关于区间预测-PICP\PINAW\CWC\SCORE\等公式

最新推荐文章于 2024-07-23 22:38:20 发布

强化学习RL

最新推荐文章于 2024-07-23 22:38:20 发布

阅读量1.9k

点赞数 1

文章标签：深度学习

本文链接：https://blog.csdn.net/m0_61015717/article/details/128350150

版权

import numpy as np

def calculate_metrics(y_true, y_pred, alpha):
    """
    计算区间预测的评价指标

    参数:
        y_true: 真实值，shape为(n_samples,)
        y_pred: 区间预测值，shape为(n_samples, 2)
        alpha: 置信水平，取值范围为(0, 1)

    返回:
        PICP: 区间预测置信度，float
        PINAW: 区间预测置信区间宽度，float
        SCORE: 区间预测得分，float
        SKILL_SCORE: 区间预测技能得分，float
        COVERAGE: 区间覆盖率，float
        WIDTH: 区间宽度，float
    """
    lower_bound = y_pred[:, 0]
    upper_bound = y_pred[:, 1]
    n_samples = len(y_true)

    # 计算PICP
    in_interval = np.logical_and(y_true >= lower_bound, y_true <= upper_bound)
    PICP = np.sum(in_interval) / n_samples

    # 计算PINAW
    PINAW = np.mean(upper_bound - lower_bound)

    # 计算SCORE
    score = np.zeros(n_samples)
    for i in range(n_samples):
        if y_true[i] < lower_bound[i]:
            score[i] = 2 * (lower_bound[i] - y_true[i]) / (upper_bound[i] - lower_bound[i])
        elif y_true[i] > upper_bound[i]:
            score[i] = 2 * (y_true[i] - upper_bound[i]) / (upper_bound[i] - lower_bound[i])
        else:
            score[i] = 0
    SCORE = np.mean(score)

    # 计算SKILL_SCORE
    skill_score = np.zeros(n_samples)
    for i in range(n_samples):
        if y_true[i] < lower_bound[i] or y_true[i] > upper_bound[i]:
            skill_score[i] = -1
        else:
            skill_score[i] = 2 * (upper_bound[i] - lower_bound[i]) / (upper_bound[i] - lower_bound[i] + (y_pred[i, 0] - y_true[i]) ** 2 + (y_true[i] - y_pred[i, 1]) ** 2)
    SKILL_SCORE = np.mean(skill_score)

    # 计算COVERAGE
    COVERAGE = np.sum(upper_bound - lower_bound) / n_samples

    # 计算WIDTH
    WIDTH = np.mean(upper_bound - lower_bound)

    return PICP, PINAW, SCORE, SKILL_SCORE, COVERAGE, WIDTH

以上代码来自“给我写一个采用PICP、PINAW、SCORE 、SKILL SCORE、区间覆盖率、区间宽度来评价区间预测的代码 - CSDN文库”

CWC作为区间预测的评价指标，论文IEEE Xplore Full-Text PDF:指出是不合适的。

时间序列预测之区间预测方法（PIs:MVE&Delta&Bayesian&Bootstrap&LUBE） | 码农家园 (codenong.com)

这个链接有每个公式的解释。

以下是我修改后的代码

''' SCORE：衡量真实值（y_true）与预测区间（由上限upper_bound和下限lower_bound定义）之间的偏差程度。'''
''' http://dx.doi.org/10.1016/j.energy.2014.06.104'''
# alpha = 0.01;0.05;0.1 置信水平（1-alpha)%，为99%,95%,90%

def calculate_S(alpha, Li, Ui, true):
    m = len(true)
    S = 0

    for i in range(m):
        di = Ui[i] - Li[i]

        if true[i] < Li[i]:
            Si = -2 * alpha * di - 4 * (Li[i] - true[i])
        elif true[i] >= Li[i] and true[i] <= Ui[i]:
            Si = -2 * alpha * di
        else:
            Si = -2 * alpha * di - 4 * (true[i] - Ui[i])

        S += Si

    S /= m
    SCORE = np.abs(S) # 确保Socre的值是正数，值越小越好
    return SCORE

alpha = 1- confidence
SCORE = calculate_S(alpha, Lower_sequence, Upper_sequence,true)
print()
print('SCORE', SCORE)

''' MPICD:'''

def calculate_MPICD(true, Li, Ui):
    m = len(true)
    MPICD = 0

    for i in range(m):
        pi = true[i]
        avg = (Ui[i] + Li[i]) / 2
        diff = abs(avg - pi)
        MPICD += diff

    MPICD /= m

    return MPICD

MPICD = calculate_MPICD(true, Lower_sequence, Upper_sequence)
print("MPICD", MPICD)