Consumer appliance‐level load shedding optimisation for real‐time application

Load shedding (LS) is implemented by distribution utilities (DUs) in addressing power supply insufficiency problems to avoid DU system damages. This is commonly implemented by the installation of LS relays in every DU feeder. However, in either scheduled or unscheduled supply disruptions, a huge amount of unnecessary de-loading is taking place in a feeder level LS implementation. In addition, the consumers connected to a de-loaded feeder are in total blackout, that is, consumers have no choice over which appliances to spare from being de-loaded. This study proposes an LS implementation that replaces feeder-level de-loading by a finer consumer-appliance-level de-loading and allows consumers to have some control over their de-loading. In this method, consumers can set an appliance priority level to their selected connected loads at a given time, to avoid a total blackout. Furthermore, to deal with the enormous data involved in this proposed method, both centralised and distributed optimisation approaches are employed to expedite the system processing response. Simulations are conducted to verify the proposed method's functionality. Lastly, economic analysis is done to assess the proposed method's viability.


Introduction
The electrical power supply and demand balance is essential in power systems. Power plant operators (PPOs), transmission system operators (TSOs), and distribution utilities (DUs) are in close coordination to maintain this balance. When the supply shortage is not large, moderate load shedding (LS) measures such as conservative voltage reduction are enough [1]. However, when the supply shortage is large, hard LS measures are necessary that cut off the power supply by LS relays in every DU feeder. Such is the case in Mindanao, the southern island of the Philippines. The recurring drought in the region impacts hydro power plants which cover 30% of the power supply [2]. DUs in this case, has to resort to the hard LS, which is simply referred to as LS hereafter, to avoid system damages.
LS is a more economical and realistic solution to the supplydemand imbalance than installing more generating capacities, which involves huge monetary investments and tedious negotiation processes between stakeholders. However, LS can be an expensive strategy [3,4]. Excessive de-loading often takes place in feederlevel LS. For example, even when the necessary de-loading is only half of the load connected to the feeder, the whole connected load is de-loaded in the current LS implementation.
To address the mentioned issues, immediate solutions were presented by several published papers that focused on the development of fast LS [5], operation optimisation of underfrequency load shedding (UFLS) [6], and under-voltage load shedding relays to reduce and optimise LS occurrences [7,8]. Shekari et al. [9] proposed an adaptive wide-area centralised LS by determining the amount of necessary demand reductions, locations of load drops, and the LS event types are optimally simulated. Potel et al. [10] proposed a clustering-based method to improve the selection of feeders in a UFLS implementation. Abdelwahid et al. [11] verified the effectiveness of UFLS scheme implemented in real-time through hardware implementation and simulations. Babaei et al. [12] developed an algorithm to handle LS implementations especially during natural disasters taking the uncertainty of contingencies into consideration. Also, Yaun and Xu [13] proposed a method for preventive-coordinated LS considering power supply uncertainties. Cavalcante et al. [14] and Shen et al. [15] introduced a self-healing scheme that minimises the unsupplied demand while maintaining the faulted network isolated.
Furthermore, the study of Hoseinzadeh et al. [16] addresses the decentralised LS by using the instantaneous voltage deviation of load buses as a parameter in determining the frequency thresholds of UFLS relays. These papers are feeder-level LS optimisation which can be made finer into a consumer appliance-level LS to maximise the available power utilisation and avoid too much LS. The study of Bhattacherjee et al. [17] explores the possibility of residential-based LS with the use of smart-meters, single-board computers, and a monitoring server to implement an efficient LS. Sigrist [18] and Yao et al. [19] proposed a residential-level LS to minimise the load losses by detecting frequency mismatch which also considers the rate of change of frequency. Moreover, appliance-level LS provides an opportunity for each consumer to experience a partial blackout, wherein not all of their electrical devices are turned-off, unlike the current LS implementation. Xiang et al. [20] developed an appliance level LS with the use of Internet of Things (IoT) technology to control individual devices through cloud servers. Azasoo et al. [21] and Jabian et al. [22] proposed an appliance-level LS that aims to minimise excessive and unnecessary de-loading that considers consumer's comfort. Azasoo et al. [21] developed a heuristic approach in an appliancelevel LS but the power demand is pre-determined in the simulation, thus, variations of power demand and power supply in real-time and fairness of implementation among consumers are not evaluated. On the other hand, Jabian et al. [22] provide consumers with a way to participate in every de-loading implementation by considering their respective appliance priority levels. Also, they introduced a fairness function to make the LS fair among consumers. However, power allocation was determined based on the forecasted load without paying attention to risks possibly incurred by the unavoidable existence of forecasting errors due to uncertainties in power systems [23]. Also, the optimisation results showed that genetic algorithm (GA) is effective in this kind of problem but they did not investigate other optimisation techniques to further elaborate its superiority. Moreover, performance evaluations regarding the employed optimisation technique and the introduced fairness function were limited only to specific sets of appliances. Furthermore, economic analysis to assess the economic impact of the proposed method was not part of the study.
In this paper, the researchers propose a method that improves the appliance-level LS implementation of Jabian et al. [22] by incorporating a measure to cope with forecasting errors based on J. Eng error analysis that improves system stability. Also, the researchers have selected binary particle swarm optimisation (BPSO), a widely used optimisation technique in power system, as a comparison to GA in order to confirm the technical viability of the latter by providing an extensive assessment in terms of optimisation quality and execution time in dealing with this kind of optimisation problem. Furthermore, a comprehensive evaluation of effectiveness in the introduced fairness function among consumer on different sets of appliances' power ratings is demonstrated. Lastly, economic analysis is performed to ascertain its economic impact from the viewpoints of both the DU side and consumer side.
The rest of this paper is organised as follows. In Section 2, the proposed LS system is described including the overall two-level system architecture, determination of load capacity to be shed, efficient and fair supply capacity allocations with consumers' priority levels considered, and economic analysis,. Results on fairness function evaluation, optimisation method comparison, economic analysis, and case studies are presented in Section 3. Finally, conclusions are given in Section 4.

