Our team has built a virtual nuclear power plant that works in real time. We set the goal to squeeze as much power out of it as possible without causing a meltdown. Which means it has to safely operate at its maximum capacity. See what happened!
AI reads reactor core temperature and controls the position of the fuel rods. After the rods are moved, reactor takes some time to settle.
Coolant Pump 1
AI reads and controls pump’s speed. This pump moves heat from the core to the heat exchanger.
AI reads heat exchanger temperature, but can’t control it directly. Heat exchanger creates steam for the turbine to generate electricity.
Coolant Pump 2
AI reads and controls pump’s speed. This pump returns water from cooling tower back into heat exchanger.
AI reads cooling tower temperature, but can’t control it directly. Cooling tower condenses steam back into water.
AI reads steam turbine RPMs, but can’t control it directly. Steam turbine turns the generator.
AI reads AC generator’s RPM, but can’t control it directly. Generator creates electrical power. AI is trying to maximize the amount of power created.
The agent sees the state as displayed on the left and sets 3 parameters below. AI learns without any data, purely by experiencing the power plant at each time step. When control parameters are changed, the reactor takes some time to settle in.
This is a nonlinear, non-stationary system with hidden environments.
The simulation is sped up – each update equals one hour of real time.