An Optimization of UAV-Based Remote Monitoring for Improving Wildfire Response in Power Systems

Wildfires lead to colossal losses on territory, local, state and federal levels, affecting critical infrastructure, the economy, decarbonization goals, social sustainability and more. Although wildfire impacts highlight the urgent need for resilience-comprehensive methods in power system wildfire response, existing techniques often focus on a single phase, usually wildfire progression. In this work, a comprehensive approach is proposed to provide optimal and real-time information toward mitigating wildfire risk in all resilience phases, necessary to decompartmentalize wildfire response. This paper focuses on the optimal routing of the remote monitoring resources for a self-sufficient low-cost wildfire mitigation model (SL-PWR), which utilizes predicted spatio-temporal wildfire probability maps of the utility service area and optimized unmanned aerial vehicle (UAV) monitoring trees to obtain input images for training the SL-PWR modules. Results show that optimizing the SL-PWR’s UAV monitoring using predicted wildfire threat parameters can improve situational awareness and rapidity of detection during wildfire incidents.


I. INTRODUCTION
T HE United States is directly and indirectly affected by wildfires.For instance, the annualized economic burden of wildfires is estimated up to $347.8 billion [1], while the increase in emissions in the 2020 wildfire year is about twice the reductions achieved from 2003 to 2019 [2], affecting the federal government's net greenhouse emissions reduction goals.Critical infrastructure such as power systems suffer colossal losses from wildfire events.Exogenous wildfires from external sources spread to and damage equipment, while endogenous wildfires caused by power equipment have adverse effects on utilities in whose service area they are ignited.For instance, electric power utilities face lawsuits and bankruptcy following endogenous wildfires [3].

II. RELATED WORK
The adverse consequences have led to the development of wildfire technology and design solutions such as the United States EPA ''Wildland Fire Sensors Challenge'' [4].Utilities are adapting endogenous wildfire mitigation tools such as the distribution fault anticipation (DFA) tools that detect circuit anomalies and notify utility operators before these anomalies spark fires [5].Utilities have also adapted other methods such as installing monitoring cameras in high threat areas, as well as lookout observers and phone calls from customers toward detecting wildfires.Unmanned aerial vehicles (UAVs) have also been utilized in wildfire mitigation efforts [6].In [7], the applicability of wireless sensor networks integrated with UAVs is studied in a natural disaster setting.In [8], a methodology is proposed for real-time UAVs to detect forest fires with a color index.Fire detection approaches using a team of UAVs and ground vehicles is discussed in [9].Reference [10] presented an image processing methodology by using infrared images obtained by UAVs for the detection of forest fires, while [11] employs a quadrotor based on smoke for fire detection.Fire and smoke detection models have been trained on satellite imagery [12], [13] like GOES-16 and MODIS, and also using UAV images [14], where convolutional neural networks (CNNs) have been seen to outperform other learning models.Tools such as the FireALERT MK by Vigilys [15] and the FIREBird [16] have been developed, with certain drawbacks addressed by the SL-PWR.To the best of authors' knowledge, no study has attempted to design and implement a resilience-comprehensive model that intelligently and automatically processes and provides highly crucial information to critical infrastructure operators and appropriate emergency response teams in real-time.
In our previous work [17], the use of quantitative data from diverse relevant sources to spatio-temporally estimate the probability of wildfire occurrence, fire spread and risk of the power grid to wildfire threats, produces a wildfire potential map of the study area.This spatio-temporal output is utilized in SL-PWR, as illustrated in Fig. 1, to optimize UAV remote monitoring.Hence, this work follows from our work in [17], going beyond prediction and estimation, further into real-time analysis for prevention, mitigation, and asset management, towards a comprehensive resilient response of the critical power grid to wildfire threats.
In this paper, we present the design and implementation of the UAV-based remote monitoring that enables rapidity of detection and localization in real-time wildfire response.To aid understanding and follow-through, we first review the functionality of the SL-PWR before discussing the optimal design and implementation of the UAV-based remote monitoring of the SL-PWR, which is the scope of this paper.
The rest of this paper is organized as follows.A review of the functionality of the SL-PWR model is discussed in Section III.The Section V then discusses the optimization of the system UAV resources for the SL-PWR model, discussing in detail the optimization process and the building of the UAV monitoring tree.The simulation and result for the SL-PWR model includes the working of the UAV optimization as furnished in Section VI.The paper is then concluded.