Proposed system design
In this section, the system topology of communication and decision-making system is first described followed by the consumer participation scheme. Then, the improved LS procedure is described. Finally, the criteria used in the economic analysis will be discussed.

System network topology
Appliance-level LS optimisation has to deal with a large number of appliances (i.e. >65,000 variables). This causes large burdens on a centralised communication and decision making system, which is not practical for low-cost implementation and real-time application. Thus, the authors decide to utilise the combined centralised and distributed topology as depicted in Fig. 1. Appliances owned by consumers are monitored and controlled by control nodes X jk , which are under the control of an aggregated appliances controller. The central station (CS) is in DU's server and in charge of power management, monitoring, data logging, and instructing appliances controllers K j [22]. Communication between these components is assumed to have a bi-directional protocol.

Consumer appliance priority level setting
In the existing LS implementation, consumers have no part in the decision making of which areas to be de-loaded, more so, of which appliance to be de-loaded. In the proposed scheme, consumers assign priority levels p to their appliances based on their comfort and prerogative. The LS scheme decides which appliances should be off so that excessive LS is minimised while respecting the priority levels, avoiding the total blackout and reducing consumers' inconvenience. Appliances of the same priority level are collectively controlled by a node X jk and therefore they are treated as if a single controllable appliance. The appliance rated power is denoted as L jk 1 , …, L jk 5 , with L jk 5 as the appliance rating with the lowest priority level [22].

Procedure
The ultimate goal of the proposed procedure is to determine in the best manner whether each of the appliances nominated by consumers is to be switched on or off. From the DU's point of view, the best way is to utilise the supply capacity maximally, or equivalently to minimise loss of revenue chances. As shown in the following subsections, this is addressed by minimisation of unallocated supply capacity. From the consumers' viewpoints, the best way is to respect their own appliances' priorities, to avoid repeated on and off over a short period of time, and to be fair among consumers. To avoid repeated on and off caused by shortsighted prospect into the future, the appropriate load forecast is performed and analysed. Consumers-specified priority levels are taken into consideration when the minimisation of unallocated supply capacity is done. Among the obtained solutions, the fairest one is chosen.

