<p>With the increasing importance of the methane gas in the context of the energy transition and the climate change mitigation, accurate forecasting of spot price of natural gas plays a crucial role in commercial policies and strategies, as well as in cost-effective market opportunities. Recently, many machine learning methodologies were used to model the gas spot price, such as artificial neural networks (ANNs)and their combination with traditional time series approaches, such as auto-eegressive moving average (ARMA) models, were implemented to boost forecasting accuracy. In this context, a new hybrid model, which integrates geostatistical tools and ANNs (i.e. the long short-term memory and the gated recurrent unit neural networks), is proposed in order to suitably reproduce the linear and non-linear components characterizing the dynamics under study. This innovative model on one hand improves time series kriging prediction, through the ANNs contribution, and on the other hand takes advantage of the use of the variogram, as a measure of temporal relationship instead of the covariance function. In addition, the supremacy of the provided hybrid model is assessed through a comparison with respect to various contender models, which goes from some baseline options (belonging to the ARMA class or the pure ANNs or the non-neural network, like the Random Forest) to another well-known hybrid model available in the literature. These models are applied to the spot price of natural gas in Italy, given by the Virtual Trading Point (VTP).</p>