GO6 - Load Forecast

 

Grid Operation

Short Term Load Forecast

Better Start With PART-1

Preface:

There are some industries that produce intermediates only. Their product is not for consumers but it is input to other industries producing final product for consumers. In the chemical sector, some industry  produce intermediates that is input for other industries like pharmaceutical, colors, detergents, cosmetics, etc. Similarly various industries produce different machine parts that are used in industries producing machines for consumers.

The objective of this write up is similar. It was observed that many academic and field experts have developed efficient models for Short Term Load Forecasting using different logics and techniques. But when applied to field data, in many cases, the results were not as accurate and consistence as expected. Reason may be discrepancy in data itself and result was GIGO.

Here I have some useful information in the matter as a practicing engineer having long experience of 26 years working at power system operation and control at SLDC. It is all about right load data, likely errors, probable cause, data filtration, data relations and effect of environment factors on system load. Hope the information will be useful to intending programmers.

1.             Load Territory: Generally load assessment and forecast is required for particular area. The area may be Industrial Zone, City, District, Area of Substation or Licensee or Distribution Company, State Grid, Regional Grid or National Grid. Forecasting requires historical data because it provides characteristic of the load variations.

2.        Load Assessment: The load of any area is not the summation of rating of connected load, contract load or operating load by all users in that area. It may be the actual power drawn by all users along with power wasted in the area. Practically it is rather than difficult, impossible to get correct data. In practice this is assessed indirectly. Total power consumed is total power delivered in that area. We use term load but actually we assess power required in the area and is referred to as power demand. This is what required for operation planning. Power delivered in the area is power injected by power generators within the area plus power received in the area from the other systems through inters connecting lines.

2.1     Injection by Power plants: This includes all power plants within area irrespective of its ownership. Power injection is power generated less its own auxiliary consumption. Sometimes gross generation is considered as power injection. But it is not correct because auxiliary consumption is associated with generating unit. Whenever generator is out its auxiliary consumption is also absent. Generator producing 500 MW has about 45 MW of auxiliary consumption. So its injection will be 500-45=455 MW. If this generator is taken out, the increase in power import will be 500-45=455 MW only. Therefore auxiliary consumption of generator cannot be considered a part of system load.

2.2    Injection by Tie line: All the power systems are interconnected operating in the grid mode. Power flows from/to other systems through interconnecting lines depending upon systems parameters. Metering is available at both ends of the interconnecting lines. Power received/sent (imported/exported) at the point within the area is considered as positive/negative injection for the load calculations.

2.3    Average MW: Normally load is expressed in MW and is measured by indicating meters. The meter shows instantaneous value of power flow. MW meter reading continuously changing. Noting and summing up of these readings my lead to inaccuracy due to time diversity in meter reading. There may be other issues also depending upon whether meter is analogue or digital. In case of analogue meter zero setting and meter reading errors are expected. But average MW over the time slot represents accurate average load. Average MW is derived from energy meter readings. Difference of MWH meter reading over the time slot divided by time interval in hours will be the average MW for that time slot. MWH reading difference for 15 minutes time slot is to be divided by ¼ i.e. multiply by four is average MW for that slot.

MWH Logging errors: Some errors are likely while operator is reading the meters and recording. So demand derived from such data is not accurate. So also forecast based on such data cannot be accurate. The errors may be due to following reasons.

2.3.1 Diversity in meter reading: There may be many control panels at station and each panel may have many meters for various parameters. Operators note down readings of all the meters at prescribed intervals.  Operator read and logs all meters one by one and moving panel to panel. It takes some time in completing the round. Assume that operator is following the same sequence in each reading cycle. In that case reading interval may be equal but data is time staggered. Load derived from such reading has no significant error during the steady load period. But it leads to significant error during the period of load rise and fall. Calculated load is more than actual during rising load period and it is less than actual during dropping load period. This error can be minimize by meter readings in two rounds. The first round may be only for selected MWH meters reading followed by second round as usual for all other readings.

