Prediction of Electric Grid Frequency for a state power generation using Artificial Intelligence

(Client name and State government details are being withheld due to non-disclosure terms put forward by the client.)

Overview

A private electricity company (in partnership with the state government in  India) was looking for a platform to predict grid frequency based on pre-processed and agreed Data buckets available corresponding to the last 5 years’ frequency data at the national level as well as regional level.

  1. Expected Generation
  2. Actual Generation
  3. Expected Drawls
  4. Actual Drawls
  5. Holidays(Binary Flag)
  6. Weather(Day Avg. Temp)

Solution  

An Artificially Intelligent software solution (prediction analysis) that provides  every 15 mins projection of frequency over the next 2 hours in 8 distinct points (corresponding to 15 mins each in the next 2-hour window)

Functional level Milestones

  • Data Preprocessing
    • Data collection
    • Collation of data into a central database

 

  • Exploratory Data Analysis (For Insights)
    • Data Manipulation
    • Plotting charts is involved wherever necessary
    • Statistical Stationarity test
    • Identifying Trend and Seasonality
    • Estimating and Eliminating (Differencing) Trend
    • Decomposing (Breaking TS into different charts/TS i.e. seasonality) ARIMA model parameter identification and modelling
    • Try different version of ARIMA models and set the baseline

 

  • Data Pipeline
    • Data Augmentation (various standardized models tried)

 

  • Network Design
    • Our Team will have to try and see different Network Architecture
    • Involves research as well
      • Simple LSTM (Starting Point)
      • Bayesian Forecasting

 

  • Training Pipeline Design
    • Training Pipeline Design
    • Training Strategy

 

  • Model Training
    • Assumed Training time for multiple passes of data

 

  • Model Hyper-parameter tuning
    • Model fine tuning with various data-lag values to have the best accuracy subjective to performance
      • Model fine tuning with network architectural changes
      • Loss Function Addition
      • Data Augmentation Implementation

 

  • BackTesting (With historical data)
    • Developing Accuracy metrics for data points
    • Test on different time series samples from the past overall performance
    • Measurement metrics

 

  • Minor Feedback based adjustments

 

  • Dashboard charting for frequency trend

 

Various other techniques were also be considered while making this predictive model.

INDUSTRY
Power Sector

TECHNOLOGY
Artificial Intelligence

PROJECT ORIGIN
India

Services we helped the client with

Technique Method Assumptions Pitfalls Model Training Frequency
Time Series Forecasting using ARIMA/ETS.
Use the existing time series of frequency values to make the predictions about the future values.
The vanilla algorithms do not use the other available data to make the forecast.
The algorithm is best suited to solve a the further we into the future we forecast, the more error is generated.
Requires frequent updates, but since the method is extremely fast, shouldn’t be cumbersome.
Regression/Decision Trees that use hand made features to make predictions.
Extract features from time series (like moving average, seasonal average) and other points like temperature changes, holiday indicator, weekday indicator etc. to make the predictions.
Regression will assume linear relationships at certain points, decision tree-based methods will be more black box approaches.
Both problem versions a and b can be solved using this class of methods, but considerable time must be spent to make the features, domain knowledge might be crucial.
Monthly might be fine if there are no major changes in the variance in data.
Forecast daily and hourly demands.
Here we break the problem as a combination of day level and hour level forecasts which are combined to make the final forecast.
The basic assumption will be that the day level demand only depends on weekday type (Monday etc.) and the hourly demands will be different for different days.
The approach will see changes very late since the model must be trained frequently, works on both a and b.
A biweekly level might work since this method is not sensitive to recent changes in the data patterns, it must be retrained to understand those.

We offers enterprise-level solutions for businesses and help you strategize your digital product

Looking to incorporate Artificial Intelligence?