A Combined Model for Cherry Greenhouse Temperature Prediction Based on LMD and Attention Mechanism

Abstract

Temperature is the most important environmental factor affecting cherry growth in greenhouses. Temperature prediction with past temperature in greenhouse is essential. Existing solutions include traditional machine learning algorithms and neural network algorithms. Nonstationary data can significantly affect accuracy of the current well-performing algorithms, such as Long Short-Term Memory (LSTM) and Gated Recurrent Unit networks (GRU). Using Local Mean Decomposition (LMD) to decompose time series can reduce the non-stationarity of the data. However, it suffers from the problem of over-smoothing so that a large amount of information is lost in the smoothing process. Therefore, this paper proposes a time series and multi-frequency combined network model based on an improved LMD and attention mechanism. Firstly, strengthen the LMD algorithm with the cubic spline interpolation to weaken the smoothing strength. Secondly, the modal decomposition of the time series is performed. Thirdly, the original time series is input into the LSTM and the decomposed frequency components are input into the GRU. Finally, we use the self-attention mechanism to extract the most useful information to improve prediction accuracy. After the experiment, compared with Recurrent Neural Network (RNN), GRU and Bi-directional Gated Recurrent Unit networks (Bi-GRU), the Mean Absolute Error (MAE) of the model in this study were reduced by 0.38, 0.53, and 0.31; the Root Mean Square Error (RMSE) were reduced by 0.46, 0.61, and 0.3; the Mean Absolute Percentage Error (MAPE) were reduced by 3.51%, 4.95%, and 3.33%; and the R <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> were reduced by 0.098, 0.127, and 0.069, respectively.