kpi.rst

Visualisation tools

Introduction

This section show how to use the KPI module. Various examples are provided with correspnding code snippet to show different functionalities. The figure below show all the capabilities at a glance.

Configuration and Inputs

The agents required for analysis should be imported at the first stage. The following code snippet shows how to load different agents:

from grid2bench.AgentsAnalytics import AgentsAnalytics

agents = AgentsAnalytics(
    data_path=input_data_path,
    agents_names= ["PPO_Agent", "MazeRL_Agent"],
) #  episodes_names= ['dec16_1', 'dec16_2']

Available KPI

Cumulative rewards and accomplished time steps

We can visualized for each agent and for each episode the cumulative reward and total survival time step that the agent accomplished. Figure below present such results for PPO_agent and MazeRL_agent

# python code snippet to reproduce the figure
agents.plot_cumulative_reward(agents_results)

agents.plot_cumulative_reward(agents_results, fig_type='AccTimeSteps')

Action frequencies

We can visualize the frequency of different actions for each agent. In figure below we can see that in both agents below 100% of the actions are topology changes.

# python code snippet to reproduce the figure
AgentsAnalytics.plot_actions_freq_by_type(
    agents.agents_data,
    col=2,
    title = "Frequency of actions based on action types"
)

Also action frequency can be visualized based on the substations, pi-chart below shows for each agent the share of each station in all the performed actions.

 # python code snippet to reproduce the figure
AgentsAnalytics.plot_actions_freq_by_station_pie_chart(
    agents.agents_data,
    col=2,
    title = "Frequency of actions by station"
)

Let's dig into more details about the impacted stations by viewing the id of the actions that impact these stations

# Select an agent to see its actions per substation
agent = agents.agent_data[agents.agents_names.index('Tutor_binbinchen')]
# Retrieve the corresponding dataframe for a given agent
df = tutor_binbinchen.get_actions_by_substation_by_id()
# Plot it
eps.plot_actions_by_station_by_id(
    df,
    title = 'Tutor Binbinchen : Frequency of actions by substation')

Impact of actions on objects

We can visualize the impacts of actions on different objects:

Impact on lines

We can visualize overloaded lines and compare the results of different agents in a bar plot. [Using function plot_lines_impact() if parameter disconnected==False, then the function will plot overloaded lines]. Figure below presents the result of above function for overloaded lines:

# python code snippet to reproduce the figure
AgentsAnalytics.plot_lines_impact(
    agents.agents_data,
    title = "Overloaded lines",
    yaxis_type = "linear"
)

And for the same function with parameter disconnected = True, we can have the visualization of disconnected lines.

# python code snippet to reproduce the figure
AgentsAnalytics.plot_lines_impact(
    agents.agents_data,
    title = "Overloaded lines",
    fig_type = "disconnected"
)

Impact on reference topology

We can also compute and compare the distance from the initial topology at each time step of selected episode.

# python code snippet to reproduce the figure
w = widgets.Dropdown(
        options=agents.episodes_names,

        description='Episode',
    )
def f(x):
    display(AgentsAnalytics.plot_distance_from_initial_topology(agents_results,
                                                                episodes_names=[x]))

interact(f, x=w)

Impact on substations

In the same way as impact of actions on lines, we can visualized the impact of actions on the substations. Figure below presents the frequency of action by stations for two different agents.

# python code snippet to reproduce the figure
AgentsAnalytics.plot_actions_freq_by_station(
    agents.agents_data,
    title = "Frequency of actions by station",
    yaxis_type = "log"
)

Action sequence length

We can visualize the action sequence length with a widget to select the episode that we want to analyze:

# Graphic parameters
min_length= 0
max_length= 15
bg = "rgba(0,0,0,0)"

# function allowing to plot the action sequence length interactively
def f(x):
    if not type(x)==list : eps = [x]
    else : eps =x

    display(agents.plot_actions_sequence_length(
        agents_results,
        episodes_names=eps,
        min_length=min_length,
        max_length=max_length,
        plot_bgcolor = bg
    )
        )

interact(f, x=w)