization was performed on the regular parameters, min-max normalization was performed on the latent parameters, and time-series alignment was completed at a 1-hour granularity to obtain a clean dataset. Next, a delayed embedding space is constructed with τ=4 hours, the neighborhood radius is dynamically adjusted and dimensionality is reduced by L1 regularization to obtain a 10-dimensional core feature set; then, the improved AESN model is loaded, the hyperparameters are optimized by the SAO algorithm, and a 60-hour Lyapunov time threshold is introduced to adapt to chaotic characteristics; in the model operation, the dataset is divided into 7:2:1 and trained until the error converges, and the training set is updated every 12 hours through a 72-hour sliding window; Finally, real-time input feature data is used to predict microclimate parameters for the next 72 hours. When MAPE > 5%, feedback correction is triggered. The corrected prediction results are output to equipment such as air conditioners, dehumidifiers, and static eliminators in the greenhouse, enabling precise control of the microclimate for storing precision components.

[0044] (III) Implementation Results In this embodiment, the model prediction accuracy is stable at MAPE≤5% (according to claim 8, the error standard is defined), the temperature prediction error is ≤0.6℃, the humidity prediction error is ≤2.5%, and the static electricity concentration prediction error is ≤50V, which is 48% higher than the existing BP neural network model (MAPE=9.2%). Through precise control, the temperature and humidity fluctuation range inside the shed is controlled within ±0.5℃ and ±2%, respectively, and the static electricity concentration is always below the safety threshold. The loss rate of precision components storage is reduced from the original 3.5% to 0.8%, and the energy consumption of ventilation and dehumidification equipment is reduced by 22%, which verifies the adaptability and superiority of the present invention in industrial storage shed scenarios.

[0045] Example 2: Seedling shed (vegetable seedling cultivation) (a) Implementation preparation This embodiment focuses on a semi-enclosed vegetable seedling shed (40m long, 15m wide, and 4m high). The core requirement is to accurately predict the temperature, humidity, light intensity, and CO2 concentration (key parameters for seedling photosynthesis) inside the shed to ensure a stable seedling growth environment and improve the seedling survival rate.

[0046] Data acquisition equipment deployment: 12 conventional environmental sensors (temperature, humidity, light intensity, CO2 concentration, wind speed, and air pressure, deployed according to seedling bed zones to improve monitoring uniformity), 3 strain gauge sensors (installed at key stress points of the greenhouse arch frame), 4 soil microbial sensors (buried 10cm deep, corresponding to different seedling areas), and 2 visual sensors (installed on the greenhouse roof to identify dynamic factors such as tree obstruction and cloud changes outside the greenhouse); 1 weather station is deployed outside the greenhouse (collecting temperature, humidity, wind direction, precipitation intensity, and light intensity); equipment operation status data (operation status of seedling lights, humidifiers, ventilators, and CO2 generators) are collected through the seedling IoT system; historical data uses hourly monitoring data from the greenhouse over the past 3 years (including seedling growth cycle records for different vegetable varieties).

[0047] (II) Implementation Steps 1. Multi-source heterogeneous data acquisition: Four types of parameters were collected simultaneously at a frequency of once every 8 minutes for 35 consecutive days to form the original dataset. Among them, the conventional environmental parameters focused on key parameters for seedling growth (temperature, humidity, light, CO2 concentration); the implicit correlation parameters included the structural stress of the greenhouse ranging from 0-4 MPa (the load-bearing capacity of the seedling greenhouse is less than that of the storage greenhouse), and the soil microbial metabolic intensity was characterized by CO2 release (mg / (kg·h)) (reflecting implicit changes in soil fertility in the seedling stage); the production activity parameters included the operating status of the seedling equipment in the greenhouse and the growth cycle of vegetable seedlings (recorded by seedling age); and the historical time series parameters included the above three types of parameters and seedling survival rate correlation data for the past three years.

[0048] 2. Multi-dimensional data preprocessing: The LOF algorithm (neighborhood parameter k=11) is used to remove outlier data (such as outliers caused by sudden changes in cloud cover in the light sensor and soil disturbance noise in the microbial sensor), and linear interpolation is used to fill in missing data (missing rate of about 2.1%). Conventional environmental parameters are standardized using Z-score, and implicit correlation parameters are normalized using min-max (mapped to the [0,1] interval, as specified in claim 3). Time series alignment is performed with a granularity of 1 hour to obtain a clean time series dataset.

