Space Station Power Forecasting with LSTMs on FPGAs
Joseph Richard Kocik · D-Scholarship@Pitt (University of Pittsburgh) · 2021
Autonomous operations of space systems is an important and difficult task that will become even more imperative as space missions become increasingly remote. Accurate prediction of telemetry data can improve system monitoring and facilitate fault detection. This thesis presents a methodology for the acceleration of short-term forecasting of power data on an embedded platform designed for space. Initially, a long short-term memory (LSTM) network is trained to forecast voltage and current values from the International Space Station. This LSTM forecasts voltage and current minutes into the future while maintaining a low error rate. This LSTM network’s weights and biases are then used to create a new accelerated network which can be deployed on the FPGA of a Zynq-7045 system on a chip (SoC). The Zynq-7045 was selected because it is the same SoC used on the SHREC Space Processor, a space computer targeted for this study. A number of networks of varying sizes and history lengths are realized in hardware and evaluated against a software baseline. These networks were designed to be deployed on the resource-constrained FPGA fabric of the Zynq-7045 while maintaining the LSTM network architecture. The best performing LSTM networks were able to achieve over 3× speedup against a software baseline with minimal increase in forecasting error.