2.3.2 Out of schedule reading: Normally readings are noted hourly. Operators have tendency to start data logging early in last hour of the shift so as to complete his work before leaving his duty. Obviously the effect is lower average MW in this hour and higher average MW in next hour. This type of error can be eliminated by shift change at mid of the hours.

2.3.3 On line Data: Most of the power systems have SCADA system. Normally digital data like breaker and isolator indications, protection system indications and analogue data of frequency, voltage, active power, reactive power and current flows are scanned. Availability of energy counters in SCADA system can be useful for the purpose. All these data are updated periodically and snapshot of required data can be best alternative to derive the load. Only validity has to be ascertained for accuracy and regular updating.

3.       Demands: Load forecast is advance estimation of future demand. Demand is power injected in the system derived as discussed here above. Demand is of four types as under.

3.1     Catered Demand: The sum total of power injection in the system by various sources is catered demand. The data of catered demand is mutilated due to variation in frequency and load shedding. Therefore catered demand data have no consistency and do not represent real system demand. It only represents how the power is delivered in the system over the day.

3.2     Computed Demand: Power drawn by frequency dependant loads in the system varies with the change in frequency. System frequency depends on power supply demand condition in the grid. System frequency may not be normal all the time. Frequency variation do not follow any regular pattern and hence unpredictable. The catered demand derived as above has deviations from true demand due to frequency dependant loads.

            So catered demand has to be corrected for each time block using formula as under

Computed Demand = Catered Demand + b×dF

Where b is system Bias and dF is frequency deviation for corresponding time block

System bias may be 3 to 4 percent per Hz of catered demand. (Bias depends on mix of load types)

Frequency deviation = (Nominal frequency – Average Frequency during time slot)

Computed Demand represents all load data at nominal frequency. This frequency correction can be negative also when block frequency is above normal.

3.3    Restricted Demand: Computed demand data corresponds to nominal frequency. But for system security and adhere to operating discipline, occasionally real time load shedding has to be imposed according to requirement. Load not so catered reflect as load drop and not represent the true demand. Therefore correction is required for concern time blocks as under.

Restricted Demand = Computed Demand + Load not supplied.

This data of restricted demand represents routine load pattern of the system at nominal frequency. Therefore data of restricted demand is used for Short Term Load Forecasting. Forecasted demand is useful for operation planning. Unit Commitment, Generation Scheduling (with merit order), ISGS Power Indent and Power Purchase from other sources is decided based on forecasted demand and estimated on line availability.

3.4    Unrestricted Demand: This represents power requirement in the system when there is no any type of constraint on use of power by any category of consumers. It is always the aim of supply authorities to cater unrestricted demand but anyhow not feasible due to various reasons. So to optimize utilization of available resources in best way, some statutory restrictions are enforced on various categories of consumers. Unrestricted demand is calculated from restricted demand by modifying respective time slot as under.

Unrestricted Demand = Restricted Demand + Load Relief due to statutory restriction in the time block.

This unrestricted demand is useful for long term forecast for power system expansion planning etc.

Statutory restrictions may be similar to hereunder.

3.4.1  Load Stagger: The objective of this is to flatten the load curve so that maximum load can be catered with available (OLC) on line capacity. Period of power requirement of various types of loads is different and operating hours are also different as per user's convenience. So unrestricted demand curve is in not flat but load varies over the time of the day. Load factor of such unrestricted demand may be in the range of 70 to 80 percent depending upon load mix. The load requiring only few hours power in a day can be shifted from high demand period to low demand period. Load for this purpose is selected such that it causes least inconvenience to minimum users.

3.4.2 Holiday stagger: Industrial power users have to observe weekly holiday as per statutory requirement. On holiday the power demand drops considerably and there is unused (OLC) requiring backing down of generation. Some time it may require to stop generator. By staggered holiday this weekly drop can be distributed on all the seven days of week. So load on all weeks is almost equal but reduce by 100/7 = 14% of weekly drop. All the industries observing weekly holiday are divided in seven groups having almost equal loads drop. Each group has to observe weekly holiday on different day fixed for the each group. Fixing different holiday may be based on some base like territory, category, etc. So exact load balancing in all groups is not feasible. So load drop on each day may not be exactly 14% but can be around this. This is soft restriction because industries continue to work six days in week without any production loss.

