📄 earth_science

Modelling São Paulo's Transport Sector Net-Zero Transition: Application of an Enhanced-FNN for HVO Biofuel Integration into the State's Energy Matrix

This study quantifies the significant CO $_2$ e emission reductions achievable through hydrotreated vegetable oil (HVO) integration in São Paulo's transport sector and demonstrates that an Enhanced Feedforward Neural Network outperforms traditional regression models in projecting these low-carbon pathways to 2050.

Original authors: Edward Iwuchukwu, Thiago Brito, Marilin Santos, Edmilson Moutinho dos Santos

Published 2026-06-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Edward Iwuchukwu, Thiago Brito, Marilin Santos, Edmilson Moutinho dos Santos

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the transportation system of São Paulo, Brazil, as a massive, hungry beast that eats fuel. For decades, this beast has been fed almost exclusively "fossil diesel," a fuel made from ancient, buried oil that spews a lot of smoke (carbon emissions) into the sky. The city wants to stop this beast from choking the planet, but it can't just starve the beast; it needs to feed it something cleaner that still makes the engine run.

This paper is like a recipe book and a crystal ball for swapping that dirty fuel for a cleaner one called HVO (Hydrotreated Vegetable Oil). Think of HVO as "renewable diesel" made from things like used cooking oil, animal fats, and soybeans. It's chemically so similar to regular diesel that you can pour it into the same tank without needing to rebuild the engine (a "drop-in" fuel).

Here is the breakdown of what the researchers did, using simple analogies:

1. The Problem: A Dirty Beast

The researchers looked at the years 2018 to 2023. They calculated how much "smoke" (CO2) the trucks and buses in São Paulo were producing.

The Old Fuel: Regular diesel is like a heavy, smoky coal fire. It produces about 2.7 kg of smoke for every liter burned.
The New Fuel: HVO is like a clean-burning gas fire. It produces only 0.7 kg of smoke for every liter.

2. The Experiment: Mixing the Drinks

The team tested what happens if you mix the two fuels together, like mixing orange juice with water.

The Mix: They looked at blends ranging from a tiny splash of HVO (13%) to a full glass of HVO (100%).
The Result: The more HVO you add, the less smoke comes out.
- If you switch to 100% HVO, the smoke drops by 74%.
- It's like turning down the volume on a loud radio; the louder the HVO gets, the quieter the pollution becomes.

3. The Crystal Ball: Predicting the Future (2030–2050)

The tricky part is guessing what will happen in the future. The researchers had a small amount of historical data (only 6 years), which is like trying to predict the weather for next year based on just one week of observations.

To solve this, they used two different "crystal balls":

Ball A (Quadratic Regression): This is a simple, straight-line guess. The paper says this method is a bit clumsy. When they tried to use it to predict far into the future (2050), it started screaming that emissions would skyrocket to impossible levels. It's like a car that accelerates uncontrollably when you press the gas pedal too hard.
Ball B (Enhanced-FNN): This is a fancy Artificial Intelligence (AI) tool. Think of it as a super-smart chef who tastes the ingredients and understands the complex flavors of the data.
- The AI was trained using a special trick (adding a little bit of "noise" to the data) to make it smarter without lying about the facts.
- The Verdict: The AI was much more stable. It predicted that emissions would rise slowly (because more people will be driving), but if we keep mixing in more HVO, the pollution stays manageable. It didn't go crazy like the simple guess did.

4. The "Credit" System: Getting Paid for Clean Air

The paper mentions a system called RenovaBio (and CBIOs). Imagine a neighborhood where if you keep your lawn clean, you get a coupon.

Every time a truck uses HVO instead of regular diesel, it "avoids" creating pollution.
The researchers calculated exactly how many "avoided pollution coupons" São Paulo could earn.
The Finding: The more HVO you use, the more coupons you get. If you go 100% HVO, the city earns a massive amount of these credits, which can be traded or used to meet environmental goals.

5. The Roadblocks and Recommendations

The paper points out that while HVO is great, building the factories to make it costs a lot of money (like building a new theme park).

The Solution: Instead of building brand-new factories from scratch, the researchers suggest retrofitting existing oil refineries. It's like taking an old, dusty factory and giving it a new, high-tech engine to start making green fuel instead of dirty fuel.
The Plan: They recommend that São Paulo's government creates clear rules to encourage this mixing (blending) and helps these factories switch over.

Summary

In short, this paper says: São Paulo can significantly clean its air by swapping dirty diesel for HVO (made from waste oils).

The Math: It works. More HVO = Less smoke.
The Tech: Using a smart AI (Enhanced-FNN) helps predict the future better than simple math, showing that this switch is a stable path forward.
The Goal: By 2050, if they keep mixing in more HVO, the city can move closer to "Net-Zero" (where they don't add any new pollution to the atmosphere), earning valuable credits along the way.

The paper doesn't promise that this will happen automatically; it just provides the map and the compass showing that if they take this route, the destination (cleaner air) is reachable.

