Title: SHORT TERM LOAD FORECASTING USING NEURAL NETWORKS AND FUZZY LOGIC
1SHORT TERM LOAD FORECASTING USING NEURAL
NETWORKS AND FUZZY LOGIC
- George G Karady
- Arizona State University
2Short Term Load ForecastingContent
- Overview of Short term load forecasting
- Introduction
- Definitions and expected results
- Importance of Short-term load forecasting
- Impute data and system parameters required for
load forecasting - Concept
- Major load forecasting techniques
- Concept of STLF model Development
- Statistical methods
- (Multiply liner regression,
- stochastic time series e.t.c)
3Short Term Load ForecastingContent
- Artificial Neural Networks
- Building block a feed forward network
- Load forecasting engine
- Results
- Numerical example
- Fuzzy logic and evolutionary programming
4Overview of Short Term Load Forecasting (STLF)
5Introduction
- The electrical load increases about 3-7 per year
for many years. - The long term load increase depends on the
population growth, local area development,
industrial expansion e.t.c. - The short term load variation depends on weather,
local events, type of day (Weekday or Holiday or
Weekend) e.t.c.
6Introduction
- The building of a power plant requires
- 10 years (Nuclear)
- 6 years (Large coal-fired)
- 3 years (combustion turbine)
- The electric system planing needs the forecast of
the load for several years. - Typically the long term forecast covers a period
of 20 years
7Introduction
- The planning of maintenance, scheduling of the
fuel supply etc. calls for medium term load
forecast . - The medium term load forecast covers a period of
a few weeks. - It provides the peak load and the daily energy
requirement
8Introduction
- The number of generators in operation, the start
up of a new unit depends on the load. - The day to day operation of the system requires
accurate short term load forecasting.
9Introduction
- Typically the short term load forecast covers a
period of one week - The forecast calculates the estimated load for
each hours of the day (MW). - The daily peak load. (MW)
- The daily or weekly energy generation. (MWh)
10Introduction
- The utilities use three types of load
forecasting - Long term (e.g. 20 years)
- Medium term. (e.g. 3-8 weeks)
- Short term (e.g. one week)
- This lecture presents the short term load
forecasting techniques
11Definitions and Expected Results(Big picture)
- The short term load forecasting provides load
data for each hour and cover a period of one
week. - The load data are
- hourly or half-hourly peak load in kW
- hourly or half-hourly values of system energy in
kWh - Daily and weekly system energy in kWh
12Definitions and Expected Results(Big picture)
- The short term load forecasting is performed
daily or weekly. - The forecasted data are continuously updated.
- Typical short-term, daily load forecast is
presented in the Table in the next page. (Salt
River Project, SRP)
13Definitions and Expected Results(Big picture)
- Typical short-term, daily load forecast from Salt
River Project. Hourly Load in MW
14Definitions and Expected Results(Big picture)
- Typical short-term, daily load forecast from Salt
River Project. Hourly Load in MW
15Importance of Short-term Load Forecasting
- Provide load data to the dispatchers for economic
and reliable operation of the local power system. - The timeliness and accuracy of the data affects
the cost of operation. - Example The increase of accuracy of the
forecast by 1 reduced the operating cost by L
10M in the British Power system in 1985
16Importance of Short-term Load Forecasting
- The forecasted data are used for
- Unit commitment.
- selection of generators in operation,
- start up/shut down of generation to minimize
operation cost - Hydro scheduling to optimize water release from
reservoirs - Hydro-thermal coordination to determine the least
cost operation mode (optimum mix)
17Importance of Short-term Load Forecasting
- The forecasted data are used for
- Interchange scheduling and energy purchase.
- Transmission line loading
- Power system security assessment.
- Load-flow
- transient stability studies
- using different contingencies and the predicted
loads.
18Importance of Short-term Load Forecasting
- These off-line network studies
- detect conditions under which the system is
vulnerable - permit preparation of corrective actions
- load shedding,
- power purchase,
- starting up of peak units,
- switching off interconnections, forming islands,
- increase spinning and stand by reserve
19Impute Data and System Parameters for Load
Forecasting
- The system load is the sum of individual load.