III. SUMMARIZING THE SL-PWR METHODOLOGY A. THE STRUCTURE OF THE CNN
The CNN is a multi-layer deep learning algorithm that takes image input and assigns parameters (weights and biases) to aspects of the image in order to learn from such images.It consists of convolution, pooling, and fully-connected layers.The convolution layers are made up of filters which basically translates to a matrix of weights called feature maps.The CNN operator at each layer completes the function in (1), where the layer under consideration is i, j is the feature map under consideration in layer i, Y x, y ij is the output located at position ( x, y) in feature map j and layer i, A(•) represents the layer's activation function, b ij is the bias term, ω pq ij denotes the weights/value of the convolution filter (F × F), at position (p, q), associated with layer i and feature map j.Hence, by convolving the filters with the input image and carrying out non-linear transformations using activation functions, N feature maps are created.The activation function used in SL-PWR is the ReLU (Rectified Linear Unit) function as in 2.

A(x) = max(0, x)
( The pooling layer(s) operate by obtaining the average or maximum value of the feature map elements where its window slides, extracting the dominant features, a dimensionality reduction of sorts which also helps to improve computational efficiency.Together, the convolutional layer and the pooling layer form the i th layer of the CNN.The fully connected layer is one that learns the non-linear combinations of these high-level features as transformed by the convolutional layer.It flattens the feature maps feeding directly into it, and then feeds into a feed-forward neural network, learning the parameters (ω, b) by minimizing the negative log-likelihood as in (3).
where I k is the correct (target) class label for the input image under consideration.This objective is optimized by applying stochastic gradient descent with back propagation using the chain rule as in (4), to training iterations over several epochs.
where µ is the learning rate, N i is the total number of layers in the network, Y n i is the output of layer i during iteration n.With this process, the model is then capable of distinguishing dominant and less-superior features in the input images, further classifying them using the Softmax function, an adaptation of the Sigmoid function used for multi-class classification, which takes in the vector of R real numbers and normalizes them into a probability distribution of N probabilities which are proportional to the input exponentials as in (5), where f c (I k ; (ω, b)) is the scores from each of the multiple classes of interest c ∈ {1, • • •, N } transformed into conditional probabilities.In order to test the CNN model after the training process described above, the output layer then predicts the label I of the image input I using the argmax of the Softmax-transformed probabilities as in (6).
The proposed SL-PWR consists of sub-modules based fundamentally on the Residual Neural Network 18 (ResNet18).

B. SL-PWR MODULES AND POWER SYSTEM WILDFIRE RESILIENCE
The SL-PWR consists of 4 modules which inform the wildfire analysis (pre-wildfire) phase, the wildfire progression (during wildfire threat), and the restoration (post-wildfire) phase, as illustrated in Fig. 1.In particular, the SL-PWR as shown in Fig. 1, receives the service area's spatio-temporal wildfire potential map, which has been divided into grid cells with centers located at g c .loc with a latitude and longitude to which UAVs can be routed to capture images [17].Hence, the g c .locs with extreme wildfire probabilities are used in optimizing the UAV monitoring trips.The functions of the modules are further enabled by sub-modules which extract spatial details for detection, localization, classification, and estimation, discussed as follows.

1) THE VEGETATION MODULE
This module is active in the ''Wildfire Analysis'' phase, and functions to: 1.) detect the type of vegetation in a service area, given that vegetation serves as wildfire fuel, this can inform the expected speed of fire spread and the vegetation management process.2.) estimate the clearance distance of vegetation [18] from power lines as mandated by Occupational Safety and Health Administration (OSHA), to prevent line sags/sways that can cause flashovers when electricity arcs from an energized line to nearby vegetation.Hence, the vegetation type module not only helps with the vegetation management plan drawn by the arborists but also helps with mapping the spread rate of the different vegetation.

2) THE POWER EQUIPMENT MODULE
A line can remain energized on the ground [19].For example, high impedance (HiZ) faults, which draw current too small to trip circuit breakers, can fall to earth and remain energized for long periods of time and can serve as sustained ignition sources.The power equipment module is active in the ''Wildfire Analysis'' and ''Wildfire Progression'' phases where the UAVs travel along their monitoring paths, which are optimized with parameters such as the equipment age and fault frequency as well as fault anticipation tools like the DFA, inspecting fault prone lines in high threat grids.Hence, the module informs the operator when the equipment risk has become a wildfire ignition i.e., ''Wildfire Progression'' phase to aid utilities route resources (e.g., repair crew vs. firefighting crew) accordingly.

