Sobre

We are seeking a skilled freelancer to develop and implement solar power forecasting model. The dataset suggested is shared in https://github.com/CDAC-lab/UNISOLAR/tree/main But I can accept suggestions from others if they are more suitable. This hybrid model needs to outperform other baseline methods: The model hybrid, which integrates physically informed and data-driven methods in deep learnin. This is the suggested composition (it is not definitive). (i) 3D-CNN Meteorological Coding Module The first module is responsible for extracting spatial and spectrotemporal features from meteorological data, such as global irradiance, ambient temperature, relative humidity, cloud cover, and wind speed. For this purpose, a three-dimensional convolutional network (3D-CNN) is used, which processes data volumes representing spatial (latitude, longitude) and temporal (forecast horizons) variations. This approach allows capturing regional patterns and complex atmospheric dynamics, essential for adequately representing the climatic context of the photovoltaic plant. (ii) Transformer-LSTM Temporal Coding Module The second module, called the temporal encoder, is composed of a combination of Transformer and LSTM layers, responsible for modeling multiscale dependencies in the temporal domain. The Transformer component employs the self-attention mechanism to identify long-range relationships between input sequences, while the LSTM captures local and non-linear dynamics. This combination ensures a robust temporal representation, uniting global attention and local sequential memory, which is particularly relevant for photovoltaic data marked by interday and intrahourly variability. (iii) Physics-Informed Layers The third module constitutes the physics-informed core of the system. In it, thermodynamic constraints, electrical equations, and empirical relationships—such as the Shockley equation, cell temperature models, and energy balances—are embedded directly into the neural architecture or loss function. These layers ensure that predictions respect fundamental physical principles, such as energy conservation, module operating limits, and realistic nonlinear behaviors. The physical regularization term works in conjunction with the data prediction error, ensuring physical consistency and generalization under unobserved conditions. (iv) Probabilistic Decoding Module (Mixture Density Network) The fourth module is responsible for decoding and quantifying uncertainties. For this purpose, a Mixture Density Network (MDN) is used, which models the predicted output as a mixture of Gaussian distributions. This component allows for the estimation of confidence intervals and the propagation of observational and model uncertainties throughout the inference process, providing not only point predictions but also probabilistic estimates of photovoltaic generation. This capability is fundamental for applications in risk management, market operation, and predictive control in energy systems. (v) Adaptive Fusion Mechanism Finally, hybrid model incorporates an adaptive fusion mechanism, responsible for dynamically combining the contributions of the physics-based and data-based modules. This mechanism is implemented through a learnable weighting layer, which adjusts the relative importance of each component based on the local reliability of the physical information and the degree of uncertainty of the observational data. In this way, the system is able to transition between regimes dominated by physical models (in scenarios with little data) and regimes driven by machine learning (in contexts of high data availability), preserving the stability and overall consistency of the model. The project involves working with various baselines, including pure physical models, data-driven models, and hybrid models, to enhance forecasting accuracy. The hybrid model should outperform the baseline models. Baselines 1. Pure Physical Models: PVLib: Standard library for photovoltaic modeling SAM (System Advisor Model): NREL model 2. Pure Data-Driven Models: Vanilla LSTM CNN Temporal Fusion Transformer (TFT) CNN-LSTM 3. State-of-the-Art Hybrid Models: Physics-guided LSTM (Li et al., 2023) Hybrid CNN-Physics (Wang et al., 2024)