- The usage of electricity by individuals is
unpredictable and varies randomly. - The system load has two components
- Base component
- Randomly variable component
20Impute Data and System Parameters for Load
Forecasting
- The factors affecting the load are
- economical or environmental
- time
- weather
- Unforeseeable random events
21Impute Data and System Parameters for Load
Forecasting
- Economical or environmental factors
- Service area demographics (rural, residential)
- Industrial growth.
- Emergence of new industry, change of farming
- Penetration or saturation of appliance usage
- Economical trends (recession or expansion)
- Change of the price of electricity
- Demand side load management
22Impute Data and System Parameters for Load
Forecasting
- The time constraints of economical or
environmental factors are slow, - measured in years.
- This factors explains the regional variation of
the load model (New York vs. Kansas) - The load model depends on these slow changing
factors and has to be updated periodically
23Impute Data and System Parameters for Load
Forecasting
- Time Factors affecting the load
- Seasonal variation of load (summer, winter etc.).
The load change is due to - Change of number of daylight hours
- Gradual change of average temperature
- Start of school year, vacation
- Calls for a different model for each season
24Impute Data and System Parameters for Load
Forecasting
- Typical Seasonal Variation of Load
- Summer peaking utility
25Impute Data and System Parameters for Load
Forecasting
- Time Factors affecting the load
- Daily variation of load. ( night, morning,etc)
26Impute Data and System Parameters for Load
Forecasting
- Weekly Cyclic Variation
- Saturday and Sunday significant load reduction
- Monday and Friday slight load reduction
- Typical weekly load pattern
27Impute Data and System Parameters for Load
Forecasting
- Time Factors affecting the load
- Holidays (Christmas, New Years)
- Significant reduction of load
- Days proceeding or following the holidays also
have a reduced load. - Pattern change due to the tendency of prolonging
the vacation
28Impute Data and System Parameters for Load
Forecasting
- Weather factors affecting the load
- The weather affects the load because of weather
sensitive loads - air-conditioning
- house heating
- irrigation
29Impute Data and System Parameters for Load
Forecasting
- Weather factors affecting the load
- The most important parameters are
- Forecasted temperature
- Forecasted maximum daily temperature
- Past temperature
- Regional temperature in regions with
- diverse climate
30Impute Data and System Parameters for Load
Forecasting
- Weather Factors Affecting the Load
- The most important parameters are
- Humidity
- Thunderstorms
- Wind speed
- Rain, fog, snow
- Cloud cover or sunshine
31Impute Data and System Parameters for Load
Forecasting
- Random Disturbances Effects on Load
- Start or stop of large loads (steel mill,
factory, furnace) - Widespread strikes
- Sporting events (football games)
- Popular television shows
- Shut-down of industrial facility
32Impute Data and System Parameters for Load
Forecasting
- The different load forecasting techniques use
different sets of data listed before. - Two -three years of data is required for the
validation and development of a new forecasting
program. - The practical use of a forecasting program
requires a moving time window of data
33Impute Data and System Parameters for Load
Forecasting
- The moving time window of data requires
- Data covering the last 3-6 weeks
- Data forecasted for the forecasting period,
generally one week
34Impute Data and System Parameters for Load
Forecasting
- The selection of long periods of historical data
eliminates the seasonal variation - The selection of short periods of historical data
eliminates the processes that are no longer
operative.
35Impute Data and System Parameters for Load
Forecasting
- The forecasting is a continuous process.
- The utility forecasts the load of its service
area. - The forecaster
- prepares a new forecast for everyday and
- updates the existing forecast daily
- The data base is a moving window of data
36Major Load Forecasting Techniques
- Statistical methods
- Artificial Neural Networks
- Fuzzy logic
- Evolutionary programming
- Simulated Annealing and expert system
- Combination of the above methods
37Major Load Forecasting Techniques
- The statistical methods will be discussed briefly
to explain the basic concept of the load
forecasting -
- This lecture concentrates on load forecasting
methods using neural networks and fuzzy logic
38Concept of STLF Model Development
- Model selection
- Calculation and update of model parameters
- Testing the model performance
- Update/modification of the model if the
performance is not satisfactory
39Concept of STLF Model Development
- Model selection
- Selection of mathematical techniques that match
with the local requirements - Calculation and update of model parameters
- This includes the determination of the constants
and - selection of the method to update the constants
values as the circumstance varies. (seasonal
changes)
40Concept of STLF Model Development
- Testing the model performance
- First the model performance has to be validated
using 2-3 years of historical data - The final validation is the use of the model in
real life conditions. The evaluation terms are - accuracy
- ease of use
- bad/anomalous data detection
41Concept of STLF Model Development
- Update/modification of the model if the
performance is not satisfactory - Due to the changing circumstances (regional
gross, decline of local industry etc.) the model
becomes obsolete and inaccurate, - Model performance, accuracy has to be evaluated
continuously - Periodic update of parameters or the change of
model structure is needed
42Artificial Neural Networksfor STLF
43Artificial Neural Networksfor STLF
- Several Artificial Neural Network (ANN) based
load forecasting programs have been developed. - The following neural networks were tested for
load forecasting - Feed-forward type ANN
- Radial based ANN
- Recurrent type ANN
44Building Blocks of a Feed Forward Network
- A Feed forward Three-Layered Perceptron Type ANN
was selected to demonstrate the short term load
forecasting technique. - The selected network forecasts
- Hourly loads
- Peak load of the day
- Total load of the day.
