Geert Scholma
Energy Forecaster / Data Analyst at E.ON Benelux
E.ON Benelux
Utrecht University
https://www.linkedin.com/in/geert-scholma-656b7120/
Source:
https://branchesofmechanicalengineering.blogspot.com//search?q=whos-who-in-energy-forecasting-geert-scholmaThursday, September 14, 2017
Who's Who inwards Energy Forecasting: Geert Scholma
I got to know Geert Scholma from NPower Forecasting Challenge 2015, where he outperformed my BigDEAL students on the leaderboard. Since then, he has been topping the NPower leaderboard every time. Recently, every bit a winner of the qualifying represent of GEFCom2017, he presented his methodology at ISEA2017.
Geert lives inwards Rotterdam, The Netherlands. He has a strong focus on information scientific discipline too the liberate energy transition, amongst a masters grade inwards physics too v years sense every bit an Energy Forecaster for Energy Retail Company too E.On spin-off Uniper Benelux.
Since 2015 he has participated inwards several online liberate energy forecasting competitions, amongst the next runway record:
What brought you lot to the liberate energy forecasting profession?
Since an early on phase of my physics pedagogy at University I stimulate got been inspired past times developments inwards the Energy Transition too stimulate got directed my career path towards it. This began amongst inquiry too internships inwards the plain of solar electricity production too liberate energy service companies. My start task was at a consultancy firm, where nosotros managed liberate energy labels too liberate energy policy for social housing firms. I too so decided to aspect for a set at a large liberate energy fellowship inwards The Netherlands, but it was a coincidence that I ended upwards every bit an liberate energy forecaster. I had never heard of the term before, but the plain has proven me to hold out really interesting.
What practise you lot practise at your electrical flow job? And what's fun nearly it?
v years agone I started my task every bit an liberate energy forecaster for Uniper Benelux. My principal focus has been the evolution of novel day-ahead forecasting models for all our customers. As our portfolio consists of electricity, gas too district heating for small, medium too large clients, this is quite a various challenge. The principal task of our squad is to deal the residual responsibleness too minimize our clients' imbalance volumes too costs. Besides having to forecast consumption too production volumes, this also agency taking into trouble organisation human relationship the effects of hierarchy / portfolio too pricing the profiles of potential novel clients. The fun business office for me is squeezing the most information out of these large data. And I gauge inwards general, working amongst numbers simply makes me a happy mortal :)
What was your start (forecasting or information mining) contest about? And how did you lot do?
My start contest was the start Great Britain Npower competition. Data were a unmarried aggregated daily electricity consumption fourth dimension serial too hence relatively slowly to deal every bit I was used to locomote amongst multiple fourth dimension serial amongst an hourly resolution. I won the competition. As forecasting much to a greater extent than than 1 twenty-four hours into the time to come was novel to me I learned to non extrapolate fourth dimension trends likewise enthusiastically into the future. All competitions I stimulate got participated so far stimulate got e'er taught me similar lessons that I wouldn't stimulate got learned every bit fast inside my daily job.
Can you lot portion amongst us the most exciting contest you've participated?
The most exciting contest so far was the recent RTE Power Consumption Forecast Challenge inwards 2017. The task was to forecast the day-ahead fifteen infinitesimal electricity consumption for all 12 French Regions. The aspect that it made it to a greater extent than interesting than the other competitions was the fact that the information was existent too the solution applicable. Also the contest much tougher. The trial was concluded amongst a seminar inwards Paris where I learned that almost all of my competitors used machine learning, where my solution was mainly based on a unmarried linear regression model.
Is at that spot a primal maiden or exciting projection you lot are working on these days?
I am working on an update for the instant business office of the French RTE Competition this winter. I am focusing on an update of my base of operations model, but also machine learning too ensemble forecasting. I am curious how the battle betwixt uncomplicated linear regression too complicated dark box machine learning methods volition cease side past times side fourth dimension when I include some novel variables I already stimulate got inwards mind. Together amongst someone from IBM nosotros are also working on a novel approach to (energy) forecasting benchmarks, but this volition notwithstanding stimulate got some to a greater extent than fourth dimension to acquire concrete.
What's your forecast for the side past times side 10 years of liberate energy forecasting field?
