A predictive model of velocity for local hydrokinetic power assessment based on remote sensing data

Abstract

Hydrokinetic power is a small-scale, zero-head form of hydropower with the potential to address renewable energy needs for many communities, including those in the Arctic. An impediment to its wide deployment is the ability to screen suitable stream reaches for adequate power generation. Remote resource assessment of streams, particularly the estimation of mean velocity, is valuable for assessing the potential of new locations prior to field visits. Many studies have investigated remote sensing techniques for stream velocity prediction through machine learning and empirical relations, but most of these studies relied at least partly on in-situ data and were not suitable for small width streams. This study addresses these gaps with a Random Forest machine learning model that uses only remotely sensible data to estimate velocities of small and large streams at locations across Alaska. The developed model, validated by velocity gage readings, had a mean prediction error of 24%, at the lower end of error rates (20–30%) found for large channel widths. Novel geometric and temporal parameters were developed and found to importantly influence predictive accuracy (e.g. slope, stream meanders). Remotely sensible data enables future model expansion by those interested in assessing new sites for community energy planning.