Multivariate temperature prediction model based on CNN-BiLSTM and RandomForest

Temperature fluctuations have profound impacts on both human society and the natural environment. However, the diversity of geographical temperature data and the non-linearity and complexity of meteorological phenomena present significant challenges to accurate prediction. Previous studies in temperature prediction have faced certain limitations, such as inadequate consideration of correlations between multiple meteorological factors in some models or limited modeling capability for non-linear and spatiotemporal relationships. To address these shortcomings, we propose a hybrid model based on CNN-BiLSTM and RandomForest for temperature prediction. Relevant meteorological indicators were selected as model inputs using a combined Pearson's correlation and mutual information analysis. Subsequently, utilizing a sequence-to-sequence framework, the CNN-BiLSTM model effectively explores spatiotemporal relationships and feature representations in time-series data as the encoder for sequence modeling. This aids in learning feature representations and sequence relationships in time series data, thus generating abstract representations for each sample in the hidden or output layers, covering key information learned by the model. Feature representations extracted from the CNN-BiLSTM model are aggregated into a feature matrix. Each sample corresponds to a row of the feature vector, with each column representing a specific feature extracted from the CNN-BiLSTM model. Finally, RandomForest, as an ensemble learning method, is capable of handling complex non-linear relationships during both model training and prediction stages, exhibiting good robustness and predictive capability. Utilizing the feature matrix as input data, it is fed into the RandomForest decoder model along with the corresponding target variables, further enhancing the prediction accuracy and stability of the model to obtain the final temperature prediction results. Simulated experiments were conducted using meteorological data from Changsha, Hunan Province, from 2017 to 2021. A significant number of experiments have demonstrated that, compared to the current leading Dliner method, the hybrid model exhibits higher prediction accuracy and stability in capturing daily temperature trends. The mean absolute error and mean squared error were reduced by 35.6 and 57.5 \(\%\) $\%$ , respectively.