I aspect real-time pricing too demand-side administration to acquire a pregnant novel factor inwards liberate energy forecasting. One of the electrical flow challenges is ofttimes notwithstanding to predict a yearly growing book of "behind the meter", renewable liberate energy (mostly solar) production. As renewable production volition acquire to a greater extent than too to a greater extent than hard to manage, marketplace set prices for to a greater extent than clients volition acquire flexible too to a greater extent than client groups volition hold out encouraged to either shop their ain production or shift their need towards off-peak fourth dimension hours. I aspect this to opened upwards a consummate novel too really interesting chapter inwards liberate energy forecasting.
How practise you lot pass your gratuitous time?
I am a existent outdoor sportsman too bask cycling too tennis. My partner is from Italy too nosotros ofttimes catch her set unit of measurement inwards Puglia where nosotros bask the food, set unit of measurement too beautiful coast too countryside.
Geert lives inwards Rotterdam, The Netherlands. He has a strong focus on information scientific discipline too the liberate energy transition, amongst a masters grade inwards physics too v years sense every bit an Energy Forecaster for Energy Retail Company too E.On spin-off Uniper Benelux.
Since 2015 he has participated inwards several online liberate energy forecasting competitions, amongst the next runway record:
- 1st place Npower Electricity Demand Forecasting Challenge 2015
- 1st place NPower Gas Demand Forecasting Challenge 2015
- 1st place BigDEAL Forecasting Competition 2015
- 1st place NPower Electricity Demand Forecasting Challenge 2016
- 1st set opened upwards runway of qualifying match, Global Energy Forecasting Competition 2017
- 4th set RTE Power Consumption Forecast Challenge 2017
What brought you lot to the liberate energy forecasting profession?
Since an early on phase of my physics pedagogy at University I stimulate got been inspired past times developments inwards the Energy Transition too stimulate got directed my career path towards it. This began amongst inquiry too internships inwards the plain of solar electricity production too liberate energy service companies. My start task was at a consultancy firm, where nosotros managed liberate energy labels too liberate energy policy for social housing firms. I too so decided to aspect for a set at a large liberate energy fellowship inwards The Netherlands, but it was a coincidence that I ended upwards every bit an liberate energy forecaster. I had never heard of the term before, but the plain has proven me to hold out really interesting.
What practise you lot practise at your electrical flow job? And what's fun nearly it?
v years agone I started my task every bit an liberate energy forecaster for Uniper Benelux. My principal focus has been the evolution of novel day-ahead forecasting models for all our customers. As our portfolio consists of electricity, gas too district heating for small, medium too large clients, this is quite a various challenge. The principal task of our squad is to deal the residual responsibleness too minimize our clients' imbalance volumes too costs. Besides having to forecast consumption too production volumes, this also agency taking into trouble organisation human relationship the effects of hierarchy / portfolio too pricing the profiles of potential novel clients. The fun business office for me is squeezing the most information out of these large data. And I gauge inwards general, working amongst numbers simply makes me a happy mortal :)
What was your start (forecasting or information mining) contest about? And how did you lot do?
My start contest was the start Great Britain Npower competition. Data were a unmarried aggregated daily electricity consumption fourth dimension serial too hence relatively slowly to deal every bit I was used to locomote amongst multiple fourth dimension serial amongst an hourly resolution. I won the competition. As forecasting much to a greater extent than than 1 twenty-four hours into the time to come was novel to me I learned to non extrapolate fourth dimension trends likewise enthusiastically into the future. All competitions I stimulate got participated so far stimulate got e'er taught me similar lessons that I wouldn't stimulate got learned every bit fast inside my daily job.
Can you lot portion amongst us the most exciting contest you've participated?
The most exciting contest so far was the recent RTE Power Consumption Forecast Challenge inwards 2017. The task was to forecast the day-ahead fifteen infinitesimal electricity consumption for all 12 French Regions. The aspect that it made it to a greater extent than interesting than the other competitions was the fact that the information was existent too the solution applicable. Also the contest much tougher. The trial was concluded amongst a seminar inwards Paris where I learned that almost all of my competitors used machine learning, where my solution was mainly based on a unmarried linear regression model.
Is at that spot a primal maiden or exciting projection you lot are working on these days?
I am working on an update for the instant business office of the French RTE Competition this winter. I am focusing on an update of my base of operations model, but also machine learning too ensemble forecasting. I am curious how the battle betwixt uncomplicated linear regression too complicated dark box machine learning methods volition cease side past times side fourth dimension when I include some novel variables I already stimulate got inwards mind. Together amongst someone from IBM nosotros are also working on a novel approach to (energy) forecasting benchmarks, but this volition notwithstanding stimulate got some to a greater extent than fourth dimension to acquire concrete.
