Attention-Infused Autoencoder for Feature Extraction withLSTM for Time Series Forecasting

In this paper, we propose a novel approach for time series forecastingby integrating an Attention-Infused Autoencoder with an LSTM (LongShort-Term Memory) network. The model first utilizes the Autoencoderfor efficient feature extraction from raw time series data, leveraging theattention mechanism to enhance the extraction process by focusing onthe most relevant parts of the input sequence. This attention-basedfeature extraction allows the model to capture complex dependencies inthe data more effectively, especially in the presence of noisy orincomplete information. Once the features are extracted, they are feddirectly into the LSTM, which learns the temporal patterns anddynamics in the sequence. The LSTM then makes accurate predictionsfor future time steps based on the enriched feature set. This hybridapproach combines the strengths of both autoencoders for unsupervisedfeature learning and LSTMs for sequential modeling, resulting in arobust model capable of handling both short-term and long-termdependencies in time series forecasting tasks. Experimental results on apower consumption dataset demonstrate the superior performance of theproposed model over traditional forecasting methods, highlighting itspotential for accurate predictions in real-world applications such asenergy consumption forecasting, as well as its applicability to otherdomains like financial markets and traffic forecasting

کلیدواژه ها:

LSTM ، Attention ، Autoencoder ، Time Series