A bidirectional temporal data-driven model for greenhouse environmental variable prediction

As key infrastructure in modern agriculture, glass greenhouses contribute significantly to enhancing crop yield and quality by offering a controllable growth environment. Consequently, accurate prediction of internal environmental parameters is vital for efficient management and precise regulation of crop production. However, greenhouse environmental data often exhibit pronounced non-stationarity and are significantly affected by noise interference, which poses substantial challenges to the accurate prediction of greenhouse environmental conditions. Traditional methods suffer from issues including insufficient suppression of noise interference, inadequate mining of time-series dependency information, and the absence of effective key-weight assignment, resulting in poor model prediction performance and difficulty in meeting the requirements of refined management for modern greenhouses. This study aimed to propose a bidirectional time-series data-driven prediction model for environmental variables in glass greenhouses. Field experiments were conducted at the Tanjiawan Cloud Agriculture Test Base in Zhejiang Province. Relying on the cloud greenhouse integrated data platform, an environmental monitoring network was established in an 80 m×104 m glass greenhouse, constructing a multi-source sensor architecture based on internet of things (IoT) edge computing. Environmental data were collected and aggregated by edge computing nodes, uploaded to base stations via 5G networks, and then forwarded to big data servers for centralized storage and management using a MySQL database. A total of 40417105 pieces of raw observation data were collected. After preliminary data cleaning, considering the nonlinear correlations among multiple variables in the greenhouse, Spearman’s correlation coefficient was used to analyze the correlations between environmental factors and evaluate the potential nonlinear relationships among them. A statistical significance level of P≤0.05 was set, and a correlation coefficient |r|≥0.5 was regarded as having practical relevance. Environmentally significant factors with significant correlations were selected as input features to reduce model input redundancy. Finally, the input feature time series was decomposed into four intrinsic mode function (IMF) components using variational mode decomposition (VMD) modal decomposition to reduce the non-stationarity and noise interference of the series and retain multi-scale feature information. The decomposed IMF feature sequences were input into the bidirectional long short-term memory (BiLSTM) model for prediction. BiLSTM was used to establish the time series features of each IMF component respectively, capture the bidirectional dynamic dependency relationships of the time series through forward and reverse long short-term memory (LSTM) layers. Subsequently, an attention mechanism was introduced to assign key weights to the hidden state vectors output by the BiLSTM, based on the correlation between the sequence data information and the current prediction task. A weighted average calculation was then performed according to this weight distribution, and the time series prediction results of air temperature, air humidity, CO₂ concentration, and light intensity were integrated and output through the fully connected layer. The test results showed that the proposed model exhibited the best prediction performance in the four environmental prediction tasks. Compared with the five control models of LSTM, BiLSTM, empirical mode decomposition-bidirectional long short-term memory (EMD-BiLSTM), complete ensemble empirical mode decomposition with adaptive noise-bidirectional long short-term memory (CEEMDAN-BiLSTM), and variational mode decomposition-bidirectional long short-term memory (VMD-BiLSTM), various indicators of the model were noticeably improved. Among them, the model had the best fitting effect on air temperature and air humidity, with determination coefficients reaching 0.986 and 0.981 respectively. The average determination coefficient (R²) of the four environmental variables reached 0.976, which was an improvement of 0.067, 0.043, 0.033, 0.026, and 0.013 compared with the control models respectively. The average mean absolute percentage error (MAPE) was 2.606%, which was a decrease of 6.279, 5.606, 3.665, 2.493, and 1.810 percentage points compared with the control models respectively. All indicators were better than those of the control models, indicating that this model could notably enhance the accuracy of environmental factor prediction. The research results will help growers more accurately and efficiently grasp the changing trends of key environmental factors in the greenhouse, timely create suitable conditions for crop growth, and thus enhance the efficiency and quality of agricultural production.