Holiday staggering

3.4.3 Recess stagger: Industries working round the clock has three shifts of eight hours starting from mid night. They have to observe recess in the mid of the shift as per statutory requirement. Recess in after noon shift may be during evening peak hours. Evening peak load is highest of the day and is critical period for system management. All the resources are on during this period to meet the demand without drop in frequency. But load drop due to industrial recess during this period causes sharp drop in demand and frequency increase for short time. After the recess, load is resumed and again demand increase and frequency drops. Similarly system demand is minimum of the day between 03 to 05 hrs in early morning with high frequency. Load drop due to industrial recess of night shift during this period causes frequency to rise very high. This phenomenon can be mitigated by staggering of the industrial recess. High load drop for short period can be made small drop for long period by staggering recess. Such industries are divided in three or four groups and advised for staggered recess timings. By this arrangement power requirement during peak hours has some relief as shown in the figure and abrupt frequency change is avoided. This is also a soft restriction.

Recess staggering

3.4.4 Agricultural load stagger: Power required for agricultural pump sets is for some hours in a day. Numbers of persons engaged per MW is very less compared to other category. Users have choice for operating hours but can be operated at other time also. Fulfilling power requirement in low demand period instead of peak demand period is very helpful for managing peak demand. Designing of power supply schedule for this purpose is very tricky because all effected users have fair treatment. Care is taken for power supply time of the day, numbers of hours, users local requirements and yet get load relief as per the system requirement. 

3.4.5 Rural load Shed: Electricity is required any time for one or other purpose. But power supply is dicontinue for some period during the time when is it is least inconvinience to minimum users. This is required during acute power shortage. Load relief this way is usefulto to supply power to other catagory of users.

3.4.6 Urban load Shed: This is similar to above but has high impect. Government and corporate offices, Malls, multi storey building, big markets, heavy traffic are effected by load shedding. Probability of problem for low and order also. So as far possible it is restricted up to town level. Load shedding in big cities and metros are avoided. 

3.4.7 Peak Restriction: Many small and medium industrial units etc are working in general shift. Restricting use of power during evening peak hours is helpful in managing peak demand. Such users can manage their working hours accordingly.

3.4.8 Half Load Lighting: Alternate pole street light is switch off. This requires such arrangment. So when required half of the lifgts are switched off. Such arrangement is also done for flood lighting, railway plateform lighting, etc. 

4         Forecasting period:

4.1   Long and Medium Term: This is useful for power system development planning, monitoring, reviewing, rescheduling of execution, material management etc. Unrestricted demand is useful for the purpose. The forecast is not fully analytical using any extrapolative methods because it is also influence by factors other than past data such as Government policy, economic situation, industrial development, development of ancillary facilities, etc.

4.2   Short Term Forecast: This forecast is mainly useful for power system operation planning. Restricted demand is useful for the purpose. It can be analytical using any extrapolative methods. It is not influence by other factors like long term forecast. But environmental parameters have impact as discussed here after. Forecasting period is less than a year.

4.2.1 Yearly: Yearly load forecast is useful for planning of network and generators statutory/schedule outages for boiler inspection, annual maintenance, capital overhaul, modification, up gradation etc. Forecast of only Peak/Off peak demand and energy requirement is sufficient for the purpose.

4.2.2   Weekly: Weekly forecast is useful to plan outages of network and generator for urgent work. It is also useful to review/modify restriction schedule in view of any change in load pattern on account of seasonal change.

4.2.3  Next day: This is most useful and used load forecast. Mostly referred as Short Term Load Forecast (STLF). It is used for operation planning activities like unit commitment, generation scheduling, ISGS power indents, power purchase/assistance from other sources. It is also useful for planning short term control measures while acute shortage of estimated OLC.

4.2.4  Run time or Corrective: This forecast is for the current day. Forecast and actual load is monitored in real time to check for any deviation. Consistent deviation indicates change in load due to unexpected abnormal condition like shutdown of major industries on account of strike, breakdown, fault, failure of captive plant of major industry, etc. This forecast is for remaining period of the day. This is useful to take corrective action for rescheduling power input in the system.