45Building Blocks of a Feed Forward Network
- The forecasting with neural network will be
demonstrated using a feed forward three-layered
network to forecast the peak load of the day. - The network has
- one output (load in kW)
- three input (previous day max. load and
temperature, forecasted max. temperature)_
46Building Blocks of a Feed Forward Network
- The structure of the Feed Forward Three-Layered
Perceptron Type ANN is presented on the next
page. - The network contains
- i 1.. 3 input layer nodes
- j 15 hidden layer nodes
- k 1 output layer nodes
47Artificial Neural Networksfor STLF
48Building Blocks of a Feed Forward Network
- The inputs are
- X1 previous day max. load
- X2 previous day max. temperature
- X3 forecasted max. temperature
- Wij weight factor between input and hidden
layer - wj weight factor between hidden layer and
output
49Building Blocks of a Feed Forward Network
- A sigmoid function is placed in the nodes
(neurons) of the hidden layer and output node. - The sigmoid equation for an arbitrary Z function
is - Y output maximum load
50Building Blocks of a Feed Forward Network
- Inputs Xi are multiplied by the connection
weights (Wij) and passed on to the neurons in the
hidden layer. - The weighted inputs (XiWij) to each neurons are
added together and passed through a sigmoid
function. - Input of hidden layer neuron 1 is
51Building Blocks of a Feed Forward Network
- The output Hj of the jth hidden layer node is
-
52Building Blocks of a Feed Forward Network
- Inputs Hj are multiplied by the connection
weights (wk) and passed on to the neurons in the
output layer. - The weighted inputs (Hjwk) to each output
neurons are added together and passed through a
sigmoid function. - Input of output neuron is
53Building Blocks of a Feed Forward Network
54Training of the Feed Forward Neural Network
- The described neural network is trained using
historical data. - Typical data set contains 2-3 years of load and
weather data. - Error back propagation (BP) method is used for
the training. - During the learning the weights are adjusted
repeatedly.
55Training of the Feed Forward Neural Network
- The output produced by the ANN in response to
inputs are repeatedly compared with the correct
answers - Each time the weights are adjusted slightly by
beck-propagating the error at the output layer
through the ANN - Equations for the training are presented in the
next page
56Training of the Feed Forward Neural Network
- The equations used for the training are
- Weight is between input and hidden layer
- Weight is between hidden layer and output
57Training of the Feed Forward Neural Network
- In the equations
- Yactual is the true value of the output load
- e is learning factor (0.3-0.8)
- n is the number of learning cycles
- Xj is the input value belongs to Yactual
- A numerical example demonstrates the use of
neural forecasting method.
58Training of the Feed Forward Neural Network
- The over training has to be avoided using the
cross validation method - The training set is divided into two parts.
(Part 1 two years data, Part 2 one year data) - Part 1 is used to train the network, by passing
the data through the network. - Few hundred times pass represents a training
period.
59Training of the Feed Forward Neural Network
- The over training of the network has to be
avoided - Part 2 is used to check the effectiveness of the
training. - After each training period the error is
calculated when the network is supplied by the
input data of Part 2. - The increase of error indicates over training,
when the training has to be stopped
60Load Forecasting Engine
- The EPRI developed a Load Forecasting Engine
using 24 Neural networks. - One network forecasts the load for each hour of
the day. - The networks are grouped into four (4) categories
depending on time of the day. - The categories have different inputs.
