How to get Wake-on-LAN working?
It sounds like some reading-up on WOL is probably in order first. Some additional reading about Ethernet, ARP, and UDP/IP is probably in order, too.
The WOL behavior in a client is triggered by a "magic packet". The magic packet can be encapsulated in any type of transport (UDP over IP, IPX, etc). The magic packet byte sequence just needs to be in the payload of a packet that the NIC to be woken-up will receive.
Sending WOL requests on the LAN is easy. Get a utility to make magic packets (like mc-wol for Windows) and fire away.
Sending WOL requests from the Internet is more problematic. You're on the right track w/ port-forwarding a UDP port from the Internet to the LAN, but there are other concerns.
You have a problem with unicast WOL requests from the Internet as follows: The MAC address of your WOL client computer won't be in the router's ARP table because, as the WOL client is powered-off, such an ARP entry would have aged out. When your router receives an IP packet port-forwarded to the WOL client's IP address while the WOL client is powered-off and its MAC address is aged out of the ARP table the router will not be able to deliver the packet to the client. You will need a router with static ARP capability to make this work.
The broadcast route is even less likely. Since you likely don't have a public subnet behind your router, but rather are using NAT to share a single public IP address on your LAN, there's no way to remotely address a packet to the subnet-broadcast address of your LAN subnet such that the router (if it could forward directed broadcasts) would generate a layer 2 broadcast with the WOL packet to allow the WOL client to "see" it.
WOL on your LAN will be easy. WOL across the Internet isn't so simple.
Static ARP is completely unnecessary
You've got the first part right. To get the WOL packet into the network, forward the packet coming from a specific port to the broadcast address of the local network (255.255.255.255) or whatever subnet range you want WOL access to be enabled on.
The Destination MAC address of the packet should be set to ethernet broadcast or FF:FF:FF:FF:FF:FF. The ethernet type of the packet should be 0x0842 (Wake On Lan).
So, where does the MAC of the computer being woken up go?
In the magic packet itself. WOL packets were only really intended to be sent across a local network. They're blind to any protocol above the link layer. To get around this limitation, they broadcast (ethernet broadcast not IP broadcast) to all the computers on the network and each computer reads the magic packet to see if they're the one being called on.
The contents of the magic packet contain 16 copies of the MAC address of the computer being woken up.
The frame structure is as follows:
DA -> SA -> Type -> Magic Packet
DA = FF:FF:FF:FF:FF:FF SA = [whatever the source MAC is] Type = 0x0842 Magic = [The actual destination MAC repeated 16x]
If you want to test that the packets coming across are in the right format, use the following filter in Wireshark:
ether dst FF:FF:FF:FF:FF:FF and ether proto 0x0842
Basically, the WOL application needs to be capable of creating a packet that spoofs the Ethernet Destination address. There are tools online that can do such a thing but I'm not familiar with them.
Note: The reason I know so much about this is because I'm the author of the WOL parser for SharpPcap (pcap wrapper in C#). If there is sufficient demand, I could extend my console application to include packet sending (it currently only sniffs) and make it available as an OSS project.
Update: @Evan Anderson made a good point that I forgot to mention. Broadcasting incoming packets on a LAN is generally a bad idea. This solution will work but it's only a hack to circumvent the limitations of the Wake On Lan protocol.
The technique I've outlined will work for any computer on the LAN the way WOL was designed but could potentially open your network up to be used for as an attack (Smurf/Fraggle/Papasmurf) amplifier if someone were to send a specially crafted packet to the WOL port.
Evan Anderson's approach is technically more secure but is limited to unicast.