Publications

2023

1. SuMob ’23

   Incident Congestion Propagation Prediction Using Incident Reports

   Mathias Niemann Tygesen, Mayank Sharan, Francisco C. Pereira, and 2 more authors

   In Proceedings of the 1st ACM SIGSPATIAL International Workshop on Sustainable Mobility 2023

   Abs PDF

   Accurate predictions of how congestion propagates are essential for mitigating its effects on traffic and the urban environment. However, the vast majority of state-of-the-art traffic prediction models focus on regular traffic scenarios and struggle to adapt to the conditions following incidents. This is particularly problematic since the irregular periods after incidents are arguably when traffic predictions are most critical. Current traffic models struggle with non-recurring congestion for two reasons: they lack inputs alerting them an incident has happened, and traffic data containing incident information is scarce. We create two new such datasets: one by simulating incidents and their congestion in an open-source microscopic simulator and another by fusing real-world traffic flow data with incident reports. We then propose a framework that integrates incident reports into deep learning models for congestion propagation prediction. Our framework utilizes the recent traffic flow data and fuses it with information from incident reports. We perform a detailed empirical comparison between recurrent and graph-based models utilizing incident reports against baselines. Our study demonstrates that our framework significantly outperforms state-of-the-art graph-based models that do not account for incident reports.

2. Thesis

   Taming the Long Tail of Deep Probabilistic Forecasting

   Mayank Sharan, and Rose Yu

   Sharan_ucsd_0033M_22111 2023

   Abs PDF

   Deep probabilistic forecasting is gaining attention across numerous applications. Fromweather prognosis, electricity consumption estimation, traffic flow prediction, to autonomousvehicle trajectory prediction. However, existing approaches focus on improving on averagemetrics without addressing the long-tailed distribution of errors. This thesis identifies longtail behavior in the error distribution of state-of-the-art deep learning methods for probabilisticforecasting. We analyze potential sources and explanations for this behavior. Further, wepresent two loss augmentation methods to mitigate tailedness pf error distributions: ParetoLoss and Kurtosis Loss. Both methods modify the loss using the concept of moments topenalize higher loss samples. Kurtosis Loss uses a symmetric measure, the fourth moment, while Pareto Loss uses an asymmetric measure of right-tailedness and models loss using aGeneralized Pareto Distribution (GPD). We demonstrate the performance of our methods onseveral real-world datasets, including time series and spatiotemporal trajectories, achievingsignificant improvements on tail error metrics, while maintaining and even improving uponaverage error metrics.

2022

1. arXiv

   Taming the Long Tail of Deep Probabilistic Forecasting

   Jedrzej Kozerawski*, Mayank Sharan*, and Rose Yu

   arXiv preprint arXiv:2202.13418 2022

   Abs PDF

   Deep probabilistic forecasting is gaining attention in numerous applications ranging from weather prognosis, through electricity consumption estimation, to autonomous vehicle trajectory prediction. However, existing approaches focus on improvements on the most common scenarios without addressing the performance on rare and difficult cases. In this work, we identify a long tail behavior in the performance of state-of-the-art deep learning methods on probabilistic forecasting. We present two moment-based tailedness measurement concepts to improve performance on the difficult tail examples: Pareto Loss and Kurtosis Loss. Kurtosis loss is a symmetric measurement as the fourth moment about the mean of the loss distribution. Pareto loss is asymmetric measuring right tailedness, modeling the loss using a generalized Pareto distribution (GPD). We demonstrate the performance of our approach on several real-world datasets including time series and spatiotemporal trajectories, achieving significant improvements on the tail examples.

2. ISUW ’22

   Short Term Load Forecasting for Dynamic Clusters using Deep Learning

   Mayank Sharan, Shreeyash Geda, Prasoon Patidar, and 1 more author

   Compendium of Technical Papers, ISUW 2022

   Abs PDF

   Electrical load forecasting is critical for utility planning, with short term load forecasting (STLF) used in short term planning like sale and purchase of power and adjusting generation to match expected demand. Current STLF frameworks are limited to static clusters where the combined load can be physically measured through a device. The wide availability of Advanced Metering Infrastructure (AMI) data enables a bottom-up approach where clusters can be dynamic. In this paper, we propose a framework capable of performing STLF for dynamic clusters. Anonymised AMI data from Nevada, USA was used for the development and testing of the framework. This framework was also evaluated on the New York Independent System Operator (NYISO) load forecasting task for universal benchmarking

2020

1. engrXiv

   Detecting Nicotine Addiction via Eye Tracking

   Mayank Sharan, and Rahul Garg

   Engineering Archive 2020

   Abs PDF

   Smoking addiction is a growing social epidemic affecting about 1 in 5 adults in the world directly and many more in the form of passive smoking. Diagnostic tools to determine the nicotine addiction are either unreliable or restrictive in their use. This makes identification and assessment less accessible for those who need it. We attempted to evaluate eye tracking paired with machine learning as an alternative to existing diagnostic tools for assessing nicotine dependence.

2. US Patent

   Low frequency energy disaggregation techniques

   Alex Shyr, Vivek Garud, Abhay Gupta, and 3 more authors

   2020

   Abs PDF

   The present invention teaches methods of performing appliance itemization based on consumption data , including : receiving at a processor the data ; determining if the data includes active signals and / or inactive signals ; upon detection of an active signal : detecting and estimating active water heating consumption and lighting consumption ; upon detection of an inactive signal : detecting and estimating passive water heating consumption , refrigerator consump tion ; and detecting vacation mode . Methods are disclosed of appliance itemization based whole house consumption data consumption from an advanced metering infrastructure device , the data being at 15 , 30 , or 60 minute intervals , including : applying disaggregation models to provide detection and estimation of any lighting , water heating , refrig eration , pool pumps , heating , or cooling appliances ; apply ing rule - based models to provide detection and estimation of any cooking , laundry , entertainment , and / or miscellaneous appliances ; wherein the disaggregation models and the rule based models provide for a near complete appliance level itemization and estimation .