The research team assembled a benchmarking framework that put an unusually broad range of forecasting architectures through their paces under strictly identical experimental conditions. Rather than comparing models across different datasets, different preprocessing pipelines and different evaluation windows — a common weakness in the machine learning literature — the authors trained and tested every model on the same photovoltaic power time-series data, sourced from a publicly available Mendeley dataset of historical generation records. The contenders spanned three generations of forecasting technology: classical recurrent and convolutional deep learning models, modern transformer-based architectures, and the newest class of so-called foundation models, which are large networks pre-trained on vast collections of time series and then adapted to specific forecasting tasks.

In the classical corner sat three workhorses of sequence modeling. The long short-term memory network, or LSTM, uses gated memory cells to decide what information from past time steps should be carried forward, allowing it to capture slow-moving trends such as the daily solar cycle. The gated recurrent unit, or GRU, is a leaner relative that merges some of those gates for computational efficiency. Alongside them ran a one-dimensional convolutional neural network, or 1D CNN, which slides small filters across the input sequence to detect local patterns — sudden ramps in output, for instance — without the sequential processing that makes recurrent networks slow to train. These models have dominated solar forecasting papers for the better part of a decade, and they provided the baseline against which the newer architectures had to justify themselves.

The transformer family brought a fundamentally different mechanism to the table. Instead of reading a sequence step by step, transformers use multi-head attention, a mathematical operation that lets the model directly compare any two points in time and weigh their relevance to the prediction at hand. The study evaluated the Temporal Fusion Transformer, or TFT, an architecture designed for interpretable multi-horizon forecasting that combines attention with gating mechanisms and variable selection, and the iTransformer, a recent variant that reorganizes how the attention operation is applied across the dimensions of multivariate time-series data. The appeal of attention in solar forecasting is intuitive: a cloudy afternoon three days ago may matter more to tomorrow’s forecast than the smooth output of yesterday morning, and attention can learn to lock onto exactly those informative episodes.

The most forward-looking entrants were the foundation models. Time-MoE, introduced at the International Conference on Learning Representations in 2025, is a billion-scale time-series foundation model built on a mixture-of-experts design, in which different specialized sub-networks are activated for different inputs, allowing enormous model capacity without a proportional explosion in computation. Mamba4Cast takes a different route entirely, using state-space models — a mathematical framework for describing how a hidden internal state evolves over time — to achieve efficient zero-shot forecasting, meaning it can predict on data it has never seen without any task-specific training. These models represent the industry’s bet that forecasting, like language translation before it, will eventually be solved by enormous general-purpose networks rather than bespoke per-site models.

The headline result of the benchmark is a clear, graded improvement across architectural generations. The LSTM, the oldest model in the lineup, achieved a skill score of 0.29 and a root-mean-square error, or RMSE, of 0.158. In forecasting parlance, the skill score measures improvement over a naive persistence forecast — the assumption that the next value equals the current one — so 0.29 means the LSTM beat that naive baseline by a modest margin. The transformer-based and foundation-model families both pushed RMSE down to around 0.120 and 0.131 respectively, confirming that attention mechanisms and large-scale pre-training do translate into measurably better solar predictions. The gap may look small in absolute terms, but in grid operations, where forecasts drive billion-dollar scheduling decisions, even percentage-point improvements compound into significant economic and emissions savings.

The outright winner, however, was neither a pure transformer nor a pure foundation model, but a hybrid. The authors’ combined CNN-foundation-transformer architecture achieved a mean absolute error of 0.071, an RMSE of 0.106 and a normalized RMSE of 12.1, corresponding to a skill score of 0.64 against persistence forecasting — more than double the LSTM’s score. The logic of the combination is technically elegant. The convolutional front end extracts local, short-scale features such as abrupt irradiance ramps; the foundation component contributes temporal representations learned from an enormous diversity of time series, giving the model a robust sense of periodicity and seasonality; and the transformer’s attention layers then integrate these multi-scale signals, deciding which features from which time steps matter most for the forecast horizon. Visual analysis of the predictions confirmed that this architecture tracked day-to-day generation peaks and variability patterns with high fidelity, precisely the behavior that matters when operators must decide how much reserve capacity to hold.

Why should a hybrid beat its individual components? The answer likely lies in the complementary inductive biases each block contributes. Recurrent and convolutional layers impose strong structural assumptions about locality and continuity, which helps when data is limited; attention imposes few assumptions and can model long-range dependencies, but typically needs more data to shine; foundation models arrive pre-armed with generalized temporal knowledge that transfers across domains. Solar power data exhibits all of these regimes at once — smooth diurnal cycles, noisy cloud-driven fluctuations, and seasonal drift — so an architecture that layers multiple forms of temporal reasoning appears to capture the phenomenon more completely than any single mechanism. The finding echoes a broader trend in the field, where hybrid CNN-LSTM and CNN-LSTM-transformer designs have repeatedly outperformed monolithic models in prior studies of solar and load forecasting.

The practical implications reach well beyond academic leaderboards. Short-term PV forecasting is a linchpin of grid stability: system operators use these predictions to schedule conventional generation, manage battery storage, and participate in electricity markets. A forecast with a skill score of 0.64, as the best hybrid achieved, means substantially fewer surprise deficits on cloudy days and less curtailment of solar output on bright ones. Moreover, the study’s unified evaluation protocol is itself a contribution. Because every model faced identical data, preprocessing and metrics, the results offer a rare apples-to-apples comparison in a literature notorious for incomparable claims, giving practitioners a defensible basis for choosing architectures rather than relying on cherry-picked benchmarks.

The study also maps the road ahead. Foundation models like Time-MoE and Mamba4Cast performed strongly even though they were not designed specifically for solar data, suggesting that general-purpose temporal pre-training is a powerful starting point that domain-specific fine-tuning could push further. The authors’ work was supported by Al-Quds University’s Najjad Zeenni Faculty of Engineering, and the data and supplementary materials accompanying the paper provide a foundation for replication. As renewable penetration deepens, the contest between forecasting architectures will only intensify — and this benchmark suggests the future belongs not to any single paradigm, but to architectures that know how to make attention, convolution and pre-trained temporal knowledge work together.

Subject of Research: Machine learning architectures for short-term photovoltaic power time-series forecasting

Article Title: A systematic evaluation of classical, transformer-based, and foundation models for photovoltaic power time-series forecasting

Article References: Salman, D., Alzeer, I., & Isayed, W. (2026). A systematic evaluation of classical, transformer-based, and foundation models for photovoltaic power time-series forecasting. Neural Computing and Applications, 38(19), Article 764. https://doi.org/10.1007/s00521-026-12498-x

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12498-x

Keywords: photovoltaic power forecasting, deep learning, transformer models, foundation models, LSTM, time-series forecasting, hybrid neural networks, Temporal Fusion Transformer, Time-MoE, Mamba4Cast, renewable energy, grid stability