5        Forecasting Day: Hourly load forecasting is most common. Load is forecasted for next day on hourly bases from 00 hrs to 24 hrs. But this is not proper. Loading pattern of most of consumers is almost similar on all days of the week. But loading pattern of industrial, commercial and office is different on holidays. Industries working round the clock start first shift at about 07 or 08 hrs in the morning. Similarly some other load like general shift working industrial units, Shops, malls, offices, banks and many others starts after  07 hrs. Means load variations due to holiday starts from 07 hrs and continue till next day morning 07 hrs. Therefore load forecast day should be from 07 hrs to 07 hrs of next day.

6       Forecasting Slots: Very simple load forecasting was manual and was block wise like peak load period, moderate load period, minimum load period etc. Normally system load is calculated hourly. So load forecasting is also hourly. But load variation is more evident when time interval is smaller. Now after implementation of Availability Based Tariff (ABT), most of the data for power transfer, generated etc are available for 15 minutes blocks. So it is feasible to have system demand data in 96 blocks of 15 minutes and hence load forecasting is feasible same way.

Therefore ideal Short Term Load Forecasting should be in 96 blocks of 15 minutes from 07 hrs to 07 hrs next day. 

7.      Load Characteristics: System load is composed of various types of loads. Quantum and characteristic of each type of load is different. Basically loads can be categories according to operational characteristic are as under.

1.       Load throughout the day.

Three shift Industries, Cold storage, Refrigerator, etc.

2.      Loads related to time of the day.

Banks, Offices, TV, Shops, Cinema, etc

3.      Loads related to day/night period i.e. sunrise/sunset.

All Lightings, Street Light, Display, etc.

4.      Load for season only.

Ice Factory, Oil Mills, Cotton Processing, Agricultural, etc.  

5.      Load related to weather/climatic conditions.

Fan, AC, Heaters, etc.

So power demand varies over the day and year. Power demand is high when more loads incidents on the system. During evening hours all types of the loads are ON except some office loads. So it is the period of highest demand of the day known as peak hours. Some loads are fixed with the time and some loads shift by time with the season. Ultimately in particular season, these loads co-incident resulting in higher but shorter peak. In other season these loads have some diversity resulting in comparatively lower but longer peak. This seasonal change occurs due to variation of different loads and its time shift. This change occures smoothly day by day and hence not visible in day to day observation. But fact is revealed if load is compared for extreme days of the seasons.

8             Seasonal Forecast: Daily load curve for winter, summer and monsoon are different. Some time season wise forecast package is recommended. But it is not logical. System load is result of all types of the loads as above having different operating characteristics influence by various factors. Various types of loads may increase or decrease, operating duration may vary and also shift in operating time. Ultimately load curve is changing. But all these changes occur gradually day by day. During long day of summer, various loads may be scattered resulting in longer but lower peak in morning and evening.  But during short day of winter, some of the loads overlap resulting in shorter but higher peaks. Therefore separate module is not required but same module should take care of gradual changes on account of seasonal changes.

9.       Modern forecasting module use AI. When data relation for future trend is unknown, fuzzy logic may be the solution. But if data relations are known, it would be better to link it for forecasting process. The results will be more accurate due to logical approach. In absence of any known link, program will try to establish links with all the data and sometimes have erratic result. Here are some hints for relationship.  

9.1    Day pattern: Load data of any hour has relation with previous block and following block. Strongest attachment is with adjoining block (previous block and next block) and weaker attachment with block before previous block and block after the next block and so on. H hours load has strong attachment with H-1 and H+1 hours load but weak attachment with H-2 and H+2 hours load. Similarly H-3 and H+3 hours load has weaker attachment and so on. This means day load curve is almost maintained with minor changes. There can not be abrupt change only in any one block. 

9.2      Forward Trend: Load trend is derived by comparing load of corresponding blocks of previous days. The difference reveals rising or dropping trend of the load. It also indicates the type of trend whether it is uniform, growth, exponential etc. This trend is useful for forecasting. 

