Comparative Analysis of ARIMA, LSTM, and TCN for Time Series Forecasting

The review summarizes ARIMA under the Box–Jenkins framework, LSTM as a gated recurrent architecture for long-term dependencies, and TCN as a causal, dilated convolutional model with residual learning. Overall, LSTM/TCN generally show advantages under strong nonlinearity and long-dependency settings, while ARIMA can remain competitive for short-horizon and relatively stable series. The findings support hybrid architectures and feature selection as promising directions for improving robustness and practical deployment. In power load prediction, studies have also shown that by optimizing the input lag term and network structure through feature selection and genetic algorithms, the prediction performance of LSTM can be further improved. However, it also reflects that the effect of deep models is highly dependent on structure selection and hyperparameter setting, requiring systematic parameter tuning or search strategies.