Technical Summary: Modelling São Paulo's Transport Sector Net-Zero Transition

Problem Statement
The study addresses the need for a reproducible, data-driven framework to quantify the integration of Hydrotreated Vegetable Oil (HVO), or "renewable diesel," into São Paulo's transport energy matrix. While global trends indicate a rise in biofuel demand, there is a scarcity of subnational studies that simultaneously integrate regional energy matrix data, policy-aligned carbon-credit logic (specifically Brazil's RenovaBio/CBIO system), and uncertainty-aware long-term projections. The research aims to bridge the gap between descriptive market narratives and quantitative frameworks capable of supporting decision-making under data constraints, specifically for a region with a heavy reliance on legacy diesel fleets and an established biofuel ecosystem.

Methodology
The study employs a distance-based fuel consumption model combined with an Enhanced Feedforward Neural Network (Enhanced-FNN) for forecasting.

Data and Baseline: Historical data from 2018 to 2023 was sourced from the Data Portal for Cities, Google Environmental Insights Explorer, and SENATRAN. The model utilizes a fleet-average fuel efficiency of 3.60 km/L.
Emission Accounting: A transparent accounting methodology calculates total fuel consumption ( $F_T$ ) based on total distance ( $D_T$ ) and efficiency ( $f_E$ ). Emission factors ($ef$) are applied based on blend composition: 2.70 kg CO2e/L for fossil diesel and 0.70 kg CO2e/L for HVO. Total blend emissions ( $E_B$ ) and avoided emissions (carbon credits, $C_B$ ) are computed deterministically relative to a fossil diesel baseline.
Machine Learning Approach: To project emissions for 2030, 2040, and 2050 under a Business-As-Usual (BAU) scenario, the authors developed an Enhanced-FNN.
- Architecture: The model features hidden layers of [256, 128, 64] with SiLU (Swish) activation functions.
- Uncertainty Quantification: Monte Carlo (MC) dropout is employed during inference to provide uncertainty estimates.
- Training Strategy: Due to the limited historical dataset (six years), a "train-only augmentation" strategy was used, where the training subset was augmented by a factor of 1,000 via noise injection while preserving the empirical distribution. Validation and test sets remained unchanged.
- Benchmark: A quadratic regression model was implemented as a parametric benchmark to test the robustness of the neural network against polynomial extrapolation.

Key Results

Historical Performance (2018–2023): The analysis confirms a monotonic reduction in CO2e emissions as HVO blending increases.
- Average emissions for pure fossil diesel (D) were 39.105 MT.
- Average emissions for 100% HVO (R) were 10.138 MT, representing a 74.1% reduction.
- Avoided emissions (credits) increased proportionally with HVO content, reaching 28.966 MT for the 100% HVO case.
Long-Horizon Projections (2030–2050):
- The Enhanced-FNN demonstrated superior performance with an $R^2 \approx 0.997$ , Mean Absolute Error (MAE) of ~0.883 MT, and Mean Squared Error (MSE) of ~1.286 MT². It produced stable, nonlinear projections that accounted for the 2020 pandemic-induced dip in fleet activity.
- The Quadratic Regression benchmark showed significantly lower explanatory power ( $R^2 \approx 0.65$ ) and exhibited instability, yielding rapidly accelerating emission trajectories outside the calibration range.
Sensitivity Analysis: The study identified three primary decarbonization levers: reducing annual vehicle mileage, improving fleet fuel efficiency, and increasing renewable blend share. The analysis showed that while fuel efficiency improvements yield nonlinear emission reductions, renewable fuel integration provides proportional emission displacement.

Key Contributions

Reproducible Framework: The study provides a transparent methodology for quantifying blended fuel emissions and avoided emissions that can be adapted to other subnational contexts with similar diesel reliance and biofuel policies.
Enhanced Forecasting: It demonstrates the efficacy of an Enhanced-FNN with MC-dropout for long-term energy forecasting in data-constrained environments, outperforming traditional polynomial regression in stability and uncertainty quantification.
Policy Alignment: The work explicitly links physical emission reductions to the generation of avoided-emission credits under the Brazilian RenovaBio framework, offering a quantitative basis for evaluating blending mandates (e.g., 13%, 15%, 30%, 50%, and 100% HVO).

Significance and Claims
The paper claims that its primary significance lies in the integration of high economic activity data, legacy fleet characteristics, and established biofuel ecosystems to create a transferable model for regional energy-matrix evidence. The authors assert that the study supports progressive blending mandates in São Paulo by demonstrating that increasing HVO content consistently reduces projected emissions and strengthens avoided-emission benefits.

The study modestly positions itself as a decision-support tool rather than a definitive prediction of future market outcomes. It highlights that while the Enhanced-FNN offers robust forecasting, future research should incorporate broader operational heterogeneity (e.g., vehicle class stratification) and sensitivity analyses regarding fuel efficiency and activity growth assumptions to further refine scenario realism. The work emphasizes that the scientific value of the São Paulo case study is its ability to serve as a template for other regions seeking to align regional energy evidence with forecast models and policy crediting rules.