3) THE WILDFIRE MODULE
This module is active in the ''Wildfire Progression'' phase and functions to detect wildfire and smoke, localize the wildfire and estimate its spread in the grid.The localization finds where the wildfire is in the grid, calculates it's spread area, and enables the rate of spread estimation in real-time without dependence on mathematical models, vegetation models, or quantitative data.The localization also enables the SL-PWR serve as a first responder, to begin containing the wildfire early on after detection, as alerts are sent to firefighters and utility crew.This can aid in extinguishing ignitions which can become wide spread before firefighting, utility, and emergency crews arrive.

4) THE BURNT EQUIPMENT DETECTION AND ESTIMATION MODULE
This module is active after wildfire suppression, and functions to detect damage type and estimate equipment damage extent in the restoration phase.This module reduces the time and cost of damage estimation conventionally done by routing manual monitoring crew, hence enhancing rapidity in infrastructure restoration and improving resilience.A major advantage is the provision of images extending transparency in cost estimation, and providing a digital twin as a blueprint for asset restoration.

IV. UAV RESOURCE INTEGRATION TO SL-PWR A. QUANTITATIVE INPUT DATA
The STWIP, presented in our work in [17], produces potential wildfire ignition maps which provide locations with their probabilities ([0,1]) of ignition readiness.A spatial location is a point i with geospatial coordinate i.loc defined by a latitude and longitude (lat,lon) at any location in a grid cell.The grid cells are g×g km polygons with a centroid with geospatial coordinates g c .loc.The STWIP provides g c .locs with levels of wildfire threat according to the potential ignition probabilities of i.locs within the grid.For instance, on the potential ignition map in Fig. 2, the grid G1 contains high threat locations while G5 contains moderate threat locations, hence G1 is an extreme threat grid, G5 is an elevated threat grid.This information, (g c .loc, threat level), serves as input for optimizing the UAV navigation.

B. UAV IMAGE CAPTURE
Conventionally, utilities perform visual inspection using manual field surveys like foot patrol crew and manned helicopters.More advanced techniques have also been employed in literature including aerial images from manned helicopters and fixed-wing platforms, land-based platforms, airborne laser scanning using Light Detection and Ranging (LiDAR), synthetic aperture radars, optical satellite, and UAVs [20], [21].Satellites orbiting at lower (500-2000 km) altitudes can detect wildfires in the early phases due to their finer resolution, although it can take several hours to days.For example, VIIRS has a 12-hour revisit time while the Landsat-8 has a 16 day revisit time hence it is rare that one of these satellites provides the first wildfire alerts [22], [23].Using satellites have additional limitations including: 1) The detected heat signatures are averaged over pixels, making it difficult to pinpoint fire location and size.2) Wildfire intensity is indicated by thermal signals which can be smoldered by smoke and hence radiate relatively less energy, causing data misinterpretation.3) At lower altitudes (up to 215m) UAVs can capture 5cm resolution imagery, which is much higher than imagery captured from satellites, 25cm resolution [24].
However, with newer advancements in satellite imagery, these limitations can be reduced, and with improved continuity in monitoring, the satellite images can also be served to the SL-PWR as input for real-time processing.Also, helicopters and airplanes can also be used as conventionally done by power utilities, however, [24] discovers that Low-flying airplanes can capture comparable imagery to UAVs, but are expensive to hire and flying at low altitudes increases the possibility of a crash, thus employing the UAV technology lowers costs and improves operator safety for such missions.
Hence in this work, we choose to use the UAVs since: 1) UAVs provide more continuous monitoring than satellites, which periodically visit specific parts of earth.2) Unlike satellites, the system would rarely be blinded by local weather conditions or smoke/dust of wildfires.3) Satellite imagery e.g., from Google Earth can only support so much detail in the image resolution before images begin to appear blurred or pixelated.4) Most UAVs do not need runways, takeoff can be from car-top launcher and recovery with parachute, if needed.
Additionally, although microwaves from synthetic aperture radars, also obtained from earth observation satellites, are capable of penetrating clouds, the UAVs are cost efficient for visual inspection and have widely been employed by power utilities, hence, will be an economic resource choice for SL-PWR since fewer investments will have to be made in terms of purchase and training the operating crew.Furthermore, discussions and studies on UAV operation during fire conditions have been thoroughly explored [25].
With the UAV-enabled SL-PWR model, the vehicles fly over the service area using the optimization model proposed in Section V.The geographical layout of the service area is as defined in Section IV-A.The image attributes (image, g c .loc (lat, lon)) are then sent to the SL-PWR processor via a communication link e.g., cellular communication or leased lines.As the UAVs fly over, images are captured from different scales/angles and taken at different times of day and weather conditions.To capture these conditions, diverse images are collected and augmentations/transformations are applied, which also serves to increase the training data.Given high temperatures during wildfires, the UAVs will fly at a certain derived altitudes, using the thermal sensor, in order to safely travel their routes.This consideration is integrated into the design of the SL-PWR prototype as in Fig. 6.

