We propose here a novel neural architecture dedicated to the prediction of time series. It can be considered as an adaptation of the idea of (GQN) to the data which is of a sequence nature. The new approach, dubbed here as the (RGQN), allows for efficient prediction of time series. The predictor information (i.e. the independent variable) is one or more of the other time series which are in some relationship with the predicted sequence. Each time series is accompanied by additional meta-information reflecting its selected properties. This meta-information, together with the standard dynamic component, is provided simultaneously in (RNN). During the inference phase, meta-information becomes a query reflecting the expected properties of the predicted time series. The proposed idea is illustrated with use cases of strong practical relevance. In particular, we discuss the example of an industrial pipeline that transports liquid media. The trained RGQN model is applied to predict pressure signals, assuming that the training was carried out during routine operational conditions. The subsequent comparison of the prediction with the actual data gathered under extraordinary circumstances, e.g. during the leakage, leads to a specific residual distribution of the prediction. This information can be applied directly within the data-driven Leak Detection and Location framework. The RGQN approach can be applied not only to pressure time series but also in many other use cases where the quantity of sequence nature is accompanied by a meta-descriptor.
Keywords: Time-series; generative model; neural networks.