2.3.1
Step 1: system voltage and frequency monitoring: DUs CS monitors the overall system voltage and frequency to ensure that these parameters are always in their respective allowable ranges. For example, according to the Philippine Distribution Code (PDC), the system is under-voltage if the system voltage is ⩽90% of the nominal voltage, and it is over-voltage if it is ⩾110%. Also, the DUs' system frequency should be between 59.7 and 60.3 Hz [24]. Hence, when CS detects that the system voltage or frequency is below its allowable lower limit for >1 min, then LS implementation is necessary.

2.3.2
Step 2: short-term load forecasting: An LS implementation is necessary when under voltage or frequency is detected. If the power reduction percentage P R , i.e. the percentage of load to be shed, is determined based on the current supply shortage, it could be too small and could cause another LS implementation within a short period because the demand may continue to grow. Thus, a 15-min ahead load forecast is done to determine P R that accommodates the demand variations within that period. This reduces the possibility of repeating an LS implementation and avoids possible flickering effect of appliances' switching. The following equation is used to forecast the load [22]: where L forecast is the forecasted demand after 15 min in MegaWatts (MW), L actual is the actual total connected load power in MW, G t avg is the average load power increase, t denotes the data logging time interval, L t d − l is the demand power at time t on day d -l in MW, d is the current day, and D is the number of available historical data.
Also, forecasting error E forecast is defined as shown in the following equation:

Step 3: power allocation for each load controllers:
The CS instructs each controller K j to reduce their load L j by the ratio of P R so that the sum of load in the forecasting horizon will be equal to the total available but presumably insufficient power supply S G provided by the TSOs. In other words, the power reduction ratio P R is defined as the ratio of S G to the total forecasted load L forecast . In any forecasting method, forecasting error is always present due to uncertainties in power systems caused by the unpredictable variations of power output from renewable sources (e.g. photovoltaic sources, wind power supply) and abrupt changes in power demand. The LS implemented based on P R defined above could be insufficient due to under-forecast of the load. To deal with this risk, the power supply capacity S j new allocated to controller K j is determined taking a margin M into consideration: Here M indicates the maximum forecasting error E forecast that is likely to happen, i.e. 95% of errors are below this value. This is to ensure that no repeated LS implementation happens over a short period of time due to insufficient LS.

2.3.4
Step 4: load controllers' optimisation: Upon receiving the new allocated power S j new , each load controller K j distributes power allocation starting with the top priority level, level 1. The proposed method verifies if the power supply capacity is large enough to energise the appliances at that priority level. If that is the case, all the appliances in that level are allocated with power and the remaining supply capacity is handed over to the next priority level as the following equations indicate: where S j p is the power supply capacity available for the appliances at priority level p under the controller K j and L jk p indicates the total load of appliances of priority level p owned by user k under control of K j . If the power supply becomes, at a certain priority level p c , insufficient, i.e. S j p c < ∑ k = 1 m L jk p c , then optimisation is executed to determine which set of enrolled appliances of priority level p c is a candidate to be turned-off or deloaded.
Objective function: The optimal switching status x jk p c defined by (8) of each appliance is determined so that as much of the supply capacity S j p c as possible is allocated to appliances by minimising (7) under the condition (9) x jk subject to This binary combinatorial problem is to be implemented in each distributed load controller K j where its designed processor has limited computing capability to perform complex calculations due to cost considerations. Also, multiple sets of equally good solutions are expected to be derived from this optimisation process for fairness evaluation as explained further in the following subsections. Thus, the authors opted to utilise a meta-heuristic optimisation method due to its low computing capability requirement, considerable convergence performance, and the possibility of getting multiple sets of equally good solutions. However, meta-heuristic methods do not necessarily give the optimal solution. So, the condition (10) is used to judge whether the derived solution is good enough where L jk min p c is the lowest power rating among the turned-off appliances at priority level p c nominated by K j for possible power reallocation.
Population generation: The optimisation problem defined previously is solved by a metaheuristic technique, typically GA or BPSO. Either algorithm usually generates the initial population of solution candidates randomly, which are called individuals in GA and particles in BPSO and are vectors of switching variables x. However, it has been observed that those initialisations often lead to unacceptable computation time. In this paper, the researchers modify the initial population generation to speed up the running time.
Initial population depends on the relationships between the available power supply S j p c and the total target load L j p c as follows: When the condition (11) holds, there is nearly sufficient power that can be distributed among the connected appliances. Then, the solution is expected to contain many switching-ON actions. Therefore, each individual in the initial population is generated so that 90% of the variables x have value 1. When condition (12) is satisfied, the population is randomly initialised with equal chances for variable values of 0's and 1's. Furthermore, in the case where condition (13) holds, the initial population should have ∼10% chance of producing switching variable value of 1's or 90% chance of producing variable value of 0's. Optimisation method: The problem formulated in the above subsection is an optimisation problem that involves a large number of binary variables. There can be a number of equally good solutions that satisfy the condition (10). Among them the 'fairest' solution will be selected as described in the next subsection. So, we need to apply an optimisation technique that can deal with this type of optimisation problem and produces a set of solutions, not a single solution. Here, the authors have chosen GA which is widely applied in the field of function optimisation, combinatorial optimisation, and automatic control [25] known for its strong robustness and general optimisation ability [26].
GA showed satisfactory performance in both qualities of solutions and computation time [22]. In this paper, the BPSO, which is used by many researchers especially in power systems application due to its known promising performance [27], is also employed to confirm and analyse the satisfactory performance of GA.