[0049] 3. Feature Extraction: Based on the characteristic of the microclimate of the seedling shed fluctuating with the seedling growth cycle, a time delay τ=2 hours is set (the value range is limited to 2-4 hours in claim 10 to adapt to the rapid response requirements of seedling growth), and a high-dimensional delayed embedding space is constructed; the neighborhood radius (0.08-0.28) is dynamically adjusted according to the local density of the data, and the sparse reconstruction weight is solved by L1 regularization constraint (regularization parameter λ=0.008), reducing the high-dimensional data from 26 dimensions to 7 dimensions, and obtaining the core feature set (including key features such as temperature, humidity, light intensity, CO2 concentration, shed structure stress, soil microbial metabolic intensity, and seedling age).

[0050] 4. Model Construction: The AESN model was improved to adapt the input layer dimension to a 7-dimensional core feature set, and the reservoir size was set to 180 (sparseness 0.88, as specified in claim 5). The output layer dimension was 4 (predicting temperature, humidity, light intensity, and CO2 concentration). The SAO algorithm was used to optimize the hyperparameters (45 iterations, 28 population size). The Lyapunov time threshold was introduced and set to 40 hours (the microclimate of the seedling shed has strong chaotic characteristics, so the threshold was appropriately reduced). The reservoir update frequency was dynamically adjusted.

[0051] 5. Model Training and Dynamic Optimization: The core feature set is divided into training set, validation set, and test set in a 7:2:1 ratio (as defined in claim 7). The model is trained using gradient descent (learning rate 0.006, number of iterations 900), and the error converges to 0.0009 (convergence threshold ≤ 0.001). The sliding window size is set to 72 hours, and the training set is updated every 20 hours (to adapt to the changes in the seedling growth cycle) and the model parameters are fine-tuned.

[0052] 6. Prediction and Feedback Correction: Input real-time monitoring data to predict the microclimate parameters in the greenhouse for the next 36 hours (to meet the short-term growth regulation needs of seedlings). The initial predicted MAPE is 3.5%. When the predicted light intensity MAPE is 5.3% (above the 5% threshold), the feedback mechanism is triggered, the SLLE neighborhood radius is adjusted to 0.22, the AESN reservoir update frequency is adjusted to 18 hours / time, and the corrected MAPE is reduced to 4.0%.

[0053] In this embodiment, the invention works as follows: First, various sensors deployed in the greenhouse and the weather station outside the greenhouse are activated to synchronously collect parameters such as temperature and humidity, light intensity, CO2 concentration, greenhouse stress, and soil microbial metabolic intensity at a frequency of 8 minutes / time. Historical monitoring data from the past 3 years and seedling growth cycle records are retrieved to form the original dataset. Subsequently, outlier data was removed using the LOF algorithm with k=11, missing data was filled in using linear interpolation, Z-score standardization was performed on the regular parameters, min-max normalization was performed on the latent parameters, and time alignment was completed at a 1-hour granularity to obtain a clean dataset. Next, a delayed embedding space is constructed with τ=2 hours, the neighborhood radius is dynamically adjusted and dimensionality is reduced by L1 regularization to obtain a 7-dimensional core feature set; then, the improved AESN model is loaded, the hyperparameters are optimized by the SAO algorithm, and a 40-hour Lyapunov time threshold is introduced to adapt to chaotic characteristics; in the model operation, the dataset is divided into 7:2:1 and trained until the error converges, and the training set is updated every 20 hours through a 72-hour sliding window; Finally, real-time input of feature data is used to predict microclimate parameters for the next 36 hours. When MAPE > 5%, feedback correction is triggered. The corrected prediction results are output to devices such as seedling lights, humidifiers, and CO2 generators to achieve precise adaptation and control of the microclimate for vegetable seedling growth.

[0054] (III) Implementation Results In this embodiment, the model prediction accuracy is stable at MAPE≤5%, temperature prediction error≤0.7℃, humidity prediction error≤2.8%, light intensity prediction error≤400lux, and CO2 concentration prediction error≤30ppm, which is 37% higher than the existing LSTM model (MAPE=8.8%). By precisely controlling the microclimate inside the greenhouse, the seedling survival rate of vegetables is increased from 82% to 95%, the seedling cycle is shortened by 7 days, and the energy consumption of seedling lamps and CO2 generators is reduced by 19%, which verifies the reliability and adaptability of the present invention in the seedling greenhouse scenario.

