最後更新日期: 2021-03-22
Table of content
以下是我在 jupyter notebook 的分析過程，由於 markdown 的限制，並沒有把所有 output 都記錄在此份 README.md
懶人包: 我的作法並沒有用到 ML ，而是觀察之後自己造一個公式去估算，原因是:
- 看不出備載容量跟其他屬性有任何高度相關
- 預測的區間小，沒辦法利用大趨勢預測
- 每年 3 月備載容量的 pattern 長的不一樣，直接餵給常用的 model 感覺誤差會很大 (雖然我沒有嘗試過)
可能我太廢了，不知道怎麼分析以及套用適合的 ML 模型，希望可以觀摩助教或其他同學如何用機器學習來做這一題
import pandas as pd
# --- plotly ---
import plotly.graph_objs as go
import plotly.express as px
import plotly.io as pio
import datetime
pio.templates.default = "plotly_dark"-
讀取台灣電力公司_過去電力供需資訊作為
df1- 區間: 2019/01/01 ~ 2021/01/31
- 包含 “日期”、”尖峰負載(MW)”、”備轉容量(MW)”, “備轉容量率(%)”...
- ”備轉容量(MW)” (operating reserve) 是預測的目標
-
讀取台灣電力公司_本年度每日尖峰備轉容量率作為
df2- 區間: 2020/01/01 ~ 2020/12/25
- 包含 “日期”、”備轉容量(萬瓩)”、“備轉容量率(%)”
- 1 萬瓩 = 10 MW
查看 df1 的所有欄位
df1 = pd.read_csv('1.csv')
df2 = pd.read_csv('2.csv')
df1.columns其中只關注全局的資訊，把個別發電廠的欄位都捨棄，將工業用電和民生用電相加儲存到 總用電(百萬度) ，並且將日期變為 datetime
df1 = df1[['日期', '淨尖峰供電能力(MW)', '尖峰負載(MW)', '備轉容量(MW)', '備轉容量率(%)']]
tmp = []
for i in df1.index:
tmp.append(datetime.datetime.strptime(str(df1.loc[i, '日期']), '%Y%m%d'))
df1['日期2'] = tmp
df1用 plotly 畫出淨尖峰供電能力(MW), 尖峰負載(MW),備轉容量(MW) 的折線圖
fig = px.line(df1, x='日期2', y=['淨尖峰供電能力(MW)', '尖峰負載(MW)','備轉容量(MW)'], title='備轉容量')
fig.update_layout(title='供電、負載、備轉', xaxis_title='日期', yaxis_title='功率 (MW)')
fig.show()
可以發現 2019 年 4 月份的 備轉容量(MW) 的數值很奇怪，跟其他時間相比特別低，這裡我透過 excel 開啟檔案，透過觀察發現:
備轉容量(MW) = 淨尖峰供電能力(MW) - 尖峰負載(MW)
以下，我們將 備轉容量(MW) 重新計算，取代錯誤的值，然後再次觀察圖表
tmp = []
for i in df1.index:
tmp.append(df1.loc[i, '淨尖峰供電能力(MW)'] - df1.loc[i, '尖峰負載(MW)'])
df1['備轉容量(MW)'] = tmp
fig = px.line(df1, x='日期2', y=['淨尖峰供電能力(MW)', '尖峰負載(MW)','備轉容量(MW)'])
fig.update_layout(title='供電、負載、備轉', xaxis_title='日期', yaxis_title='功率 (MW)')
fig.show()
- 由上圖可以發現天氣越炎熱用電量越高，但是備載容量增加的沒那麼明顯
- 供用電呈現密集的震盪，我初步猜測震盪的高峰代表平日，低谷代表假日
我在下兩張圖將每個日期用圓點標示，觀察平日和假日用電的差異，並且我只擷取 3 至 4 月觀察，我認為其他月份的 pattern 參考意義不大
# begin day and end day
begin = 59
end = 119
fig = go.Figure()
fig.add_trace(go.Scatter(x=df1['日期2'][begin:end], y=df1['淨尖峰供電能力(MW)'][begin:end], mode='lines+markers', name='淨尖峰供電能力(MW)'))
fig.add_trace(go.Scatter(x=df1['日期2'][begin:end], y=df1['尖峰負載(MW)'][begin:end], mode='lines+markers', name='尖峰負載(MW)'))
fig.add_trace(go.Scatter(x=df1['日期2'][begin:end], y=df1['備轉容量(MW)'][begin:end], mode='lines+markers', name='備轉容量(MW)'))
title_str = '供電用電狀況 (' + str(df1['日期2'][begin]) + ' - ' + str(df1['日期'][end]) + ')'
fig.update_layout(title=title_str, xaxis_title='日期', yaxis_title='功率 (MW)')
fig.show()
# begin day and end day
begin = 425
end = 485
fig = go.Figure()
fig.add_trace(go.Scatter(x=df1['日期2'][begin:end], y=df1['淨尖峰供電能力(MW)'][begin:end], mode='lines+markers', name='淨尖峰供電能力(MW)'))
fig.add_trace(go.Scatter(x=df1['日期2'][begin:end], y=df1['尖峰負載(MW)'][begin:end], mode='lines+markers', name='尖峰負載(MW)'))
fig.add_trace(go.Scatter(x=df1['日期2'][begin:end], y=df1['備轉容量(MW)'][begin:end], mode='lines+markers', name='備轉容量(MW)'))
title_str = '供電用電狀況 (' + str(df1['日期2'][begin]) + ' - ' + str(df1['日期'][end]) + ')'
fig.update_layout(title=title_str, xaxis_title='日期', yaxis_title='功率 (MW)')
fig.show()
- 由上圖可以發現震盪的波峰通常都維持 5 天，對照日曆也確定是週一到五，而低谷則是星期六日和國定假日
- 備載容量卻沒有明顯的波峰波谷，而且通常備載容量通常在比前幾日高出許多後又會降回來
