Improving Day-Ahead Market Price Forecasting through a Spatially Informed Clustering Framework

Day-Ahead Market (DAM) price depends on demographic, social, economic and meteorological variables that are often used as features in price forecasting. Among them, meteorological forecasts are generally disregarded, proxied, or included indirectly. To the best of our knowledge, studies rarely include weather forecasts directly in a DAM price forecaster. Nevertheless, these works use linear models with trivial spatial sampling based on k-means clustering, opening a new research direction on more complex models and more informative sampling methods.

To address this gap, we propose a novel spatially informed clustering framework for weather forecast sampling. Rather than using mere clustering, we apply a state-of-the-art spatial clustering method, using the factors influencing the DAM price while imposing geographical constraints on clusters. Once the clusters are defined, meteorological forecasts are sampled on a predefined mesh and averaged over the same clusters, to retain as much information as possible without overfitting the model.

We then create an experiment campaign aiming at i) assessing the effect of the proposed spatial sampling, regardless of the specific model, and ii) identifying the best-performing model over a set of multi-step forecasters. We compare our approach, using two years of hourly prices from the northern bidding zone in Italy, with three other sampling methods: no weather forecasts, proxy (single location) and baseline k-means clustering. First, we find that including weather forecasts improves DAM price forecasting accuracy for each model. Second, combining our sampling method with any of the proposed models outperforms simple baselines such as naive persistence. Third, our sampling method exhibits the highest accuracy on the best-performing model.

In conclusion, the proposed spatially informed clustering framework improves DAM price-forecasting accuracy and consistently outperforms alternative strategies when coupled with multi-step forecasters. Future research will investigate its generalizability, to evaluate the proposed method across different bidding zones.