[0055] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in the present invention, and these should all be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for constructing a small climate prediction model for a shed factory based on big data analysis, characterized in that, Includes the following steps: Step 1: Multi-source heterogeneous data acquisition, synchronously collecting routine environmental parameters, implicit correlation parameters, production activity parameters and historical time series parameters of the shed factory to form the original dataset; Step 2: Multi-dimensional data preprocessing, cleaning, normalizing and time-series alignment of the original dataset to obtain a clean dataset; Step 3: Based on the feature extraction of improved sparse local linear embedding, construct a delayed embedding space, adaptively select the neighborhood, and reduce the dimensionality of high-dimensional data through sparse reconstruction to obtain the core feature set; Step 4: Construct a prediction model based on the improved autonomous echo state network, integrate the snow ablation optimizer to optimize the model hyperparameters, adapt to the chaotic characteristics of microclimate parameters, and build a prediction model. Step 5: Model training and dynamic optimization. Divide the dataset and train the model. Update the training set periodically and fine-tune the parameters using a sliding window mechanism. Step 6: Prediction and Feedback Correction. Input real-time feature data to predict microclimate, correct model parameters based on prediction error feedback, and output the corrected prediction results.

2. The method according to claim 1, characterized in that, The implicit correlation parameters mentioned in step 1 include the structural stress of the greenhouse, the metabolic intensity of soil microorganisms, and the surrounding dynamic shading data. The structural stress of the greenhouse is collected by strain gauge sensors, the metabolic intensity of soil microorganisms is indirectly characterized by CO2 release collected by soil microorganism sensors, and the surrounding dynamic shading data is collected and identified by visual sensors.

3. The method according to claim 1, characterized in that, The data cleaning described in step 2 uses a density-based outlier detection algorithm to remove outlier data, and linear interpolation is used to fill in missing data. The data normalization uses Z-score standardization for common environmental parameters and min-max normalization for implicit correlation parameters.

4. The method according to claim 1, characterized in that, In step 3, the neighborhood size of the improved sparse local linear embedding algorithm is dynamically adjusted according to the local density of the data. When the local density is high, the neighborhood radius is reduced, and when the local density is low, the neighborhood radius is expanded. The sparse reconstruction weights are solved by L1 regularization constraints.

5. The method according to claim 1, characterized in that, In step 4, the reservoir of the improved autonomous echo state network adopts a sparse random matrix with a sparsity of 0.85-0.95; the snow ablation optimizer is used to optimize the reservoir spectral radius, reservoir size and input weight hyperparameters without the need for manual initialization.

6. The method according to claim 1, characterized in that, In step 4, a Lyapunov time threshold is introduced to quantify the chaotic characteristics of microclimate parameters and dynamically adjust the update frequency of the model pool.

7. The method according to claim 1, characterized in that, The dataset mentioned in step 5 is divided into training set, validation set and test set in a ratio of 7:2:1; the model training adopts the gradient descent method to minimize the mean square error, and the convergence threshold is ≤0.001; the sliding window size is set to 72 hours, and the training set is updated regularly.

8. The method according to claim 1, characterized in that, The prediction error in step 6 is represented by the mean absolute percentage error. When the mean absolute percentage error is greater than 5%, a feedback mechanism is triggered to adjust the neighborhood radius parameter in step 3 and the reserve pool update frequency in step 4.

9. The method according to claim 1, characterized in that, The conventional environmental parameters mentioned in step 1 include the temperature, humidity, light intensity, CO2 concentration, wind speed, air pressure inside the greenhouse and the temperature, humidity, wind direction, and precipitation intensity outside the greenhouse; the production activity parameters include the operating status of the environmental control equipment inside the greenhouse and the crop growth cycle or product storage density.

10. The method according to claim 1, characterized in that, The greenhouse factories include agricultural planting greenhouses, industrial storage greenhouses, and seedling greenhouses; in step 3, the time delay τ is adaptively adjusted according to the type of greenhouse factory, with a value range of 2-4 hours.

Images