2.3.5
Step 5: fairness function evaluation: It is not fair that a specific appliance of a certain consumer is repeatedly turned off while other appliances are kept on. To exclude these unfair solutions, Jabian et al. [22] introduced a fairness measure as shown in following the equations: R jk, p c ON = n jk, p c ON n jk, p c ON + n jk, p c OFF , where n jk, p c ON is the number of turning-on switching operations of the consumer k's appliance of priority level p c over a certain past period, n jk, p c OFF is the number for turning-off switching operations, and thus R jk, p c ON is the ratio of the number of turning-on operations to the total number of switching operations. Equation (14) becomes large when appliances which have been frequently switched on are switched on again, which is an unfair situation. Therefore, (16) is a good measure to evaluate quality of solution in both terms of efficiency (small unallocated supply capacity) and fairness when the weights a and b are properly determined by the DU.
Evaluation of the fairness function is always performed after every optimisation process at CS and K j to ensure that no consumer experiences repeated appliance de-loading.

2.3.6
Step 6: central station's optimisation: The CS collects the remaining unallocated power capacity after optimisation at each K j and redistributes them to appliances nominated by load controllers, where (17) is the objective function: x j cs = 1, on 0, off , and the constraint is where S T unallocated is the accumulated S j unallocated reported by each K j . After the execution of this final optimisation, CS communicates the switching status of the nominated appliances to each K j which in turn transmit the allocation to those nominated appliances. This optimisation is accomplished to maximise the power utilisation of the unallocated power S j unallocated from each K j . The final LS implementation commences as soon as the load controllers K j receive the switching combination from CS for the nominated appliances.

Economic analysis
To validate the feasibility of this research, economic analysis is conducted. Estimated DU loss, device cost, installation and maintenance cost, and consumer investment are considered. Although consumer comfort and satisfaction is one of the important values that the proposed LS scheme provides, its monetary evaluation is not given because of its difficulties.

Estimated DUs loss:
In the existing feeder-level LS implementation, every LS execution costs at least one or more feeder capacity to be de-loaded even if the necessary power reduction is only half of the feeder's capacity. This excessive LS can be avoided and become an additional DUs revenue when power utilisation is maximised.

