An Initial ANN Approach to LMP - PowerPoint Presentation

Slide1

An Initial ANN Approach to LMP Classification & Prediction

Honghao ZhengUniversity of Wisconsin – Madison

©

Honghao

Zheng

2010Slide2

Motivation

Locational Marginal Price (LMP), which is usually referred to as “shadow price” of the power grid, gives efficient measurement of power production and the consumption of energy at the different bus nodes.The prediction of LMP at different zonal price could benefit the individual biding for the electricity at different nodes in the power system.If we could locate the feature vector, then we could use ANN method to predict the PMU value at certain time in certain place.Slide3

Previous WORK

[1] Zonal Prices Analysis Supported by a Data Mining Based Methodology, J. Ferreira, S. Ramos, Z. Vale and J. P. Soares. IEEE Conference Proceedings, 2010.[2] Zone Clustering LMP with Location information using an Improved Fuzzy C-Mean, Se-Hwan Jang, Jin-Ho Kim, Sang-Hyuk Lee and June-Ho Park, IEEE Conference Proceedings.[3] High value wind: A method to explore the relationship between wind speed and electricity locational marginal price, Geoffrey McD. LewisSlide4

The first step of the project would be manually filtering the large amount of LMP hourly data into different groups.The LMP data is downloaded from the website of Midwest Independent System Operator (Midwest ISO).

Filters: Time = April, Value = LMP, Type = LoadZoneMethodologyStep No.1Slide5

MethodologyStep

No.2Step No.2 mainly concerns with the feature vector selection.Major Issue that may influence the value of LMP: 1. Grid structure; 2. Weekday or Weekend (7 days in one week); 3. Different period in a day (Morning/Noon/Evening)Generate physical position of different load zones;Grant different weights to the seven days;

Choose 4 hours to be one period, all have high LMP.

Feature Vector Dimension: 4Slide6

ALTEALTW

AMRNCILCCINCONSCWLDCWLP311

35

3

13

21

3

3

DECO

DPC

EKPC

FE

GREHEIPIPL24611592102IGEEMDUMGEMPNIPSNSPOTPSIGE4418426154SIPCTVAUPPCWECWPSWR TOTAL111521 238

MethodologyStep No.2(Cont’d)

Generate Geographic Location:Slide7

MethodologyStep

No.3The Step No.3 Using MLP Mapping to Test the dataClassification Criterion: <35 LOW LMP, 35~50 MID LMP, >50 HIGH LMPSeparate the 28 days in Apr into 4 weeks, labeled W1, W2, W3, W4.Formulate 3 tests: Training Set (W1, W1&W2, W1&W2&W3), Testing Set (W2&W3&W4, W3&W4, W4)Here the testing set functions as the prediction, because in the future if we know the feature vector, we could use MLP to predict the LMP value directly.Slide8

Simulation Result

Ways of TrainingLayer = 3, Neurons/Layer = 5Layer = 4, Neurons/Layer = 6Training RatePrediction Rate

Training Rate

Prediction Rate

T.1

71.47%

55.12%

83.87%

50.26%

T.2

67.98%

54.89%

67.467%53.12%T.365.55%53.20%63.15%61.50%TrainingTraining SetTesting SetT.1W1W2, W3, W4T.2W1, W2W3, W4T.3W1, W2, W3W4Comment:Training Rate does not have necessary relationship with the Prediction RatePrediction Rate (Testing Rate) is not that high as expected.The randomly-generated location may result in the inconsistency.Slide9

DiscussionDisturbance & Ways to Improve

Disturbance:Inconsistency in the locationThe classification of the LMP may be too rough to determine the exact position of LMP.Possible feature difference not quite clear.Ways to Improve:Acquire actual geographic location (longitude, latitude).Classify the LMP value range smaller.

To make the range difference between the features to be obvious.Slide10

Conclusion

ANN: quite a useful tool in the power system, yet the application of prediction for LMP value is rare.The result that has the best performance (63%) is roughly acceptable, yet not the expected value.Outlook: make the model more realistic; trying to get the location data from the government; change MLP algorithm to better suitable for LMP Prediction