Transcript

SUNSEED project is partially funded by EC FP7 programme under grant agreement #619437.

Big Data StreamMining• Maintain summaries of the streams, sufficient to answer the

expected queries about the data:• Summaries can be in various forms: clusters (flat orhierarchic, statistical aggregates, …)

• Maintain a sliding window of the most recently arriveddata operations on a sliding window mimic moretraditional database/mining operations

• Sampling ‐ obtain representative data sample (i.e., enabling toperform correctly required operations on data)

• Smart sampling (x % from stream of multiple datasources; alternative ‐ take y % of selected data sources)

• Similarity comparison – smart indexing• Incremental updating of predicting models

M. Skrjanc, B. Kazic{Maja.Skrjanc, Blaz.Kazic}@ijs.si , 

Jozef Stefan Institute, Jamova ul. 39, Ljubljana, Slovenia

Forecasting in Smart Grids• Types of forecasting problems:

• Electricity load (short term, medium term, long term)• Renewable sources generation• Electricity prices• Costumer segmentation

• Input sources:• Historical load variables: used for learning models and

detecting short term trends• Meteorological data: known to be correlated with load

(depends on location)• Static data: such as special calendar data (holidays,

summer season), and topology of electrical grid• Methods used:

• Naive approach: Localized averages, previous values.Computationally non demanding, fast, robust and easyto maintain. Can work surprisingly well.

• Classical approaches: Autoregressive (ARMIA),regression‐based statistics methods. Based on historicaldata. Can take advantage of seasonality trends, butusually don’t include other data sources.

• Computational intelligence approaches: Artificialneural networks, support vector machines. Data drivenapproach that can take advantage of variousheterogeneous data sources.

• Hybrid methods: combine two or moredifferent approaches in order to take advantageof specific methods benefits and overcometheir drawbacks.

Recommended