Consumers' investment:
DU facility enhancements that directly benefit the consumers can be fully recovered from the consumers upon approval by the ERC [28]. So, the total cost in implementing the proposed method can be divided among the consumers, depending on their subscription and assumed as their corresponding investment. In the Philippines, specifically in Iligan City, residential consumers are categorised into two, namely, the regular residential consumers and the lifeline residential consumers. Lifeline residential consumers are those who have minimal power consumption (i.e. lighting fixtures only) due to their social economic status and are ∼40% of the total registered residential consumers. Hence, it is important to identify the consumer's subscription category because the device cost also varies depending on their subscriptions. Consumers under the lifeline category are assumed to have a single relay located before their main circuit breaker as the control point to minimise the investment cost. Consumer investment formulas are given below for regular residential consumers and lifeline residential consumers, respectively: where C dev T reg is the total device cost for all n regular regular consumers, C fhs T is the total cost for firmware, hardware, and software, C m T is the total maintenance cost, and C dev T lifeline is the total device cost for all n lifeline lifeline consumers. In line with ERC, that mandate to fairly recover costs from the consumers, they are only charged those devices and accessories that they benefit from. Since all consumers equally benefit from the DUs overall software, firmware, server and maintenance costs, the total cost of C fhs T and C m T is divided to all consumers to recover its costs. On the other hand, devices' cost for a certain consumer group is also divided among themselves.
Consumer investment is evaluated using the future value formula (22) and annuity payment formula (23) given respectively as follows: where FV is the future worth value, PV is the present value, C p is the monthly payment for every consumer, r is the rate of interest, τ is the compounding period in one year, and φ is the number of months payable.

Performance evaluation
In this section, several simulations are conducted to show effectiveness of the proposed LS implementation in efficiently utilising the available power supply.

Simulation model
The simulation model is based on a local DU in Iligan City, Philippines. The local DU has the average demand of ∼20 MW with more than ten installed feeders for residential and commercial consumers. Out of 65,000 total registered consumers, ∼60,000 of them have residential classification. Currently, the local DU employed supervisory control and data acquisition (SCADA) in its distribution network for transformer and feeder control and monitoring only. However, in this research, it is assumed that CS is integrated in the local DUs' SCADA for continuous overall system voltage and frequency monitoring, data logging and overall power management.
In these simulations, the connected loads of the regular and lifeline residential consumers are assigned to a distributed load controller K j . Each K j is assumed to have a maximum connected load of 150 kW, where all K j 's are to perform load optimisation simultaneously. The appliance power ratings owned by the regular consumers are randomly selected based on the actual appliance power rating provided by the local DU. Also, priority levels are assigned depending on the type of appliances (i.e. lightings, kitchen appliances, ventilation, entertainment etc.) and ranked based on typical consumer must-haves during low power supply situations. On the other hand, lifeline residential consumers are assumed to have only one type of appliances (i.e. lightings) and assigned to have the top priority level. Fig. 2 shows a histogram of a typical 1-week forecasting error percentage calculated by (3). The results indicate that the errors are <6% with only two outliers which are caused by unexpected power curtailments.

Short-term load forecasting
To get the margin M, a cumulative error density analysis is performed as shown in Fig. 3. This indicates that errors smaller than 2% happen at a probability of 95%, and, therefore, we set 2% as the margin to be used in (4).