C. UAV CONTROL AND ROUTING
UAV control is performed in the mission planning ''ground station'' software to upload optimal routes to UAVs wirelessly, launch the UAVs, monitor trip progress and issue landing commands.In this tool, the aerial image of the service area to be monitored is highlighted within a rectangle, producing a preview of the proposed flight paths using waypoints which signify the UAV turning points in the trip.Typically, ''no waypoints zones'' (e.g., close to major airports) are also indicated by the software so as to mitigate the UAV flying into restricted airspace.After confirmation, the trip is uploaded wirelessly to the UAV via it's datalink which creates a communication bridge between the control software and UAV.The UAV can then be launched, its trip monitored through each waypoint, and automatically landed upon trip completion via the software.

D. IMAGE ACQUISITION AND PROCESSING
Image acquisition and processing are significant efforts in training the SL-PWR.A few databases [17], [26] exist for wildfire detection however these are limited to quantitative NASA data (NDVI and LST) for forecasting fires [26], wildfire hotspot images detected by NASA satellites and the Fire Information for Resource Management System (FIRMS) [27], and smoke dataset [28].However, there is no known database that captures the image set required by SL-PWR including utility equipment data, wildfire firesmoke data, vegetation type data, vegetation clearance data, and wildfire spread data.Hence, one of the contributions of this paper which is highly important in the field is the provision of these datasets [29].The SL-PWR ''image collection'' python scrapper code was developed for online search engines for RGB image data, and relevant images were retained.

V. OPTIMIZATION OF SYSTEM UAV RESOURCES
Most UAV path planning have been done using algorithms that are sampling-based, node-based, bio-inspired, mathematic-model-based, and multi-fusion based [30].In the sampling based algorithms, pre-known information of the UAV environment is used to map the environment as a set of nodes or to search randomly to find an optimal path.The node-based algorithms, like the Dijkstra's algorithm, generate paths based on a set of nodes The mathematic-model-based methods model the UAV space considering the kinematic and dynamic constraints and then bound the cost function with all the inequalities or equations to achieve an optimal solution.The bio-inspired algorithms emulate biological behavior to obtain path solutions, while the multi-fusion based algorithms combine two or more of the priors.In this work, we employ a multi-fusion based methodology which integrates sampling-based and node-based algorithms as in Algorithm 1, and our UAV routing strategy is an exponential distribution-based given that our strategy uses a novel dynamic dispatching rule with priority given to the grids with the dynamically and continuously highest fire risk, taking into account both historical data and other factors that may influence wildfire occurrence [31].
The SL-PWR provides a vital solution to pro-actively mitigate both endogenous and exogenous wildfires, provide real-time intelligence to actively contain progressing wildfires and enable response in emergency conditions, and aid rapid system restoration.However, the SL-PWR requires optimal monitoring routes for the UAVs in order to ensure efficient use of UAV resources and accurate monitoring strategy.Toward this, the SL-PWR utilizes STWIP's high accuracy potential ignition maps.First the STWIP sends the i.locs with different wildfire threat levels, where normal threat level i.locs have probability ≤ 0.5, elevated threat level: 0.5 ≤ probability ≥ 0.8 and extreme threat level: ≥ 0.8.The forecasted i.locs in the potential ignition map tend to have a ''cluster-like'' attribute, i.e., an i.loc of a specific threat level is more likely surrounded by i.locs of similar threat levels.Hence, the UAVs can be stationed in the center (g c .loc) of the grid with the cluster under consideration.Consider Fig. 2, illustrating the potential ignition map from STWIP where grids G1 and G5 have the extreme and elevated threat levels respectively, and Vg i is the vegetation type (crown Vg 1 , grass Vg 2 , litter Vg 3 ) associated with the grids.Furthermore, PE i is the weight associated with the amount, age, fault frequency, and fault anticipation of power equipment within the grids, for instance, PE 1 has a higher criticality/weight (4 lines, a generator and 3 buses), than PE 2 (4 lines and 2 buses) and PE 3 (2 lines and 2 buses), assuming all equipment age, equipment fault frequency/anticipation, and loads served in each area are equal i.e., L1 = L2 + L3 = L4 + L5.
The goal is to route UAV 1 to grid G1 to monitor the extreme threat for the length of time the threat is viable.However, in order to get to G1, the UAV 1 gets to travel along a path, and since the UAVs are limited resources, the operator wants to maximize the monitoring of critical grids without compromising with the risk posed by G1.To achieve this, a weighting/criticality factor is also assigned to the grids i.e., criticality of G1 ≥ criticality of G5 ≥ criticality of (G2 − G4, G6 − G9) on the potential ignition maps.A logical optimal route would be {G8 − G5 − G1} because in G8 there exists crown vegetation and higher amount of power equipment, increasing the likelihood of there being tall encroaching vegetation towards power lines.In G5 there are also 4 power lines with grass vegetation making the rate of spread of fires rapid if power lines sag or fall to the ground.Thus the operator will obtain more information via this path.
Since UAVs are airborne, trips between g c .locs are via a direct line of flight, as opposed to mobile crews going through a road network.The UAV routing problem is a bi-level one illustrated in Fig. 3 and formulated as follows:

