IIRC, 15kW parts are okay for lower levels of load dump, maybe not the worst case you've selected. There are 30kW parts out there too. The lower voltage rating helps.
Note that Fig.1 must eventually flatten out: the 8W continuous rating will take some 10s of s to stabilize at. The sub-sqrt slope of the plot suggests limited thermal diffusion (exactly, diffusion would go as t^(-1/2)). It's not clear if that would get steeper over package-scale time scales (i.e., it takes ~10s ms for heat to diffuse from die to lead frame and package). There's nowhere to go beyond the package, so the curve should steepen at that point, until lead conduction and convection dominate, then finally flatten out at the 8W figure (in the ~10s s range).
If we assume the curve remains flat, then we can take its slope of ~ -0.443, and... let's see, the rated pulse is a 10/1000us waveform, but that's the 50% figure; helpfully, the curve shows it reaching 10% at about 4ms. The load dump waveform takes 400ms to reach 10%. If we simply equate these, then we need a 100x time figure, which would be 100^(-0.443) = 7.7 times lower power, or 1.95kW.
The load dump peak power is Vc * (Us - Vc) / Ri or 4.8kW, so it seems this part is undersized by ~2.5 times.
Which hey, that meshes with my expectation that it's fine for lower ratings, but not the full thing. And in fact an even bigger one is needed then (probably two 30kW in parallel would be safe enough? -- three 15kW would be too marginal I would say, but four or more is probably fine).
Contrast with MOVs, for example:
https://www.yageo.com/upload/media/product/productsearch/datasheet/cpc/mov/20D_1.pdfNothing particular about this brand, I think they're all the same materials; ratings/specs are more or less industry standard, curves always look the same.
Notice 18-68V types are only rated 20Apk, and the 82V part clamps at the same voltage as the 68V but at 5x the current! (LV MOVs rather stink.)
Looking at the clamping curves, the 18V and 22V parts clamp at 100A and 40/50V respectively, which is... useful enough, but the energy ratings are tiny. (A much bigger disc would afford more energy, but likely not enough.) The derating curves suggest for N = 2 about a t^(-1/2) slope, and the pulse is the same type so let's assume 10 times the highest (10ms) point or sqrt(10) less current, or about 13A. Or 22A for the J version ("high surge").
They don't say here BTW, but the maximum current spec I believe is for 8/20us, which, let's see... lines up on the derating plots for single event, yeah. So that's why that figure is so high.
So, for about a 100A surge at this level, you'd need more like a stack of 5 of these in parallel, preferably 7-10 for better sharing/reliability. Or larger discs, but you'll still need multiple in parallel I think.
Note they aren't rated for leakage; that's probably a bad thing for automotive.
As for relays, you're not thinking of using one to disconnect the load, are you? They don't respond fast enough (~1 to 10s of ms).
Tim