Better prediction of greenhouse extreme temperature base on improved loss function

Extreme greenhouse temperatures can lead to irreversible damage to crops inside the greenhouse, resulting in yield reduction and even crop failure. Predicting such extreme temperatures and intervening in advance can mitigate the economic losses caused by these conditions. Existing models demonstrate relatively accurate predictions within the normal temperature range of the greenhouse, but they exhibit significant deviations when forecasting extreme temperature intervals, leading to narrow temperature prediction ranges, which hinders their ability to address the aforementioned scenarios effectively. In this paper, we propose a novel approach that combines the weighted idea for handling class imbalance and introduces a loss function suitable for multiple models. By ensuring the accuracy of normal temperature predictions, our proposed method significantly enhances the precision of predicting extreme greenhouse temperatures and expands the model’s temperature prediction range. Experimental results demonstrate the effectiveness of this loss function in various models such as LGB (LightGBM), LSTM (Long Short-Term Memory), and BPNN (Backpropagation Neural Network), leading to a significant reduction in false positive and false negative predictions of extreme temperatures.