Accurate forecasting of variable renewable energy (vRE) generation is essential for the reliable operation of modern power systems with high shares of wind and solar energy. Recent advances in machine learning have introduced tokenised representations of time series, enabling the application of language-model-based architectures and the development of generalist forecasting models. However, their applicability to energy system forecasting remains largely unexplored, particularly in combination with exogenous predictors such as meteorological variables.

This work investigates the use of tokenised time series embeddings for regionally aggregated wind and solar power forecasting in Germany. Building upon recent approaches of discretisation, two embedding strategies are implemented: (i) direct value tokenisation through scaling and binning, and (ii) learned discrete representations of temporal patterns using self-supervised encoding of time series segments. These embeddings are evaluated in two complementary modelling frameworks. First, classical neural networks are trained on the embedded representations and compared against previously developed Bayesian-optimised multilayer perceptron models. Second, transformer-based sequence models inspired by natural language processing are trained on the tokenised sequences to assess their suitability for vRE forecasting tasks.

A particular focus is placed on the integration of exogenous meteorological predictors into the tokenised modelling framework. Unlike existing generalist model approaches, which primarily consider endogenously driven time series tasks, this study proposes discrete temporal embeddings with continuous weather features derived from reanalysis data and numerical weather predictions.

The analysis is conducted on multi-year datasets of hourly wind and solar power generation aggregated at the national level. First results indicate that tokenised representations can capture relevant temporal structures, classical models trained on embeddings show comparable accuracy to their continuous-input counterparts, while transformer-based models benefit from the sequential structure of tokenised inputs, particularly for longer forecast horizons. The inclusion of exogenous predictors is found to be essential for achieving high forecast accuracy in all approaches.

The findings demonstrate that tokenised time series representations show a promising extension of existing forecasting methodologies. They highlight the necessity of incorporating domain-specific information, such as meteorological drivers, when applying generalist models to vRE forecasting. The presented work contributes to the ongoing development of hybrid modelling approaches that combine data-driven representation learning with physically meaningful inputs.