WWWUniversityUnityMachine LearningIndustryGaming : MixGaming : MinecraftGaming : From The Depths : ModdingGaming : From The Depths : FilmingGaming : From The Depths : DesignsArtworkArchitecture

Machine Learning

Unity : Power Grid : Genetic Neural Network

Posted: 2020-05-25 12:45:37

Project goal is to simulate behavior of power grid (very simplified), using Evolving Neural Network approach.

The concept of using Neural Network for controlling power production, based on demand and associated cost, is to prove the concept, if such system can respond accordingly to many variables, ensuring that power demand is met, yet reducing power overproduction across multiple power plants.

Due to short deadline for this project, power grid design has been extremely simplified, in respect to initial goal.

Network of substations has been narrowed into single substation. Hence, no step up/down power grid is considered. Power level across system is an arbitrary value as Mega Watts for simplification, rather than realistic power grid values of M/G Watts, on each system point. There are no other sub stations, hence no power redirection, no line overload, or no power grid failure scenarios, in this state.

2 consumers are power delivery points. Their produce a power demand, with two peak times spikes, morning and afternoon.

System includes 3 power plants:

Solar Power Station, which for simplification, produces energy only during day time, with increasing output at morning and increasing output in the evening. Additional weather events and its forecasting are not taken int consideration.

Large Power Station has relative slow and lagging response to the power demand change. Either significant overproduction is required, or supporting power station, with faster response.

Small Power Station is supporting power station, to its larger version. Its maximal power output is much lower, but its response is much faster, to meet volaille power demand changes. However, power production cost is also much higher. So it would be desired, to reduce its output, when adequate power demand is lower.

Click for more details ...

Project uses Evolving Neural Network - NEAT, based on Evolving Neural Networks through Augmenting Topologies.pdf (NEAT Paper by Kenneth O. Stanley and Risto Miikkulainen)


Author: Antypodish
Posted: 2020-05-25 22:57:22

Default power grid system response without neural network

System reacting immediately to power demand change

Red dashed line, indicates changes in power demand, in span of 24 hours.
Light blue line indicates total power generation, with visible power deficiency at peak times 5-10 AM and 15-19 PM, caused by late response of power generation, by large (blue dashed line) and small (purple dashed line) power stations.
Each power station has own time response, of how fast it can change its output.

While for untrained system example, total power generation cost is lower, than for neural network trained system, in normal cases, positive power demand (green dashed / pink lines) would cause blackouts.
Possible solution, is to increase absolute margin, at which power must be generated. Or encourage system, to increase power production earlier, before peaks.

NN can come with an aid, to assist with a problem.

Trained data using Genetic Neural Network

It need to be noted, that there is place for improvement, on training this system.

Neural network trained systems shows, that no positive power deficiency is present, at cost of higher power output.
However, system attempts to respond to power demand changes, specially by predicting peak times.
System also reacts, to changes of solar power generation output, by reducing power generation from small and large power station.

Example of 80 simultaneously trained Power Grid systems.

NN uses 3 sets of layers. 26 nodes in input layer, which takes current power production - power demand, total power generation cost, and 24 inputs as on / off state per each hour.

2 nodes in output layer, which corresponds to power generation setpoint.

Hidden layer, of vary count of nodes, depending on trained net. Number of hidden layer nodes is independent from design definition and its setup can dynamically change during training. NN Training depending on complexity and parameters, may last from few minutes, to few hours.