Changes for TODO.md: 11 added lines, 0 removed lines.
Original line number
Diff line number
Diff line
Research:
* Finding temporal causal relationships in a time series. Instead of instantaneous `input-output` correlation measures, how to ascertain `input-delay-output` relationships in data.
Analysis:
* Generate new neural network model based on full-year's data.
Changes for docs/4-preprocessing.md: 3 added lines, 1 removed line.
Original line number
Diff line number
Diff line
@@ -115,7 +115,9 @@ These operations are defined in `thermo.py` and called in `preprocess.py` to cle
### Missing/zero power values
In several measurements power is recorded as 0. However this may not reflect the actual state of the system and may simply be an error in logging/networking or a temporary fluctuation. Because the cause of a zero measurement is indeterminate, such records are not considered during analysis.
~~In several measurements power is recorded as 0. However this may not reflect the actual state of the system and may simply be an error in logging/networking or a temporary fluctuation. Because the cause of a zero measurement is indeterminate, such records are not considered during analysis.~~
Using the flag `--keep_zeros` for `preprocess.py` will keep rows where power values are 0.
In essence, the system is analyzed only at the states when all components (chiller, water pumps, fans) are operational.
Changes for docs/5-dataset.md: 1 added line, 1 removed line.
Original line number
Diff line number
Diff line
@@ -69,5 +69,5 @@ And the following derived fields:
## Datasets
For each of the 2 chillers (1 & 2), the data were recorded for six months from Spring Equinox (March 20, 2018) through Fall Equinox (September 23, 2018). Measurements were recorded at a 5 minute interval.
For *Chiller 1* at the Engineering Science Building, parameter readings were recorded from January 1, 2018 through December 31, 2018. Measurements were recorded at a 5 minute interval.
Changes for docs/6-trends.md: 7 added lines, 1 removed line.
Original line number
Diff line number
Diff line
@@ -3,7 +3,7 @@ title: 'Trends and metrics'
order:6
---
An aggregate analysis of measurements over days reveals trends in sensor readings. The following trends were obtained by averaging measurements for a time across all days in the dataset.
An aggregate analysis of measurements over days reveals trends in sensor readings. The following trends were obtained by averaging measurements for each week over the year. Blank periods correspond to invalid data that was cleared out during pre-processing.
## Temperature
@@ -50,3 +50,9 @@ During warmer parts of the day, the gap between ambient and wet-bulb temperature
A plot of the control variable (fan speed) and the resulting power consumption shows that while power consumption generally tracks fan speed, additional variables also come into play. A moving average over a 1 hour window of mean fan A and B consumptions is shown for illustration.
This also shows that both fans in a cooling tower are identically controlled.
## Distribution of fan speed controls
Fan modes overwhelmingly are either close to 100% power or switched off.
Changes for docs/7-relationships.md: 3 added lines, 5 removed lines.
Original line number
Diff line number
Diff line
@@ -19,13 +19,11 @@ Normalized [mutual information][3] was calculated for each pair of variables. A
### Fan speed and power consumption
## Causality
*Hypothesis*: Fan power, `PowFan[A | B]` depends on ambient temperature `TempAmbient`, relative humidity `PerHumidity`, and fan speed setting `PerFreqFan[A | B]`.
### Granger causality test
A pipeline was set up where first the three features were normalized to [0-1] range. Then they were trained on a 90-10 training testing split. The coefficient of determination, $R^2 = 0.04$ indicating that the model simply predicts the mean power consumption. There is no strong linear relationship between the features and power consumption. *However*, this model is not conclusive as the data contain measurements mostly for when fan speed is near 100%. The model, as it is, simply shows that it cannot capture the noise in measurements - which is to be expected.
