-
Notifications
You must be signed in to change notification settings - Fork 0
/
README.txt
34 lines (30 loc) · 2.81 KB
/
README.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
REPO for GitHub / GitLab and Google Cloud
Set settings for QDS (quasi dynamic load flow simulation in a power grid) and elements and save results to file to
create a data set executing a QDS.
At first the grid is prepared and scenario settings are set. Then samples are created from raw data.
These samples are time series of voltages at points of connection of households and photovoltaics (PVs) of a low
voltage distribution network. Finally a deep learning approach is compared to a linear classifier to either
determine if a sample is from a term with PV (class 1) or no PV (class 0) or from a term with a regularly
behaving PV (class 0) or a PV with a malfunctioning reactive power control curve (class 1).
Additionally a dummy dataset can be created that only consists of samples that are constant over the entire
timeseries (class 0) and samples that are not (class 1). Randomly chosen samples of either classes are plotted
along with execution at default.
See framework diagrams for a better overview.
Test files are in the project folder.
Choose experiment (dataset and learning settings) in experiment_config.py
Predefined experiments vary the dataset type (dummy, PV vs no PV, regular PV vs malfunctioning PV) as well as the
timeseries length of samples (1 day vs 1 week) and the number of samples (too little, 'just enough', sufficient to
produce a meaningful output after training with the basic network design used, i.e. no Fscore ill defined because only
always one class predicted in any run of cross validation; note that 1 day vs 7 days also means increasing the amount
of data points, therefore redundant experiments (i.e. increasing the sample number even more for 1 day timeseries
experiments was neglected to allow for a better orientation between experiments)
The experiment also defines the network architecture (in the predefined experiments this is a simple 2 layer Elman
RNN with 6 hidden nodes in each layer). Multiple options are available such as changing the mini batch size, early
stopping, warm up, controlling the learning rate...
Metrics: Deep learning approach should perform better than linear classifier (which just guesses between 0 and 1 class)
meaning that a higher Fscore should be achieved
Experiment configs state if this goal can be fulfilled with the experiment settings
Task Dataset collection ANN design ANN tuning Results Report Presentation
Time planned: (Hours) 15 7.5 15 7.5 10 4
Time spent: (Hours) ~20 25 ~15 5 to be seen
Conclusion: It took much longer than planned to actually get the RNN running and producing meaningful outputs