Multi-horizon net load forecasting for residential households: evaluating statistical and deep learning models with recursive and direct strategies

• Comparison of multi-horizon net load forecasts using real data from Cypriot PV prosumers. • Univariate models (both statistical and RNN) are effective when meteorological data is unavailable. • LSTM achieves top accuracy for medium-term univariate net load forecasting. • ARIMA and TSLM outperform RNN models when exogenous inputs are included. • Exogenous weather data improves Transformers, harms RNNs in forecasting tasks. Considering the European Union (EU) climate neutrality objectives, the increasing adoption of residential photovoltaic (PV) systems presents new challenges for grid reliability, especially in isolated electricity systems such as in islands. This study evaluates recursive and direct multi-horizon forecasting strategies of household net load (defined as electricity consumption minus PV generation), using real-world 30-min data from 68 PV households in Cyprus (one year). Forecasts were produced for horizons from 1 to 73 days, and benchmarked across recursive statistical models, including seasonal autoregressive-integrated-moving average (ARIMA) and Time-series Linear model (TSLM), and direct multi-output deep learning (DL) models, including Long Short-Term Memory (LSTM) and Transformer architectures. Models were evaluated in both univariate (net load-only) and multivariate settings, with the latter incorporating exogenous variables such as solar irradiance, air temperature, and humidity. The results show that for models without exogenous parameters, ARIMA with seasonal adjustment had the best performance in the short-term (RMSE = 145 W at 5 days), while LSTM outperformed in the medium-term forecasts (RMSE = 433 W at 66 days). When exogenous parameters are included, statistical models, particularly ARIMA and TSLM with calendar–weather interactions consistently outperformed across all forecast horizons, including the medium-term (RMSE = 364 W at 73 days). Prediction-interval analysis further indicates horizon-widening uncertainty for recursive statistical forecasts, whereas direct deep learning ensembles tend to produce comparatively stable interval widths, with regime-dependent changes during PV-active periods. These findings provide practical guidance on horizon-dependent model selection and the conditional value of exogenous inputs for planners and operators managing PV-rich residential systems.