Time series forecasting still awaits a transformative breakthrough like that happened in computer vision and natural language processing. The absence of extensive, domain-independent benchmark datasets and standardized performance measurement units poses a significant challenge for it, especially for photovoltaic forecasting applications. Additionally, since it is often time domain-driven, a plethora of highly unique and domain-specific datasets were produced. The lack of uniformity among published models, developed under diverse settings for varying forecasting horizons, and assessed using non-standardized metrics, remains a significant obstacle to the progress of the field as a whole. To address these issues, a systematic review of the state-of-the-art literature on prediction tasks is presented, collected from the Web of Science and Scopus databases, published in 2022 and 2023, and filtered using keywords such as “photovoltaic,” “deep learning,” “forecasting,” and “time series.” Finally, 36 case studies were selected. Before comparing, a state-of-the-art demonstration of key elements in the topic was presented, such as model type, hyperparameters, and evaluation metrics. Then, the 36 articles were compared in terms of statistical analysis, including top publishing countries, data sources, variables, input, and output horizon, followed by an overall model comparison demonstrating every proposed model categorized into model type (artificial neural network units, recurrent units, convolutional units, and transformer units). Due to the mostly utilization of specific private datasets measured at the targeted location, having universal error metrics is crucial for clear global benchmarking. Root Mean Squared Error and Mean Absolute Error were the most utilized metrics, although they specifically demonstrate the accuracy relative to their respective sites. However, 33% utilized universal metrics, such as Mean Absolute Percentage Error, Normalized Root Mean Squared Error, and the Coefficient of Determination. Finally, trends, challenges, and future research were highlighted for the relevant topic to spotlight and bypass the current challenges.

Accurate photovoltaic (PV) power forecasting is essential for grid integration, particularly in maritime climates with dynamic weather patterns. This study addresses high-dimensional meteorological data challenges by systematically evaluating 32 variables across four categories (solar irradiance, temperature, atmospheric, hydrometeorological) for day-ahead PV forecasting using long short-term memory (LSTM) networks. Using six years of data from a 350 kWp solar farm in Scotland, we compare satellite-derived data and local weather station measurements. Surprisingly, downward thermal infrared flux—capturing persistent atmospheric moisture and cloud properties in maritime climates—emerged as the most influential predictor despite low correlation (1.93%). When paired with precipitation data, this two-variable combination achieved 99.81% R², outperforming complex multi-variable models. Satellite data consistently surpassed ground measurements, with 9 of the top 10 predictors being satellite derived. Our approach reduces model complexity while improving forecasting accuracy, providing practical solutions for energy systems.

Accurate 24 h photovoltaic (PV) power forecasting is essential for day-ahead scheduling, storage operation, reserve planning, and market participation. However, published deep learning comparisons are often difficult to reproduce and interpret because they use inconsistent datasets, forecasting horizons, baselines, evaluation metrics, and leakage-control procedures. From a software engineering perspective, this limits the trustworthiness, comparability, and practical adoption of AI-based forecasting systems. This paper presents a controlled and reproducible benchmarking framework for evaluating AI-driven forecasting software. The framework is applied to nine deep learning architectures, three non-deep learning reference models, and two persistence baselines for hourly PV-power forecasting at a 350 kWp rooftop installation near Edinburgh, Scotland. All models were evaluated under a consistent experimental protocol, including the same chronological train–validation–test split, a 32-feature meteorological and solar-geometry input set, a 24-step forecasting horizon, capacity-normalised mean absolute error (NMAE), and Bayesian hyperparameter optimisation. The results show that TCN-LSTM achieved the best aggregate H24 performance with 7.22% NMAE, narrowly outperforming CPWformer-DEC at 7.28% and CT-PatchTST at 7.31%. LightGBM ranked fourth at 7.35% with fixed hyperparameters, outperforming six of the nine deep learning models. The top three models differed by only 0.09 percentage points, indicating that architectural superiority cannot be established reliably without significance testing and operational diagnostics. Per-horizon analysis showed that CT-PatchTST and S-Mamba performed best at the nearest forecast steps, whereas TCN-LSTM provided the most stable far-horizon profile. Peak-power diagnostics further revealed that aggregate NMAE can mask operational shortcomings, as Naive Persistence outperformed all deep learning models in high-output peak detection. The findings highlight the importance of reproducible benchmarking, leakage safeguards, horizon-aware evaluation, and operationally meaningful diagnostics in trustworthy AI software evaluation. The novelty of this work lies not in proposing a new architecture but in a controlled, reproducible framework that benchmarks fourteen forecasters under identical conditions, with explicit leakage safeguards, per-horizon reporting, and operationally meaningful peak diagnostics, enabling claims of architectural superiority to be made trustworthy rather than merely favourable. Architecture selection for PV forecasting should therefore consider not only aggregate accuracy but also reliability, interpretability of evaluation outcomes, and deployment-relevant performance behaviour.