MARK PEARL

Renewable Energy Client

• Lead data scientist for Wind Farm predictive maintenance initiative for large wind farm in Southern Ontario. Developed data & feature engineering pipelines to bring in petabyte scale dataset containing various SCADA sensor readings for both turbine and farm performance. Develop unsupervised clustering algorithms to group turbines based on key contributing factors to performance like rotor rpm, wind speed, wind direction and turbulence intensity. Developed supervised algorithms to measure active power production, and assess if static yaw misalignment could be detected in a given turbine. Through a power curve analysis, the loss in AEP for underperforming turbines was quantified in order to inform the operations team which turbines to assess and monitor
• Data science lead for engagement to optimize scheduling of surplus water flows for hydro station in Pennsylvania. Lead a team of several data scientists and data engineers, and designed end to end architecture on AWS. Optimization was responsible for utilizing developed price forecasts, and displacing additional flows to most lucrative market (regulation vs real-time). This project contributed to an additional 700k USD in revenue to date for it's 6 months in production
• Lead data scientist for time series forecasting engagement for various wind assets in ERCOT (Texas). Scope was to forecast both the day-ahead and real-time (spot market) nodal prices for 2 hub and 2 wind farm locations. Quantile regression forecasts were developed using XGBoost, LightGBM and TemporalFusionTransformers. For the supervised methods, time-series cross validation was leveraged in the model training and tuning process. Hyperopt was used for hyperparameter tuning using bayesian optimization framework. End-to-end MLOps framework was developed in MLFLow for engagement to track key validation metrics in the training runs. This process and code was developed into a reusable artifact leveraged by other teams across the firm. Each trained model was spun up with a serving endpoint, in order for short-term traders to experiment with the forecasted values for various features, and develop what if scenarios for under or over performing wind, load and solar generation. The forecasts with model serving were exposed in Tableau, and used for virtual trade submissions every day by the trading team

Canadian Government

• ML engineer for engagement to build classification engine to flag narcotics and firearms for Canadian border services client. This new model provided a 68% increase in effectiveness of contraband detection compared to the previous business process. Work included ingestion and preparation of required data sources, feature engineering, model development and model validation. The top performing models were XGBoost and RandomForest which were later deployed to production in AWS Sagemaker through a custom docker image and endpoint. AWS Lambda function was then developed to be used to retrieve model inference for front end mobile application via API Gateway trigger

US NHL Team

• Lead engineer for initiative aimed at building the next generation hockey analytics platform for a US based NHL team. Built pipelines in combination with Azure Databricks and Azure Batch for streaming and batch sources to build out lambda ingestion architecture. Refactored expected goals model into spark ml to provide real-time information for game outcomes. This platform was used to conduct analysis that informed player trades, matchups, and upcoming draft decisions.

Mentored entrepreneurs looking to take their business to the next level, and required advisement in data science and cloud architecture.

Consultant working across several analytics projects across various insurance, telecommunications and government clients.