A. THE UAV OPTIMIZATION PROBLEM 1) UPPER LEVEL
This determines the UAV path by maximizing criticality across service area's PE, Vg, and G layers as in Fig. 2. The objective ( 7) is to maximize grid criticality towards optimizing the UAV travel path to a destination grid.
The equation is maximized where PE i is the information of power lines (count, age, fault frequency/anticipation) in grid i, Vg i is the growth/spread rate of vegetation predominant in i, Cr i is the criticality of potential ignition in i (normal, elevated and extreme probability grids), and t i is the amount of time since the grid was last visited by a UAV.Hence, the objective of ( 7) is to obtain optimal UAV paths given that high vegetation growth locations quickly encroach power lines and need to be visited often, grids with high power equipment density are more likely to cause endogenous wildfires, and the grid criticality ensures that even elevated wildfire threat grids can also be visited even if not as often as the extreme grids.This is because, occasionally wildfires can occur outside the predicted extreme area as was the case with the famous Campfire.Lastly, t i ensures that grids of sufficient criticality are not overlooked for too long and are visited occasionally.
where N is the number of grids in the chosen path, s UAV is the UAV travel speed, t fuel UAV is time to fuel depletion, and p d i is distance it takes to get to grid i from the preceding grid i − 1 in the trip path.This constraint ensures that the UAV path is feasible with its fuel state.The output of this level is a set of selected paths P UAV = {p 1 , p 2 , • • •, p N } for every UAV.

2) LOWER LEVEL
The lower level problem is a maximum UAV monitoring coverage one taking in the output of the upper level, In each path p i , there are grids from i = {1, •• •, I }.The UAV optimization has to ensure that the UAV does not spend undue and valuable time monitoring/flying through the paths (I − 1) leading to the assigned destination grid I .Hence, this level ensures that the UAV-assigned extreme threat grid gets maximum monitoring coverage in and within the appropriate time. max P UAV y j ≤ 1 (10) In ( 9), the goal is to ensure that the selected paths with higher criticality is prioritized while ensuring that the UAV spends appropriate amount of it's travel time monitoring the assigned destination grid I , the last grid in the selected path, where y j is a binary variable indicating if a path is selected (y j = 1) for UAV j, τ I j is the amount of time spent at I by UAV j, and Cr p i sum is the path criticality i.e., weight of all grid nodes in the path.(10) ensures that for each UAV, no more than a travel path is selected however, once the UAV arrives at its I , another route can be added to its trip if the feasibility constraints allow.(11) ensures that the travel time through the path before the UAV arrives at I is less than the predicted start time of wildfire threat at I , T I s , and (12) ensures that the UAV keeps monitoring I for the threat duration, where T I e is the threat end time i.e., time until which wildfire threat is viable.

B. UAV OPTIMIZATION PROCESS
A graph-theoretic algorithm is proposed, a three-step procedure to determine the optimal UAV monitoring strategy: 1.) build UAV monitoring trees forming paths based on grid criticality 2.) path selection 3.) solve for maximum monitoring coverage at all I .The first step identifies paths to destination grids, where a UAV could have different monitoring paths forming a monitoring tree, with that UAV as root node.The second step, path selection, limits selected paths to feasible paths.The third, maximizes UAV monitoring coverage over assigned destination grids, which ensures that the UAV travels the most critical path, capturing informative images at grids in the path, while not spending crucial ''destination monitoring time'' on other grids in the travel path.Hence ensuring path travel time does not exceed start time of wildfire risk forecast at I , and that the UAV monitors I till the forecast end time.