61Load Forecasting Engine
The construction of the engine shows the four
groups of neural networks.
62Load Forecasting Engine
- The four categories are
- Category 1. Nine neural networks. Forecasts the
load between 1-9AM . - Category 2. Nine neural networks. Forecasts the
load between 10AM -2PM and 7 -10PM . - Category 3. Four neural networks. Forecasts the
load between 3-6 PM . - Category 4. Two neural networks. Forecasts the
load between 11-12 PM .
63Load Forecasting Engine
- The inputs in four categories are
- Category 1. Forecast for early morning
- general input
- load in the last three-four hours
- temperature in the last three-four hours
- Category 2. Forecast for off peak hours
- general input
- forecast temperature of previous hours
- yesterdays load and temperature of hours close
to this hours
64Load Forecasting Engine
- The inputs in four categories are
- Category 3. Forecast for afternoon peak hours
- general input
- forecast temperatures of previous and feature
hours close to this hours. - yesterdays load and temperature of hours close
to this hours - Category 4. Forecast for late night hours
- general input
- forecast temperatures for the four proceeding
hours
65Load Forecasting Engine
- The general input variables are
- same hour load, temperature and humidity of one
day ago. (3) - same hour load, temperature and humidity of two
days ago. (3) - same hour load and temperature seven (7) days ago
(2)
66Load Forecasting Engine
- The general input variables are (continuation)
- same hour forecast temperature and relative
humidity of the next day (2) - day of the week index (Sunday 01, Monday 02 etc.)
67Load Forecasting Engine
- The load forecasting engine has one output for
the hourly load - The extended forecast uses the forecasted values.
E.g. The two-day ahead forecast uses values
obtained by the one-day ahead forecast. - The forecast can be updated each hour using the
recent load and weather data
68Load Forecasting Engine
- The weights in the neural network are adjusted
daily. - The retraining uses the actual load and weather
data of the past few days. - The retraining helps to follow the trends,
changes of weather patter e.t.c
69Load Forecasting For Holidays
- The load during the holidays has different
patterns and is significantly reduced. - The forecast is inaccurate because of the small
number of historical data. - The holiday is treated as
- Saturday if the shopping centers are open
- Sunday if the shopping centers are not open
70Hourly Weather Forecast
- The weather service provides forecasts for
- daily maximum and minimum temperature
- daily maximum and minimum relative humidity
- rain and fog
- maximum wind speed
- No hourly data are provided.
71Hourly Weather Forecast
- EPRI developed an hourly temperature and humidity
forecasting engine. - Single neural network with
- 28 inputs
- hourly temperature of the previous day)
- high and low temperature of the two previous day
- 24 outputs expected hourly temperatures.
72Load Forecasting Results
73Load Forecasting Results
Comparison of forecasted and actual loads
74Load Forecasting Results
- Accuracy less than 3 for the next days
forecast is considered good - The longer term forecast accuracy is less (7-8)
75Appendix 1
- Derivation of Learning Algorithm
76Derivation of Learning Algorithm
- The output of the hidden and output layer and the
error function are
77Derivation of Learning Algorithm
- The update of the weight factors require
iteration - For the calculation of the derivative the
following substitutions are used
78Derivation of Learning Algorithm
- After substitutions the equations are
79Derivation of Learning Algorithm
- The derivation of the error function results in
- The derivative of the u function is
80Derivation of Learning Algorithm
- The derivative of the output function is
81Derivation of Learning Algorithm
- The derivation of the auxiliary function Y2
gives - The derivative of Y1 function is
82Derivation of Learning Algorithm
- Substituting the results in the equations
83Derivation of Learning Algorithm
- The rearrangement of the output equation results
in - The final equation for the update of dwj is
84Derivation of Learning Algorithm
- The derivative of the hidden layer function is
85Derivation of Learning Algorithm
- The derivation of the auxiliary function h2
gives - The derivative of h1 function is
86Derivation of Learning Algorithm
- Substituting the results in the equations
87Derivation of Learning Algorithm
- The rearrangement of the output equation results
in - The final equation for the update of dWij is
88Derivation of Learning Algorithm
- Substituting the results in the equations which
is used to iterate the wj value
89Derivation of Learning Algorithm
- The two training algorithms are