The thing with procedural generation is that it's 90% defined for the specific application it is going to be used for. There's no way to tell you how to do it a certain way to make it turn out like how you want it to turn out, without you describing in detail how you want it to turn out.
Have you looked the advanced random tutorial?
construct.net/en/tutorials/getting-started-advanced-30
See the pictures of noise? Procedural generation is taking those, and defining what you want the black and white parts to be. The difference between the first picture which is pure randomness/noise, and the second perlin noise (what is normally used as a basis for many procedural algorithms), is that each location in a perlin noise map takes into account it's neighbors, sot that it will not jump from 0 (white) to 1 (black) abruptly. Different types of noise will give you different patterns
For example, you could take perlin noise, and decide that any x/y coordinate that has a value of >0.6 is water. So all the dark splotches are now water. You then layer other noise definitions on top of that. For example, after you decide what is water and land, you could say that if any location is land, check the same location on a random noise map and if it is 1, then that's a tree.
The other key about procedural generation is that the "random" values will be the same every time with any given input (usually x/y coordinate values, for a top down map), as long as you keep the same seed. But the values will change if you change the seed. So you could use perlin noise for terrain elevation, and then use another perlin noise with a different seed to stack different features, like resources, weather, trees, ect.
The very basic principle is that the advanced random noise function will give you a value from 0-1 for every x and y (possibly z as well, if you want) value input, and nearby values will not differ too much from each other. It is up to you to decide what that 0-1 value means in any given situation.