1) BUILDING THE MONITORING TREES
The service area is modeled as an undirected graph G = (V, E) with PE, Vg, and G layers.The set of nodes V represent grid centers which carry grid attributes through the layers.The set of edges E represent inter-grid UAV flight path which is assumed to be a direct line of flight.A source node is the UAV take off node while a destination node is a critical/high risk node which a UAV has been assigned to monitor.Each node has a weight w whose value is the combined criticality of the grid across all layers.Each path has a weight Cr p i sum which is the sum criticality of nodes on the path.The UAV-monitoring path is that with highest criticality/weight that gets to the destination node within critical time.A modified Dijkstra's algorithm is used to obtain the paths from UAV source node to assigned destination nodes to form the monitoring tree and via pruning, infeasible paths are eliminated.The pseudo code for the building the UAV monitoring tree is furnished in Algorithm 1.
From the algorithm, the UAV monitoring paths are returned as a tree whose root node is sourced from s. Furthermore, for a node v ∈ V, v.dist is the distance from s to v which is the weight/criticality Cr p v sum of the nodes in the monitoring path from s to v, and v.dist will be updated to equal the weight of the monitoring path when a monitoring path is found.The q ← EXTRACT-MIN-DIST(S) 17: for node v ∈ adjacent to q do 18: ▷ Modified Dijkstra relaxation operation 19: Cr tot i = M-Cr Check the monitoring trip feasibility from s to v according to the UAV charge state, weather conditions on grids in the path, and travel time s to v should not surpass the start time of the wildfire threat forecast in T 43: if monitoring path is feasible from s to v then 44: ▷ Add nodes in the monitoring path including v to V tree , and add edges to E tree according to the sequence of the nodes in the monitoring path.Drop extra paths (e.g., path to destination already assigned to another UAV) if any, leaving one feasible path in v.paths 45: while v / ∈ V tree do 46: V tree ← V tree ∪ {v} 47: end while 50: end if 51: end for 52: ▷ Return 53: Tree ← (V tree , E tree ) criticality of each grid is adjusted to M-Cr tot * i in order to maintain the shortest path attribute of the algorithm since our objective is to maximize the weight of the chosen path and the Dijkstra's algorithm does not work well with negative weights, where M is a fixed number bigger than Cr tot * i across all UAVs.Additionally, v.path is the set that will contain the predecessors of v forming individual paths which are then appended to v.trip.A priority sequence S is used to store nodes that have not been explored by the algorithm and also to manage the nodes which form key-value pairs with the node's distance.The nodes are explored by extracting from S, the node with the minimum distance and adding such a node to the v.path for that UAV which should run from its source node to the assigned destination if a path exists.
Hence, Lines 2-10 initializes the parameters for the nodes towards implementing the modified Dijkstra's relaxation operation where distance value (dist) of all nodes are set to infinity while the source node's is set to 0 (Line 10).For each v, the v.path is a null set where the predecessors q of v are appended.In Line 8, a fixed number, M, bigger than Cr tot * i across all UAVs is defined in order to maintain the minimum distance attribute of the Dijkstra's shortest path.In Line 11, all nodes are inserted into sequence S, the set of destination nodes D is defined in Line 10, while the set of potential wildfire risk start time in each of the grids T is defined in Line 13.Since adjacent grid centers are equidistant, the time taken to travel a path can be easily obtained given the UAV speed.In the while loop of Lines 15-37, a modified Dijkstra's algorithm is utilized to find the UAV monitoring paths to the destination node.For each destination node, the node q, with minimum distance (in the first iteration, source node with s.dist = 0), is extracted from S and explored.After extracting q, the relaxation operation is applied to the nodes adjacent to q as seen in Lines 17-25, and it is removed from S. This is modified in this paper, since we seek the path with the maximum criticality for the UAV, hence the fixed number M is introduced for which Cr tot i must be positive across all UAVs.If q is a destination node, then a monitoring path is found for that destination node and q is removed from the set of destination nodes (Lines 26-27), v.path is appended to v.trip (Lines 28) while the UAV moves along to another destination node in the same trip if feasible.This ensures that in a UAV trip it could be able to get to more than one destination node if possible.This is enabled by Lines 31-36, the node set in G is appended back to S if the sequence becomes empty before D becomes empty (signifying the end of a trip).This ensures that each destination node is explored and reached once if a path exists, and while all destination nodes have not been explored but S is empty, the UAV monitoring path is renewed from s to form several other paths (and possibly trips) to explore the destination nodes that have not been explored.In Line 32, the elements of set D * are the leaf nodes of each UAV tree as in Fig. 4. The search for the monitoring path ends when all destination nodes have been explored if path found (D = ∅) or otherwise e.g., when there is an obstacle such as storm that prevents the UAV from traveling through a grid.Lines 38-53 build the UAV monitoring tree, where a feasible trip is chosen for each UAV from source node s to destination node ∈ D * .The graph Tree which is made up of (V tree , E tree ) represents the UAV monitoring tree where V tree is the set of nodes of the tree, and E tree is the tree's edge set, not including infeasible paths or dropped trips.

