The increasing frequency of anomalous climatic events exposes energy utilities to significant volumetric risk, particularly through revenue shortfalls during mild winters. To hedge this exposure, firms rely on Over-the-Counter weather derivatives, such as Heating Degree Day (HDD) options, structured as weighted spatial baskets that reflect distributed customer loads. Pricing and risk-managing these multi-site contracts remains challenging due to spatial dependence structures in temperature dynamics. Existing pricing frameworks fall into two categories. First, modelling each geographical location independently ignores the spatial correlation of weather phenomena. This creates a false sense of portfolio diversification, cancelling out local shocks and underestimating portfolio risk. Second, to manage spatial dependence, the literature relies on linear techniques driven by empirical covariance matrices. While mathematically convenient, these linear transformations struggle to capture asymmetric tail dependencies and climatic extremes, implying theoretical hedging baskets that are untradable in illiquid weather markets. These frameworks are bound to discrete meteorological stations or fixed grids, inducing basis risk when evaluating portfolios tied to off-grid locations, highlighting an unmet need for continuous spatial representations. To address these limitations, we propose a neural network framework designed to price temperature derivatives under interdependent risks. Our network architecture performs non-linear dimensionality reduction, compressing the high-dimensional field into latent risk factors. By utilising a continuous spatial approach, we map these latent factors to any geographic coordinate, solving the spatial discretisation problem. This design preserves explainability essential for risk management, while capturing asymmetric spatial interactions and tail dependencies that linear covariance models miss. Empirical analysis on NASA MERRA-2 temperature data demonstrates that our interpretable approach captures spatial risk dependence more accurately than linear models. This leads to improved pricing stability, reliable portfolio Value-at-Risk estimates, and hedging performance for multi-site HDD options, highlighting the critical role of explainable non-linear spatial modelling in climate-related financial risk.