Air pollution is among the leading environmental risk factors for human health. Worldwide, accelerated economic development and urban activity have been accompanied by rising pollutant concentrations. In the European context, fine particulate matter (PM \(_{2.5}\) ), nitrogen dioxide (NO \(_2\) ) and ozone (O \(_3\) ) are the pollutants responsible for most of the premature deaths each year caused by pollutants. Therefore, the prediction of pollutants is of great importance to human health. In this article an encoder-decoder Long Short-Term Memory (LSTM) neural network model is developed in order to forecast the concentrations of PM \(_{2.5}\) , NO \(_2\) , and O \(_3\) one- and 24-hour-ahead. A specific case study is addressed: the pollutants forecasting in Madrid, using observations from 24 monitoring stations. The encoder-decoder LSTM is trained and evaluated across urban and suburban sites to compare location-specific behavior. In addition, a neural network variant is implemented: a multivariate encoder-decoder LSTM model that includes meteorological covariates. Mean absolute error (MAE) and root mean squared error (RMSE) are used to evaluate and compare models, considering a single-layer perceptron model as a reference baseline for results comparison. The proposed encoder–decoder LSTM achieved up to 60% lower MAE than the baseline perceptron model for one-hour forecasts, and still provided 11–17% MAE reductions at 24 h ahead across the three pollutants.