VI. SIMULATION AND RESULTS
In this section, we evaluate the results obtained from the UAV optimization of the SL-PWR.A gridded area which consists of 16-grid cells superimposed on the standard IEEE 33-bus test system is used to test the UAV optimization process and monitoring trees in order to validate the effectiveness of the proposed method.The one-line diagram of the test system is shown in Fig. 5.The layer 1, layer 2, and layer 3 weights are chosen off a random distribution [0,1] for the different grids, t i is assumed to be 1 ((t 1 i − t 0 i ) = (1 − 0)) since the first optimization time step is illustrated, while M is chosen to be 1.The same graph structure is used for the modified Dijkstra's relaxation to get the UAV tree, and then the graph attributes are modified to inter-grid travel time, to select feasible UAV trips.For adjacent grids, the travel time is 1, while for diagonal grids, the travel time is 1.41.The trip feasibility based on the travel time of the paths which make up the trip, is informed by: 1.) Forecasted threat start time: when STWIP or any utility prediction technique indicates a wildfire threat.From this time, if that grid is exposed to an ignition source, a wildfire can be sustained.Hence, the UAV must arrive at the grid on/before this time.2.) Monitoring time: that the wildfire threat in a grid is viable and hence the UAV must monitor this grid for this threat duration before routing to another destination if feasible in the same trip.
We consider two scenarios which inform the mission planning software as in Tables 1 and 2, a) The Lax case: where monitoring time is an estimate, e.g., extreme threat expected from noon to within 5-7pm when temperature, a wildfire  contributing variable, goes down.b) The Strict case: where utility's confidence in the wildfire threat forecast model is high, and UAVs are routed strictly to that forecast.The difference being that in the strict case, the UAV must monitor till the end of the threat period while in the lax case, the UAV can leave the grid before estimated threat end time since, if a wildfire did not occur within about 90% of the threat duration, chances are that it would not occur in that grid.Hence, the UAV saves some time and routes to the next destination.
Here, [G7, G14, G15, G8, G9, G12] are chosen as high threat grids, ordered according to forecasted threat start times (T I s ), while the UAVs all source from [G1].In Table 1, the first destination grid is G7 with T G7 s = 3, the earliest time of UAV arrival is at T = 2.41 < T G7 s = 3, so that the UAV can positioned to monitor before the wildfire risk begins.After the UAV arrival, the time the UAV should remain in the grid for monitoring τ G7 1 = 3, which means this is the estimated duration of the wildfire threat.The UAV leaves G7 at T = 2.41 + 3 = 5.41 and checks the grid cells that are reachable from the current G7 given the T I s .G8 is the next feasible destination from G7 as by the time the UAV gets to G8, the time would be 5.41 + 1 (adjacent grids) = 6.41, and T G8 s is 6.5, hence the UAV can make it in time to its second destination, the UAV then spends τ G8 1 = 3 and leaves G8 at 9.41.At that time, the UAV cannot reach any other destination grids in its current trip, making UAV 1 have 2 paths in its trip.Next, UAV 2 picks up from ''G1 -G14'' with earliest arrival time at 3.41 (1.41 + 1 + 1) it then spends τ G14 2 = 5 and leaves ''G14'' at 8.41 by which time it cannot make it to any other grids before their T I s .The trip to G9 is interesting as the earliest arrival time would be 2.41, however, the threat start time T G9 s = 7, hence, the operator starts routing the UAV on/before T = 4.59 with the leave time approximately T = 9.Same applies to ''G12''.Table 2, illustrates that the threat forecast model confidence can change/improve the UAV routes.For instance, if the forecasting model accuracy for T I s and τ I j is closer to 100%, then the UAV strictly follows these times and will only leave a grid at T = T I s + τ I j .In this case, the UAV 1 routes from G7 to G12.The strict case would be best with more UAV resources that trip time conservation can be overlooked.Runtime for the UAV optimization code is 0.0182 secs.
Furthermore, the UAV optimization spatio-temporally considers grid weather.A work-around would be avoiding routes with high wind gusts, but there may be only one efficient route given forecast start time.Hence the optimal SL-PWR technique schedules medium/large UAVs for the trips if high wind gust will occur during the times the UAV is flying through.Similarly, the total trip time also influences the UAV type scheduled.For instance, for a 10 hour-trip, a medium UAV can be scheduled.Table 3 shows the basic attributes which can be improved, e.g., Penguin B (medium UAV) with an optional 7.5 L capacity fuel tank and an 80W on-board generator, improves flight time from 6 hours to >20 hours.
Thus, the larger UAVs, as illustrated in Table 3, can aid initial firefighting efforts.Once SL-PWR detects a fire condition, an alarm is raised to appropriate utility and emergency response teams with visual and real-time information.Given the UAV payload capacity for carrying firefighting fluids, the early-detected-fire and hence less-spread wildfire, can be simultaneously contained by applying firefighting fluids on the wildfire boundary localized by the SL-PWR, helping curb wildfire spread pending when response teams arrive.