9.3    Weekly trend: System load data with staggered holiday requires different treatment. All consecutive days load have slight positive or negative variation as seen earlier in holiday staggering. Therefore the consecutive days load data is not suitable to derive load trend as above. In such cases load trend is derived comparing same block load with corresponding day of previous week. These differences are expected to be large due to weekly difference. Differences of all week days with respect to corresponding day of previous week is useful to ascertain load trend for future.

Power system facing acute power shortage has to implement regular load control measures. Power supply to various categories and/or cluster like agricultural, villages, towns, talukas, districts, industrial zones or area of city is selected to switch off power as per schedule. The schedule may be for few hours every day or more hours once in a week. Some category of consumers observe different holidays on their own like saloons and parlors on Saturday, shops on Monday, offices on Sunday etc. In such case daily load curve may not be similar. So load forecast should follow logic of weekly trend as above in staggered holidays.

Load variation on festival holidays is different for each festival. On HOLI the load is normal till evening and drop start thereafter but DHULETI is full light load day. JANMASTHAMI has no load drop but next day is light load day. In Gujarat DIWALI days load drop starts in the evening and minimum load on next day. Load growth thereafter is day by day and normal load reaches after 5 days. Hence load forecast for such festival holidays can be derived using data of similar holidays in previous years.

Generally understood that larger the past data, better is tuning/training of program and forecast will be perfect. This may be true for other matters but not right for load forecast.  As discussed above, the load pattern is continuously changing and very old data have no significance and may lead to errors. Only three to four weeks immediate past data should be used for forecast.

10    Data Filtration: Past load data is used as base for forecasting. Forecast may not be perfect if input data is not proper. Manual data can have errors of reading, logging, transferring, calculating etc. Online data can have errors due to calibration of sensors or its supply failure, noise on channels or transmission failure, etc. Perhaps errors could have been rectified at the data sources. However data has to be filtered / rectified before processing. Logic at 9 above can be used for the purpose.

11       Accuracy Check: Hundred percent perfect forecasting is not possible. Small deviation is expected and is allowable. But wide deviation in forecast and actual load indicate inappropriate data, filtration or forecast. Therefore validity check is important before implementation in the field. Forecasted load data is compared with actual load data and deviation is found for each block. Validity check should be performed for numbers of days to ascertain consistency of results. Various ways for validity check are as under.

11.1    Simple Average: This simple average of deviation in all blocks of the day. Positive and negative deviations cancelled in this method. So it is not a good check method.

11.2   Maximum Deviation: Small deviation is expected in any forecast and is immaterial in operation planning. But wide deviation will mislead for operation planning. So the maximum deviation is considered as validity check. Some time maximum positive and maximum negative deviation is also checked. But in this case other wide variations remain unnoticed under the cap of maximum. So it may not project the true accuracy of the program. This method is not perfect but better than above.

11.3   Unsigned Average: Average deviation is derived from unsigned deviations of blocks of the day. Here positive and negative deviations do not cancel. This is moderate method for accuracy check.

11.4   RMS Deviations: RMS is root mean square of deviation in all blocks of the day. This is presented as percentage of average actual load of the day. This is the best accuracy indicator of forecast.

           For training / tuning and accuracy check huge data base is used. For example, load for Thursday is forecasted using actual load data up to Wednesday as input. Than forecast of Thursday is compared with actual load of Thursday and error is derived. Next load for Friday is forecasted using actual load data up to Thursday as input. Than forecast of Friday is compared with actual load of Friday and error is derived. In this way past load data base is build up adding one day data each time. Here gap is of one day between latest actual load data available and day of load forecast.

But in field implementation it differs because current day actual data is not available. Forecast for tomorrow is done using actual load data of yesterday. For example, on Wednesday load is to be forecasted for Thursday. But actual load data available is up to Tuesday only. So there is gap of two days between latest actual load data available and load to be forecasted. Therefore this aspect is to be considered in load forecast program for field implementation.      