What's your forecast for the side past times side 10 years of liberate energy forecasting field?
I aspect real-time pricing too demand-side administration to acquire a pregnant novel factor inwards liberate energy forecasting. One of the electrical flow challenges is ofttimes notwithstanding to predict a yearly growing book of "behind the meter", renewable liberate energy (mostly solar) production. As renewable production volition acquire to a greater extent than too to a greater extent than hard to manage, marketplace set prices for to a greater extent than clients volition acquire flexible too to a greater extent than client groups volition hold out encouraged to either shop their ain production or shift their need towards off-peak fourth dimension hours. I aspect this to opened upwards a consummate novel too really interesting chapter inwards liberate energy forecasting.
How practise you lot pass your gratuitous time?
I am a existent outdoor sportsman too bask cycling too tennis. My partner is from Italy too nosotros ofttimes catch her set unit of measurement inwards Puglia where nosotros bask the food, set unit of measurement too beautiful coast too countryside.
=================================================
http://blog.drhongtao.com/search/label/npower
Tuesday, Dec 20, 2016
Winning Methods from npower Forecasting Challenge 2016
RWE npower released the finally leaderboard for its forecasting challenge 2016. I took a shroud shot of the exceed teams. Interestingly, the international teams (colored inwards red) took over all of the exceed vi places. Unfortunately, some of those top-notch Great Britain charge forecasters did non bring together the competition. I'm hoping that they tin shipping away demo upwards at the game to defend the country's legacy:)
In each of the previous 2 npower competitions, I asked my BigDEAL students to bring together the contest every bit a team. In both competitions, they were ranked exceed too beating all Great Britain teams (see the weblog posts HERE and HERE). We also published our winning methods for HERE.)
OK, plenty bragging...
I asked the exceed teams portion their methodologies amongst the audience of my weblog every bit what nosotros did in BFCom2016s. Here they are:
1st Place: Geert Scholma
My forecast this fourth dimension consisted of the next elements:
- linear regression models seperated per thirty infinitesimal menstruation amongst 78 variables each
- 4th grade yearly shapes per weekday every bit a base of operations shape
- an intercept, vi weekdays too 22 holiday, bridgeday too schoolholiday variables
- daylight savings too a linear timetrend, each seperated for weekdays too weekends
- a shift at september 2014 too a dark variable
- conversion of temperature to windchill
- 3rd grade windchill polynomials for cooling too heating amongst dissimilar impacts
- iii moving averages amongst dissimilar periods for temperature effects occurring at dissimilar timescales
- dissimilar radiations variables depending on fourth dimension of twenty-four hours amongst upwards to vi hourly too moving average radiations variables interacted amongst a instant grade polynomial of the twenty-four hours of yr for peak hours
- 1 hourly too 1 moving average rainfall variable
- manually exclusion of outliers too filling of whatever atmospheric condition gaps
2nd Place: Devan Patel
Model: Multiple linear regression approach was used during the NPower forecasting competition. The basic model was Tao’s Vanilla Benchmark model. Influenza A virus subtype H5N1 major alter was made inwards the cast of subject variable Energy Consumption. Influenza A virus subtype H5N1 Box-Cox transformation of Energy Consumption was taken based on the develop information distribution. Polynomials of Humidity too Wind Speed were added into the Base model. With the assistance of this changes the performance of the benchmark vanilla model was improved. During testing inwards a higher set changes were successfully able to ameliorate the accuracy of vanilla model past times around 1.5% on the scale of MAPE.
Data: Two dissimilar approaches were used inwards society to develop the model. During wintertime (Round 1 too Round 3) model was trained using whole year’s data. During summertime (Round 2) entirely summertime month’s information was used during model training. Scatter plots across dissimilar months were helpful to empathize the distribution of liberate energy consumption.
Explanatory information analysis: The missing values of the hours were replaced past times previous day's hours. Scatter plots of temperature, humidity too air current speed were used to set their relationships amongst liberate energy consumption.
Error matrix: MAPE was used every bit a base of operations mistake matrix inwards society to evaluate the accuracy of the forecast during model validation.
Software: RStudio was used every bit a principal software for model building, validation too forecasting. MS Excel was used to laid upwards the information files which tin shipping away hold out used inwards RStudio.
