Environmental Parameter Prediction and Variable-Frequency Data Acquisition Strategy Optimization for Greenhouses Based on LSTM

To address issues such as data redundancy, acquisition lag, and insufficient prediction accuracy in the collection of environmental parameters within solar greenhouses, a collaborative method combining environmental parameter prediction and variable-frequency acquisition based on Long Short-Term Memory (LSTM) networks was proposed. Taking a solar greenhouse in Chaoyang, Liaoning Province, as the research object, multi-source environmental data both inside and outside the greenhouse were collected, and main influencing factors were screened through correlation analysis. LSTM prediction models for greenhouse temperature, humidity, and CO2 concentration were constructed, and the numbers of hidden layers, learning rate, and prediction horizon of the models were systematically optimized. The results showed that when the number of hidden layers was 3, the training set root mean square error (RMSE) of the temperature, humidity, and CO2 concentration prediction models reached their minimum values of 0.91, 2.21, and 25.00, respectively; when the prediction horizon was set to 5 min, the models exhibited optimal prediction accuracy. Further multi-interval verification showed that within the 1–30 min sampling interval range in the experimental greenhouse, the models demonstrated excellent prediction performance, with coefficients of determination R2 maintaining above 0.85. Among them, the temperature and humidity models exhibited good cross-greenhouse adaptability in an analogous greenhouse, whereas the transferability of the CO2 model was relatively weak. On this basis, a variable-frequency data acquisition strategy explicitly driven by LSTM short-term prediction outputs was developed. Its core logic dynamically adjusts the sampling frequency based on the rate of change between the current measured temperature and the predicted temperature 5 min ahead. Verification results indicated that compared with the 1-min fixed-interval collection, the variable-frequency acquisition strategy reduced the data volume from 453,024 to 296,180 entries—a reduction of 34.6%—while maintaining prediction accuracy, thereby effectively reducing redundant data volume and potentially lowering transmission and storage requirements. The “acquisition–prediction–variable-frequency acquisition” collaborative optimization framework constructed in this study achieves a balance between monitoring accuracy and data utilization efficiency. Compared with traditional fixed-interval or reactive adaptive methods, it possesses systemic advantages such as proactive tracking of abrupt changes and potential reduction in node energy consumption, providing a theoretical basis and technical reference for high-quality environmental data collection and supporting precise environmental management in solar greenhouses.