12.    Environmental Effect: Forecast procedure discussed above is without consideration of impact of environmental variations. Following three parameters have influence on the power system demand.

1.       Temperature

2.      Wind Speed

3.      Rainfall 

12.1   Ambient Temperature: There is no appreciable variation in system demand during normal temperature range. But system demand will increase when temperature is higher or lower than the normal range. Ambient temperature influence system load in two ways

12.1.1Cooling load: The cooling load is added when ambient temperature rises above normal range. It is due to operation of Fans, Air Coolers and Air Conditioners. This load varies with variation in temperature. There is no cooling load up to certain temperature. Thereafter cooling load increase slowly with rise in temperature up to a limit because very few consumers use cooling devices in this temperature zone. But cooling load increase faster when temperature rise beyond the limit when more and more consumers use cooling devices. Again cooling load growth is slow after certain high temperature because all the installed and available cooling devices are in service at this stage. Beyond this stage load increase is slow with further rise in temperature. This is because no more gadgets are added but is due to long operation of compressor in ACs, fans at high speed and reduced diversity of operation of cooling devices. 

Cooling Load Variation

12.1.2Heating Load: This load is added when ambient temperature drops below normal range. It is due to operation of heaters. This load varies with variation in temperature. There is no heaters load up to certain drop in temperature. Thereafter heaters load increase slowly with drop in temperature because only few consumers use heaters in this temperature range. But heaters load increase faster with further drop in temperature beyond the limit when more and more consumers use heating devices. Thereafter no appreciable rise in heater loads because all the available heating devices may be in service. However meager load increase is observed with drop in temperature due to heaters operation continue for more hours.

Heating Load Variation

 

This load is area specific. Cold zone has heaters load only whereas hot zone has cooling load only but other areas may have load variation due heaters in cold days and due to coolers in hot days.

Implementation: Normal load forecast includes cooling and heating loads. Because past data used for forecasting is inclusive of respective load as per season and area. Therefore no correction is required in forecasted load for routine temperature change. However correction can be applied to forecast when temperature variation is abnormal. Exact temperature load relation can be established from past incidences and applied. The correction can be positive or negative depending upon the temperature drift. This method is applicable for cooling as well heating load.

12.2  Wind speed: Wind affects the above loads but effect is different. During hot days, operation of fans, coolers and ACs reduces during windy atmosphere as people opt to enjoy natural breeze avoiding use of artificial cooling device and also to save energy. But during cold days, operation of heaters increases during windy atmosphere.

12.3   Rain: Rain fall causes drop in demand. It is due to two reasons.

    1.   Fault on distribution feeders. System demand drops due to loss of load on account of tripping of distribution feeders due to fault on line. Load drop is large during first rain of the season. This load drop continue till fault on feeders are rectified and resumed. 

    2.    Non operation of agricultural pump sets. Lift irrigation by pump set is not required after natural irrigation by rain. Therefore heavy load drop in the area of rain fall as pump sets are not operated. Rain is local event and hence its effect on load drop is also local. Therefore rain data from numbers of strategic points all over the area is required. Pumping load may re-incident in case raining pattern is not as per irrigation cycle required according to the crop in concern area.

Agriculture pump load pattern is related to type of the crop in the area. Each crop irrigation requirement is peculiar and user convinience also differs. Farmers prefer to run pump sets in the evening hours during hot summer days for their convenience and also to have less water evaporation. But prefer to run pump sets during noon time in the winter. However farmers in the fields near the forest, prefer to work during day time irrespective of the season because of fear from wild animals during night hours. Farmers in the area like Saurashtra where water is deep and inflow is slow, cannot operate pump sets at a stretch. They have to operate pumps with intervals. Proportion of agricultural load and other load is different in various areas. Load drop depends on amount and repeat frequency of rain in respective areas. Time, duration and amount of drop is also different in various areas.

Load drop due to rain may radically change power flow pattern in the system. Network loading needs close monitoring during the season. So load forecast is very complex due to erratic behavior of users, erratic rain and erratic load drop. Therefore very accurate load forecast for monsoon is not feasible.