VII. DEPLOYABILITY OF THE SL-PWR
The SL-PWR is designed so it can be deployed in different ways including Energy Management/ Outage Management/ SCADA rooms, or deployed in hardware.

A. ENERGY/OUTAGE MANAGEMENT SYSTEM-CONTROLLED SL-PWR
Depending on utility preferences, the SL-PWR can be deployed as in Fig. 1, where the processing software GUI is managed from the operation room.Already existing utility monitoring methods such as service area cameras and satellite imagery can be processed by the SL-PWR to provide realtime qualitative information to appropriate wildfire response teams.This deployment method may add additional time to the processing of the input images depending on the time it takes the internet service provider (ISP) communication service to transfer the captured images to the SL-PWR software at the operator's base.Some utilities have also begun integrating private communication networks to mitigate dependence on third party service providers.With the UAV-enabled SL-PWR model, the UAVs fly over the service area using the SL-PWR's spatio-temporal optimization model.The image attributes [image,UAV geographical location (lat, lon)] are then sent as output from the UAVs and input to the SL-PWR processor via a communication link and follows as described in the earlier sections.

B. DEPLOYMENT POTENTIAL IN COMPACT HARDWARE
The SL-PWR could also be embedded to function in a device that would be mounted on the UAVs as in Fig. 6, depending on the requirements of the utilities.This mode of deployment will retain the 40msecs time of detection, while the system operator deals with verification of alerts sent in by the SL-PWR by looking at the images associated with the alerts and following protocol thereafter.In this mode, the operator can still control and manage the ground station software of the UAVs if the utility requires.The following illustration visualizes the compact mode of deployment.

VIII. CONCLUSION
This paper proposes the optimization of the UAV resources for SL-PWR, a self-sufficient low-cost wildfire mitigation model that benefits critical infrastructure.The proposed UAVbased approach is a proactive solution for both endogenous and exogenous wildfires that provides real-time intelligence to help actively contain progressing wildfires, facilitate response in emergency conditions, and expedite system restoration.The design and simulated implementation of the UAV-based remote monitoring optimization, using a data-driven and spatio-temporal technique is designed and implemented.Results show that optimizing the UAV monitoring routes using spatial and temporal wildfire threat parameters would aid early wildfire detection, improve situational awareness with efficient mitigation strategies, and provide effective use of limited UAV resources.With SL-PWR's optimized remote monitoring, a high level of automation is achieved, continuity in situational awareness can be maintained, and UAVs are optimally deployed which results in more monitoring trips completed within threat time, thus reducing risks from late wildfire detection through timely wildfire detection and comprehensive information for resilient response to wildfires.

FIGURE 2 .
FIGURE 2. Illustrating the UAV resource optimization problem.

FIGURE 3 .
FIGURE 3. Illustrating the UAV resource optimization problem.

VOLUME 10, 2023 683
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

Algorithm 1
Building the UAV Monitoring Tree 1: ▷ Initialization of parameters 2: for node v ∈ V do 3end for 10: s.dist ← 0 11: S ← V 12: D ← Set of UAV destination nodes 13: T ← Set of wildfire threat start time of elements of D 14: ▷ Obtain UAV monitoring paths 15: while D ̸ = ∅ do 16:

FIGURE 4 .
FIGURE 4. Illustrating UAV trips.With destination nodes {D 2 , D 3 , D 5 , D 6 }, the first trip may be {s, D 1 , D 2 , D 3 } before S is depleted, then S is replenished with V to make another trip which may contain {s, D 4 , D 5 }, and so, till trip {s, D 6 }.More efficiency as the UAV covers more destinations in a trip, as opposed to returning to s after each I.