Multivariate multi-output LSTM for time series forecasting with intermittent demand patterns

Compared to other data sources, demand time series are easily available in the industrial context. Providing accurate forecasts based on this kind of data for components with intermittent demand patterns is thus fundamental in many applications. Examples include optimizing inventory levels and the selection of the best tradeoff between holding and stockout costs in the spare parts management context. Recently, deep learning and machine learning models have been proposed to address this need. Compared to the more traditional ones, these methods better model nonlinear patterns in data. On the other hand, they require more effort in the parameter tuning phase, making it difficult to optimize at an item level in real life. In addition, relying only on single time series has some limitations. This study proposes a new approach based on a multivariate multi-output long short-term memory neural network to reduce time spent tuning and capturing interactions between different items' consumption data. The model is tested on a real spare parts dataset of a mechanical company. Croston's method and its variations, together with a multi-layer perceptron neural network, are used to compare the results.