Online Fine-Tuning of Carbon Emission Predictions using Real-Time Recurrent Learning for State Space Models
Julian Lemmel, Manuel Kranzl, Adam Lamine, Philipp Neubauer, Radu Grosu, Sophie Neubauer · 2025
Read Online Fine-Tuning of Carbon Emission Predictions using Real-Time Recurrent Learning for State Space Models by Julian Lemmel, Manuel Kranzl, Adam Lamine, Philipp Neubauer, Radu Grosu, Sophie Neubauer (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-00804.txt.
Introduction
Julian Lemmel and co-authors published a computing paper in 2025 on fine-tuning state space model predictions online as new data arrive. Such models are efficient with long-range patterns but usually freeze after training. Their method updates the weights during use through real-time recurrent learning, a classic trick revived for modern recurrent units. The test task was carbon emission prediction from embedded car hardware.
On that small real dataset, online updates steadily cut prediction error during inference. The authors present the scheme as a fit for dynamic, resource-tight settings where conditions drift after deployment. It shows a static model can keep learning on the job.
Catalog record
| CREATOR | Julian Lemmel, Manuel Kranzl, Adam Lamine, Philipp Neubauer, Radu Grosu, Sophie Neubauer |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 23 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-00804
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Online Fine-Tuning of Carbon Emission Predictions using Real-Time Recurrent Learning for State Space Models Julian Lemmel, Manuel Kranzl, Adam Lamine, Philipp Neubauer, Radu Grosu, Sophie Neubauer Abstract This paper introduces a new approach for fine-tuning the predictions of structured state space models (SSMs) at inference time using real-time recurrent learning. While SSMs are known for their efficiency and long-range modeling capabilities, they are typically trained offline and remain static during deployment. Our method enables online adaptation by continuously updating model parameters in response to incoming data. We evaluate our approach for linear-recurrent-unit SSMs using a small carbon emission dataset collected from embedded automotive hardware. Experimental results show that our method consistently reduces prediction error online during inference, demonstrating its potential for dynamic, resource-constrained environments. I INTRODUCTION Vehicle emission prediction is a critical area of research due to its implications for environmental sustainability and public health. Predicting emissions from data readily available on modern automobiles can help estimate effects on air quality and climate alike. The advent of deep learning has revolutionized emissions prediction by enabling data-driven approaches that leverage large-scale datasets from chassis dynamometer tests and on-road measurements. Recurrent neural networks (RNNs) are capable of modeling temporal dependency in time-series data and have been successfully applied to emission prediction tasks, such as forecasting nitrogen oxides (NOx) emissions from vehicles under varying traffic conditions [ 15 , 16 ] . Recently, state space models (SSMs) such as LRUs [ 3 ] , have revitalized research on RNNs, as foundational large language models. Continual learning approaches, such as incremental retraining, are essential to ensure that emission prediction models remain accurate as new data become available [ 16 ] . Online learning techniques, which process data in real-time [ 12 , 7 ] , are particularly relevant for applications in intelligent transportation systems and digital-twin platforms. These technologies enable models to adapt to evolving traffic conditions, vehicle technologies, and environmental factors, ensuring consistent performance over time. In this work, we show how LRUs, pretrained on a small carbon emission data-set collected from embedded hardware, can be fine-tuned online through real-time recurrent learning [ 12 ] . The fine-tuning successfully refines the emission predictions. Our results are very encouraging and generalizable to fine-tuning large language models during inference time. II RELATED WORK Traditional models, such as COPERT and EMFAC, have been widely used for macro-level analysis, providing estimates based on aggregated data [ 10 , 2 ] . However, these models often lack the granularity required for real-time applications and fail to account for dynamic driving conditions and regional variability. Modal models like MOVES and CMEM offer higher temporal resolution but rely on static assumptions, limiting their adaptability to real-world scenarios [ 16 ] . Artificial neural networks have been widely applied to model complex relationships between vehicle