3rd Place: Masoud Sobhani
For the start round, the model was Tao's Vanilla model amongst recency effects (by adding extra lagged temperature to the original model). The model uses MLR method too the predictors are calendar variables, temperature, lagged values of temperature too cross effects betwixt them. The model was implemented inwards SAS. For the instant round, I tried to ameliorate Vanilla model past times adding to a greater extent than predictors beyond the temperature. Humidity was added to the model past times using the method introduced in Xie too Hong 2016. The novel model was an improved model having temperature too relative humidity every bit atmospheric condition related predictors. Since nosotros didn't know the location of the utility, I tried to alter the novel model to select the perfect model amongst the best results. For the 3rd round, the model used inwards previous circular was improved past times adding some lagged values of relative humidity. In each round, the model pick was done past times cross validation method.
 |
| RWE npower Forecasting Challenge 2016 Final Leaderboard (top 12 places) |
In each of the previous 2 npower competitions, I asked my BigDEAL students to bring together the contest every bit a team. In both competitions, they were ranked exceed too beating all Great Britain teams (see the weblog posts HERE and HERE). We also published our winning methods for HERE.)
OK, plenty bragging...
I asked the exceed teams portion their methodologies amongst the audience of my weblog every bit what nosotros did in BFCom2016s. Here they are:
1st Place: Geert Scholma
My forecast this fourth dimension consisted of the next elements:
- linear regression models seperated per thirty infinitesimal menstruation amongst 78 variables each
- 4th grade yearly shapes per weekday every bit a base of operations shape
- an intercept, vi weekdays too 22 holiday, bridgeday too schoolholiday variables
- daylight savings too a linear timetrend, each seperated for weekdays too weekends
- a shift at september 2014 too a dark variable
- conversion of temperature to windchill
- 3rd grade windchill polynomials for cooling too heating amongst dissimilar impacts
- iii moving averages amongst dissimilar periods for temperature effects occurring at dissimilar timescales
- dissimilar radiations variables depending on fourth dimension of twenty-four hours amongst upwards to vi hourly too moving average radiations variables interacted amongst a instant grade polynomial of the twenty-four hours of yr for peak hours
- 1 hourly too 1 moving average rainfall variable
- manually exclusion of outliers too filling of whatever atmospheric condition gaps
2nd Place: Devan Patel
Model: Multiple linear regression approach was used during the NPower forecasting competition. The basic model was Tao’s Vanilla Benchmark model. Influenza A virus subtype H5N1 major alter was made inwards the cast of subject variable Energy Consumption. Influenza A virus subtype H5N1 Box-Cox transformation of Energy Consumption was taken based on the develop information distribution. Polynomials of Humidity too Wind Speed were added into the Base model. With the assistance of this changes the performance of the benchmark vanilla model was improved. During testing inwards a higher set changes were successfully able to ameliorate the accuracy of vanilla model past times around 1.5% on the scale of MAPE.
Data: Two dissimilar approaches were used inwards society to develop the model. During wintertime (Round 1 too Round 3) model was trained using whole year’s data. During summertime (Round 2) entirely summertime month’s information was used during model training. Scatter plots across dissimilar months were helpful to empathize the distribution of liberate energy consumption.
Explanatory information analysis: The missing values of the hours were replaced past times previous day's hours. Scatter plots of temperature, humidity too air current speed were used to set their relationships amongst liberate energy consumption.
Error matrix: MAPE was used every bit a base of operations mistake matrix inwards society to evaluate the accuracy of the forecast during model validation.
Software: RStudio was used every bit a principal software for model building, validation too forecasting. MS Excel was used to laid upwards the information files which tin shipping away hold out used inwards RStudio.
3rd Place: Masoud Sobhani
For the start round, the model was Tao's Vanilla model amongst recency effects (by adding extra lagged temperature to the original model). The model uses MLR method too the predictors are calendar variables, temperature, lagged values of temperature too cross effects betwixt them. The model was implemented inwards SAS. For the instant round, I tried to ameliorate Vanilla model past times adding to a greater extent than predictors beyond the temperature. Humidity was added to the model past times using the method introduced in Xie too Hong 2016. The novel model was an improved model having temperature too relative humidity every bit atmospheric condition related predictors. Since nosotros didn't know the location of the utility, I tried to alter the novel model to select the perfect model amongst the best results. For the 3rd round, the model used inwards previous circular was improved past times adding some lagged values of relative humidity. In each round, the model pick was done past times cross validation method.
You powerfulness also like: