(373e) Optimal Production of Solar Photovoltaic Panels in Mexico through a Hybrid Machine Learning and Mathematical Programming Approach

Nowadays, solar energy technology represents the third most widely applied renewable energy source in the world, behind only hydroelectric and wind power. Electricity generated from fossil fuels generates CO₂ emissions ranging between 400 to 1,000 gCO₂ eq/kWh, while the emissions from silicon-based solar panels range from 98.3 to 149.3 g CO₂ eq/kWh, with an average value of 123.8 g CO₂ eq/kWh. Hence, photovoltaic (PV) technology represents promising opportunities to meet future global energy demand in an environmentally friendly manner. In recent decades, the uptake of solar PV has grown considerably, establishing a strong global market for solar panels capable of producing clean energy. Furthermore, it is projected that photovoltaic electricity will dominate the world's energy supply before the end of the 21st century. The last decade has been exceptionally significant for solar PV, with the worldwide installed capacity for solar power generation comparable to other energy production technologies. Indeed, solar energy has contributed more new capacity than the combined capacity of both nuclear and fossil fuel power generation.

Latin America, with Mexico in a prominent position, looms as a region full of promising opportunities in the PV market. This potential is driven by its abundant solar resources, rapid growth in electricity demand, and a progressive legislative environment conducive to the widespread adoption of renewable energy. Mexico, in particular, plays a crucial role in this context, positioned alongside its Latin American colleagues, because it combines significant levels of solar irradiation with an attractive market that captures the attention of potential investors. Brazil has also charted a unique path in the region by effectively integrating solar PV into energy auctions, which has led to a remarkable increase in the number of solar PV installations that are seamlessly connected to the grid. However, while other countries in the region are gradually adopting PV in their energy matrices, a proactive effort is underway to implement programs and mechanisms aimed at accelerating the proliferation of PV projects.

However, with the increase in the number of solar PV installations, the challenge of achieving large-scale, sustainable solar panel production becomes even more crucial. In Mexico, the development of a local industry value chain remains a significant obstacle. Currently, the country has 10 local solar PV panel manufacturers that collectively contribute to an annual production capacity of 1.5 GW. This highlights the growing importance of strengthening a strong domestic supply chain to meet the demands of solar energy expansion in Mexico. It underscores the growing momentum of the solar sector in Mexico, although there is still potential for further growth and to establish a strong domestic supply chain to support the expansion of the industry.

Surprisingly, there is a significant paucity of models reported in the literature that employ Machine Learning for efficient optimization of renewable energy systems. These strategies can be used to generate models with various applications, such as classification models, regression models, or models capable of predicting future data from data sequences.

Therefore, in this work, a novel approach has been developed and implemented in Mexico. An Artificial Neural Networks (ANN) model has been built, that estimates raw material quantities, costs, water, and energy usage, and polycrystalline silicon production for PV panel plants. Sophisticated optimization challenges have been met by meticulously integrating the ANN model into an advanced Non-Linear Programming (NLP) framework, skillfully implemented and run in the Pyomo library. However, the genuine innovation lies in this pioneering approach that allows seamless integration of the ANN model into the optimization process, with the support of the Optimization and Machine Learning Toolkit (OMLT) library. This seamless integration facilitates the creation of individually tailored surrogate models for each PV production plant, considering a wide range of variables and constraints, such as production schedules and lead times. This improves the accuracy and applicability of the methodology. This work introduces several key innovations. Firstly, it presents a multi-objective mathematical model that thoroughly examines the optimal production process of solar PV panels, covering the entire production sequence from polycrystalline silicon to wafers, cells, and modules. Additionally, an innovative optimization approach incorporating both ANN models and mathematical programming models is applied. In particular, the creation of ANN models for polycrystalline silicon production plants is particularly noteworthy, involving meticulous fine-tuning of hyperparameters such as the number of hidden layers, the number of neurons in the hidden layers, and the learning rate. Furthermore, the study assesses the Mexico's potential and capacity to enter the global solar panel production market. Finally, the research successfully predicts the waste production attributed to solar PV panels currently installed throughout Mexico.