attributes, driving patterns, and emissions. Advanced architectures, such as ST-MFGCN [ 13 ] , further enhance spatial-temporal correlations by integrating graph-based learning and attention mechanisms. These models address the limitations of traditional methods by dynamically adapting to spatial and temporal variations in emission data [ 16 ] . RNNs are used for modeling temporal dependencies in the data. They are different from classical feedforward neural networks due to recurrent connections resulting in a hidden state that is maintained across entire trajectories. These models are particularly effective in scenarios where emissions are influenced by sequential driving behaviors, such as acceleration, deceleration, and idling [ 16 , 8 ] . For instance, [ 15 ] demonstrated the effectiveness of LSTMs in predicting vehicle emissions by integrating weather variables, driving patterns, and historical emissions data [ 15 ] . Similarly, GRUs have shown efficiency in handling shorter forecasting horizons, making them suitable for real-time applications [ 8 ] . To the best of our knowledge, none of the works discussed above has addressed the problem of online fine-tuning the carbon-emission predictions of SSMs, in particular of LRUs, through real-time recurrent learning. III DATASET We empirically test our approach on a small dataset that was recorded from specialized embedded sensor hardware developed by the Edge Intelligence Research Institute Nanjing. The dataset consists of a total of 34244 samples and 12 columns. The columns are: (1) Timestamp (s), (2) Engine speed rpm (rounds/min), (3) Hourly fuel consumption (litre/hour), (4) Coolant temperature (°C), (5) Speed (kilometre/hour), (6) Instantaneous fuel economy (kilometre/litre), (7) NO ppm (parts per million), (8) NO2 ppm, (9) NOX ppm, (10) CO2, and (11) CO ppm. The dataset comprises 5 recording sessions with a total of 12:02:58 of recorded time. The median time delta between consecutive data points on any given day was 1.0 second, while the mean time delta was ~1.267 seconds. This implies that a considerable number of measurements are missing. Fig. 1: Plot of the raw data for one recorded drive. For our carbon emission prediction problem, the last five columns form the target values, while the remaining columns are the explaining variables. For fitting a model, we selected the first 4 days as our training set and the remaining days as the validation set. This corresponds to a 80/20 split that is commonly found in the literature. Fig. 1 shows the raw data of the recording used for validation. Automobile carbon emissions naturally depend on weather conditions. For example, low and high temperatures may require the use of heating and air conditioning, respectively, both leading to an increase in power demand and subsequently carbon emissions. In order to capture the dependency of carbon emissions on the current weather, we acquired Nanjing regional weather data for the days of recording from https://www.visualcrossing.com/ . It includes hourly measurements for a range of indicators such as temperature, precipitation, or a categorical description of the weather. The weather data was concatenated with the emission data by matching the timestamp. IV METHODS Carbon emission prediction is a time-series prediction task that can be naturally approached using sequence models such as RNNs. For this work, we implemented a carbon emission model using SSMs, and later fine-tuned them during inference using real-time recurrent learning (RTRL) [ 12 , 7 ] . In the following, we will first introduce SSMs and RTRL. We will give an intuition as to why SSMs have a strong computational advantage when trained using BPTT, and also how they allow for computing the exact gradient with reasonable computational complexity when using RTRL. IV-A Sequence Models When doing Machine Learning on sequential data it is highly beneficial for the model to retain a memory of preceding inputs. Sequence models such as Recurrent Neural Networks (RNNs) can do just that by persisting a hidden state that is used when computing future outputs. Due to the evolving hidden state (”memory”), they exhibit temporal dynamic behavior which enables them to accurately model temporal dependencies in time-series data. RNNs are traditionally trained using backpropagation through time (BPTT), where an input sequence of fixed length is fed to the network, producing a corresponding sequence of outputs. The outputs are used for computing the loss and its gradient where inputs at the beginning of the sequence influence all succeeding outputs. In the following we introduce two instantiations of RNNs that were used for this project: Linear Recurrent Units and Continuous-Time RNNs. IV-B Linear State Space