Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
SlideShare a Scribd company logo

Time series forecasting with machine learning

•Download as PPTX, PDF•
•

An introduction of developing and application time series forecast models with both traditional time series methods and machine learning techniques. Case study for a challenging very short-term electrical price forecasting project was presented.

1 of 23
Time Series Forecasting with Machine Learning See the future with a crystal ball Dr Wei Liu Data Science Lead
Time series forecasting with machine learning
29 October 2018 | 3 What is forecasting
Time Series and its Forecasting • A time-series is a set of observations on a quantitative variable collected over time. • Examples • Stock: Dow Jones Industrial Averages • Marking: sales, inventory, and customer counts etc • Economics: Interest rates, GDP, and employment etc. • Energy (Electricity, Gas, Oil, and Solar) demands and prices etc. • Weather: e.g., local and global temperature etc. • Sensors: Internet-of-Things • Businesses are often very interested in forecasting time series variables. • In time series analysis, we analyze the past behavior of a variable in order to predict its future behavior 29 October 2018 | 4
Approaches for Time Series Forecasting Classical Time Series Analysis Methods • Naïve, SNaïve • Seasonal decomposition (+ any model) • Exponential smoothing • ARIMA, SARIMA • GARCH • Dynamic linear models • TBATS • Prophet 29 October 2018 | 5 Machine Learning Methods • Generalised Linear Modelling (GLM) • Gradient Boost Machine (GBM) • Random Forest (DF) • Deep Learning (DL) • Automated Machine Learning (AutoML)
The Classical Approaches: Pros and Cons 29 October 2018 | 6
The Classical Approaches: Notable Disadvantages 29 October 2018 | 7
Machine Learning Approaches: Pros and Cons 29 October 2018 | 8
What we have done with time series forecast modelling? A Real and Challenging Project Our clients engages us to develop a short-range electricity price forecasting tool for optimising their operation/production. The problem is quite challenging as the data is really dynamic with rapid variations without clear trends and seasonality. Our solution is to investigate intensively with available classic and ML approaches for the problem and identify the best approach with the most accurate forecastings. The accuracies of less than 10% for forecasting electricity price 24 hours in 5-minute interval in advance with the best model identified has achieved. 29 October 2018 | 9
Forecasting Examples 29 October 2018 | 10
Forecasting Examples 29 October 2018 | 11
Forecasting Examples 29 October 2018 | 12
Forecasting Examples 29 October 2018 | 13
Forecasting Examples 29 October 2018 | 14
What we have done with time series forecast modelling? A powerful Cloud-Computing Tool for Time Series Forecast Modelling • Easy to use with interactive graphic user interface • Interactive data exploring and visualization • Process and prepare data for forecast modelling • State-of-the-art classical and machine learning time series forecasting algorithms. • Automatically tuning learning parameters using repeated cross-valdidation. • Benchmark experiments with different models and measures 29 October 2018 | 15
More Forecasting Tool Screenshots 29 October 2018 | 16
More Forecasting Tool Screenshots 29 October 2018 | 17
More Forecasting Tool Screenshots 29 October 2018 | 18
More Forecasting Tool Screenshots 29 October 2018 | 19
What we have done with time series forecast modelling? A Production Level Interactive Tool for Implementing Time Series Forecasting Models • Load input data from multiple sources (csv data file, Google Spreadsheet and cloud database) • Carrying out forecasting analysis with high- performance cloud computing server • Interactive view of forecasting results • Export and download forecasting results 29 October 2018 | 20
What we have done with time series forecast modelling? A Production Level Interactive Tool for Implementing Time Series Forecasting Models 29 October 2018 | 21
Use cases for future opportunities 29 October 2018 | 22
www.ghd.com/digital

More Related Content

Time series forecasting with machine learning

  • 1. Time Series Forecasting with Machine Learning See the future with a crystal ball Dr Wei Liu Data Science Lead
  • 3. 29 October 2018 | 3 What is forecasting
  • 4. Time Series and its Forecasting • A time-series is a set of observations on a quantitative variable collected over time. • Examples • Stock: Dow Jones Industrial Averages • Marking: sales, inventory, and customer counts etc • Economics: Interest rates, GDP, and employment etc. • Energy (Electricity, Gas, Oil, and Solar) demands and prices etc. • Weather: e.g., local and global temperature etc. • Sensors: Internet-of-Things • Businesses are often very interested in forecasting time series variables. • In time series analysis, we analyze the past behavior of a variable in order to predict its future behavior 29 October 2018 | 4
  • 5. Approaches for Time Series Forecasting Classical Time Series Analysis Methods • Naïve, SNaïve • Seasonal decomposition (+ any model) • Exponential smoothing • ARIMA, SARIMA • GARCH • Dynamic linear models • TBATS • Prophet 29 October 2018 | 5 Machine Learning Methods • Generalised Linear Modelling (GLM) • Gradient Boost Machine (GBM) • Random Forest (DF) • Deep Learning (DL) • Automated Machine Learning (AutoML)
  • 6. The Classical Approaches: Pros and Cons 29 October 2018 | 6
  • 7. The Classical Approaches: Notable Disadvantages 29 October 2018 | 7
  • 8. Machine Learning Approaches: Pros and Cons 29 October 2018 | 8
  • 9. What we have done with time series forecast modelling? A Real and Challenging Project Our clients engages us to develop a short-range electricity price forecasting tool for optimising their operation/production. The problem is quite challenging as the data is really dynamic with rapid variations without clear trends and seasonality. Our solution is to investigate intensively with available classic and ML approaches for the problem and identify the best approach with the most accurate forecastings. The accuracies of less than 10% for forecasting electricity price 24 hours in 5-minute interval in advance with the best model identified has achieved. 29 October 2018 | 9
  • 15. What we have done with time series forecast modelling? A powerful Cloud-Computing Tool for Time Series Forecast Modelling • Easy to use with interactive graphic user interface • Interactive data exploring and visualization • Process and prepare data for forecast modelling • State-of-the-art classical and machine learning time series forecasting algorithms. • Automatically tuning learning parameters using repeated cross-valdidation. • Benchmark experiments with different models and measures 29 October 2018 | 15
  • 16. More Forecasting Tool Screenshots 29 October 2018 | 16
  • 17. More Forecasting Tool Screenshots 29 October 2018 | 17
  • 18. More Forecasting Tool Screenshots 29 October 2018 | 18
  • 19. More Forecasting Tool Screenshots 29 October 2018 | 19
  • 20. What we have done with time series forecast modelling? A Production Level Interactive Tool for Implementing Time Series Forecasting Models • Load input data from multiple sources (csv data file, Google Spreadsheet and cloud database) • Carrying out forecasting analysis with high- performance cloud computing server • Interactive view of forecasting results • Export and download forecasting results 29 October 2018 | 20
  • 21. What we have done with time series forecast modelling? A Production Level Interactive Tool for Implementing Time Series Forecasting Models 29 October 2018 | 21
  • 22. Use cases for future opportunities 29 October 2018 | 22