Fairness and optimisation method evaluation
GA showed satisfactory performance in both quality of solutions and computation time [22]. To further verify its superiority in this specific optimisation problem, we employed BPSO for comparison. As mentioned, these techniques are widely used in power systems applications. The two methods are evaluated in terms of the achieved unallocated power and corresponding CPU time after 1000 iterations. In this comparison, a single iteration in GA comprises an evaluation of 50 individuals in a population, individual ranking and selection, and reproduction using crossover and mutation strategy. On the other hand, a single iteration in BPSO consists of evaluating 50 particles in a swarm, evaluation of local best and global best to update the velocity function in order to produce new particles. Each of the particles in BPSO and individuals in GA is composed of 50 binary numbers that represent consumer appliance switching conditions.
After the selection between the mentioned optimisation techniques and a reasonable good solutions have been derived, they are checked for fairness using (16). The 'fairness' of derived solutions is evaluated by means of their corresponding coefficient of variation (CV) as shown in the following equation: where μ is the mean of turning-ON frequencies R jk, p c ON of appliances and σ is the standard deviation. A lower CV value signifies that the derived set of switching operation is fairer.
Based on several simulations, it is observed that the fairness function and optimisation methods vary their performance depending on how the appliance power rating distributions are diversified. Hence, case studies are conducted for the appliances combinations with low and high diversities of appliance power ratings.
3.3.1 Case 1: appliances with low diversity of power ratings: Low diversity of appliances possessed by consumers means that they have appliances with similar power ratings. This is usually the case for appliances with relatively high priority levels because consumers have similar needs for those appliances.
Here, we assume that there are only two kinds of appliances at priority level 3 owned by 50 consumers under a certain K j . Actual power ratings are assigned to appliances randomly as shown in Fig. 4.
Optimisation method comparison: Fig. 5 shows the convergence process of GA and BPSO where 1336 W of supply capacity is distributed to 50 appliances. After 1000 iterations, GA derived an unallocated power of ∼20 W and BPSO achieved ∼100 W, which indicates that GA has the advantage over BPSO in terms of effective allocation of the supply capacity. Fig. 6 shows processing time of both GA and BPSO. GA implementation converges faster than BPSO; its CPU execution time is approximately <30% of the BPSO's.
Fairness evaluation: Based on the previous subsection, GA shows superiority in terms of processing time and achieved good solutions. Hence, we used GA's achieved solutions and evaluate their corresponding fairness values. Fig. 7 shows the CV of the switching combinations derived with and without the fairness function. From the figure, it can be deduced that the CV is lower when the fairness function is considered and that the fairness function is effective to produce fairer solutions.

Case 2: appliances with high diversity of power ratings:
Appliances with high diversity of power ratings appear at lower priority levels. Fig. 8 shows the appliances combination with highly diversified power ratings. This time, consumers are   Optimisation method comparison: Fig. 9 shows the convergence performance of GA and BPSO for the appliances with high diversity of power ratings. Here, BPSO leads slightly compared to GA based on their individual achieved value. Fig. 10 shows both GA and BPSO processing time after 1000 iterations. GA has shorter CPU execution time compared to BPSO which is approximately <30% of BPSO's.
Fairness evaluation: As shown in the previous subsection, BPSO performs slightly better than GA in terms of their achieved unallocated power. However, GA is found to be superior over BPSO in terms of CPU processing time. Thus, GA's achieved solution is used to determine the CV of switching combinations obtained with and without the fairness function as shown in Fig. 11. By inspection, the switching combination with the fairness function have lower CV values and thus are fairer.

Conclusions on fairness and optimisation method evaluation:
Both GA and BPSO showed exemplary results in minimising unallocated supply capacity. GA is more time-efficient and more suitable for real-time applications. These support our previous finding that GA is appropriate for this type of problem. The superiority of GA in computation time can be explained as follows: in GA, there is a possibility that the new solution directly   inherits an excellent parent trait by skipping the mutation and crossover process because the process is applied probabilistically, whereas, in BPSO, every time a new particle is generated it is different from its parent and loses the excellent trait.
The CV values confirm the effectiveness of the proposed fairness function.

LS optimisation performance evaluation
Now that we have confirmed that GA is an appropriate optimisation method for this problem and that the fairness function is useful, we closely investigate optimisation performance at both CS and K j levels by GA with fairness function.

Distributed load controllers' optimisation:
The actual appliance power ratings for residential consumers posted in the DU's website are considered in this research. Fig. 12 shows a sample of randomly selected appliance power ratings for the optimisation process in K j at priority level 4. Fig. 13 shows the fairness values of the possible switching combinations derived after employing GA. As shown, the seventh switching combination provides the least fairness value or the fairest among the switching combinations, hence, the seventh switching combination is selected as the initial switching combination for an LS implementation. Fig. 14 shows the power allocation for every priority levels in K j . In priority levels 1-3, demand power was matched with allocated power. This means that appliances in these priority levels will be energised. On the contrary, in priority level 5, demand power has no or zero allocated power. This means that all appliances will be de-loaded.
The optimisation begins in the case of priority level 4 wherein the demand power is greater than the allocated power S j 4 .    Optimisation process using GA is employed at priority level 4 wherein out of 23,100 W power to be allocated, S j 4 , the final unallocated S j unallocated is 21 W. The resulting unallocated power S j unallocated is already negligible since it is lower than the lowest power rating of the appliance connected at this priority level. However, to further optimise the system, this unallocated power S j unallocated is reported to CS for the second optimisation process done at CS for final power reallocation. Fig. 15 shows the power reallocation for the nominated appliance L jk nom by each load controller K j .

