Temporal hierarchies are being increasingly used for forecasting purposes over the past years. They have been shown to produce accurate and coherent forecasts which are beneficial for enterprises. Reconciling forecasts of different aggregation levels to achieve coherence, supports aligned decisions between different organizational levels. Current research focuses on analytical reconciliation methods which have shown to be more beneficial than conventional Bottom-Up and Top-Down approaches. However, such methods rely on a number of assumptions, primarily due to estimation requirements. This work proposes a novel approach for forecasting with temporal hierarchies. It results in a non-linear reconciliation method inspired by the architecture of an encoder-decoder deep neural network. A trainable encoder combines base forecasts into the reconciled bottom level predictions, while a fixed, non-trainable decoder reconstructs the forecasts across all hierarchical levels. Two different reconciliation architectures are presented based on different optimization procedures. They both ensure coherence. This thesis suggests two alternative usages for the reconcilers. One, to replace analytical expressions and reconcile base forecasts produced by models such as Exponential Smoothing. Second, as a part of a deep neural architecture DTH-28, which mimics the general framework for forecasting with temporal hierarchies. The proposed framework outperforms established benchmarks on real data. Furthermore, this work discusses the general effect of coherence on forecast accuracy. Coherence affects accuracy in two ways. One as a regularizer and second as a stepwise function. Exploiting each usage offers different accuracy benefits.