Central station's optimisation:
In this simulation, the DUs CS receives 2618 W total unallocated power S j unallocated reported by each load controller K j . After GA implementation, a total of 2617 W of the appliance power rating were supplied and only 1 W is the remaining overall unallocated power. The corresponding CPU processing time was 103 ms.

Economic analysis
3.5.1 Estimated DUs loss: Currently, 14 feeders are installed with the maximum feeder capacity of 6.7 MW and a minimum feeder capacity of 1.81 MW. Even when one of the feeders is switched-off, it is already a huge amount of loss for the DU and discomfort on the part of the affected consumers for the duration of ∼2 to 4 h.
Based on the optimisation results of the proposed LS implementation, the unnecessary de-loading or unutilised energy can be reduced to a negligible level, thus, can maximise the DU's revenue.

Consumer's investment:
Initial costing for the lifeline consumers is ∼Php3916 which composed of a wireless module, a digitally controlled relay, power supply, and a device casing. On the other hand, the regular consumer's initial device cost is ∼Php13,327 which composed of a Wi-Fi and a radio-frequency (RF) module for wireless communication, five smart outlets, and a house appliance controller board. The estimated costs for device components are based on the off-the-shelf devices that can be purchased online.
The total systems cost is based on the actual connected consumers of the local DU. The estimated systems cost is ∼Php46.69M for 3635 distributed load controllers K j and the DUs CS server. The estimated maintenance cost is ∼10% of the total systems cost [29].
The consumers' investment were calculated based on (20) and (21) for lifeline and regular residential consumers, respectively.
For lifeline residential consumers, the total initial cost is ∼Php3967 while regular residential consumers has the initial cost of ∼Php13,822. Assuming that the DU shoulders the overall initial investment and obliges the consumers to partially pay the invested amount in a monthly basis for about 5 years with 12% interest rate, by using the future worth formula (22) and annuity (23), the monthly payable for lifeline subscription is ∼Php155 and ∼Php540 for regular consumers. Note that the average residential consumer in the Philippines consumes ∼400 kWhr every month which is about Php4800 [30]. This cost will result to an increase of ∼11% of the average monthly bill of a regular consumer.

Conclusion
In this paper, we proposed an LS implementation that aims to maximise power utilisation by replacing the feeder-level LS with appliance-level LS. Also, the proposed method provides an opportunity to consumers to take part in the decision making in every LS implementation by means of assigning a priority level to their chosen connected loads.
The appliances' historical switching activities were evaluated using the developed fairness function, as introduced in the previous paper, to avoid repetitive de-loading. This is further verified with the resulting CVs. Moreover, power allocation during LS implementation is improved by incorporating short-term load forecasting errors to avoid insufficient de-loading and repeated LS in a short period due to underestimate of power supply shortage. The comparison of GA with BPSO confirms that GA is an appropriate optimisation technique when dealing with this kind of problem because GA has a chance to preserve excellent parent resulting in fast convergence. Furthermore, performance evaluations of the introduced fairness function and the optimisation techniques were completed using different appliance power ratings distribution diversity.
Based on the above results, closer evaluation and improvements on the proposed method confirms the promising outcomes with significant accuracy. Also, the proposed method is a sound investment since the un-utilised energy could be an additional revenue for the DU. On the other hand, consumer satisfaction will be enhanced since they can take part in deciding which appliances to turn off first during LS implementation. In effect, DUs will have a fair, automated, and efficient LS process which improves system stability, reliability and power quality.

Acknowledgment
The researchers thank and acknowledge the contribution of the Republic of the Philippines, Department of Science and Technology, Engineering Research for Development and Technology (DOST-ERDT) for the scholarship support.