- 一樣，備載容量有時在假日時會增加，但沒那麼規律，基本上可以視為與尖峰負載無關
由此我得出一個結論: 備載容量在假日時有機會微幅提升，但大部分時候是穩定的
接下來我新增一個欄位，將資料標記星期一至星期天 (1 ~ 7)
day = 2
tmp = []
for i in df1.index:
tmp.append(day)
if day == 7:
day = 1
else:
day = day + 1
df1['星期'] = tmp
day = 5
tmp = []
for i in df2.index:
tmp.append(day)
if day == 7:
day = 1
else:
day = day + 1
df2['星期'] = tmp再看看 df2 (台灣電力公司_本年度每日尖峰備轉容量率) 的資料，先把備轉容量單位換成 MW
df2['備轉容量(MW)'] = df2['備轉容量(萬瓩)'] * 10
tmp = []
for i in df2.index:
tmp.append(datetime.datetime.strptime(str(df2.loc[i, '日期']), '%Y/%m/%d'))
df2['日期2'] = tmp
df22/27, 2/28, 2/29 是國定連假，會造成雜訊，故先把這三天 drop 掉
tmp = 0
while True:
if df2['日期'][tmp] == '2021/02/27':
break
else:
tmp = tmp + 1
df2 = df2.drop([tmp, tmp+1, tmp+2])
df2 = df2.reset_index(drop=True)觀察每日往前回推 4、7、14 日的均線
def avg_line(n):
lst = []
for i in range(n, len(df2)):
lst.append(sum(df2[i - n:i]['備轉容量(MW)'].tolist()) / n)
return lst
fig = go.Figure()
fig.add_trace(go.Scatter(x=df2['日期2'], y=df2['備轉容量(MW)'], mode='lines+markers', name='備轉容量(MW)'))
#fig.add_trace(go.Scatter(x=df2['日期2'][14:], y=avg_line(14), mode='lines', name = '14 日均線'))
fig.add_trace(go.Scatter(x=df2['日期2'][14:], y=avg_line(14), name='14 日均線', line = dict(dash='dash')))
fig.add_trace(go.Scatter(x=df2['日期2'][7:], y=avg_line(7), name = '7 日均線', line = dict(color='yellow', dash='dash')))
fig.add_trace(go.Scatter(x=df2['日期2'][4:], y=avg_line(4), name = '4 日均線', line = dict(color='green', dash='dash')))
title_str = '備轉容量 2021'
fig.update_layout(title=title_str, xaxis_title='日期', yaxis_title='功率 (MW)')
fig.show()
觀察 3、5、7 日的 moving averages
def avg_line(n):
lst = []
for i in range(n, len(df2)):
lst.append(sum(df2[i - n:i]['備轉容量(MW)'].tolist()) / n)
return lst
fig = go.Figure()
fig.add_trace(go.Scatter(x=df2['日期2'], y=df2['備轉容量(MW)'], mode='lines+markers', name='備轉容量(MW)'))
fig.add_trace(go.Scatter(x=df2['日期2'][1:], y=avg_line(3), name='3-period MA', line = dict(dash='dash')))
fig.add_trace(go.Scatter(x=df2['日期2'][2:], y=avg_line(5), name='5-period MA', line = dict(dash='dash')))
fig.add_trace(go.Scatter(x=df2['日期2'][3:], y=avg_line(7), name='7-period MA', line = dict(color='yellow', dash='dash')))
由均線圖和 moving averages 圖可以發現 moving averages 的趨勢要比前 n 日平均更準，不知道有沒有機會拿這兩樣資訊來預測
觀察每個星期一的備轉容量，list_by_day 可以觀看每周 x 的備轉容量， x 介於 1 至 7
def list_by_day(df, x):
date = []
lst = []
for i in range(0, len(df)):
if df['星期'][i] == x:
date.append(df['日期'][i])
lst.append(df['備轉容量(MW)'][i])
print(date)
print(lst)先觀察每周一的備轉容量
list_by_day(df2, 1)['2021/01/04', '2021/01/11', '2021/01/18', '2021/01/25', '2021/02/01', '2021/02/08', '2021/02/15', '2021/02/22', '2021/03/08', '2021/03/15']
[3032.0, 3093.0, 3157.0, 3269.0, 2938.0, 3047.0, 2645.0, 2989.0, 3110.0, 3074.0]
可以發現除了國定假日 2/15 (農曆新年)和 3/1 (二二八)之外，星期一的備轉容量都介於 2900 和 3200 之間，而且 3/8 和 3/15 分別為 3110 和 3074 ，再考慮之後天氣不會變冷，我認為 3/22 的備轉容量有很大的機率落在 3100 上下
接下來用相同的方法分析其他星期:
list_by_day(df2, 2)['2021/01/05', '2021/01/12', '2021/01/19', '2021/01/26', '2021/02/02', '2021/02/09', '2021/02/16', '2021/02/23', '2021/03/02', '2021/03/09', '2021/03/16']
[3067.0, 3116.0, 3159.0, 3219.0, 3045.0, 2880.0, 3002.0, 3109.0, 3221.0, 3048.0, 3087.0]
list_by_day(df2, 3)['2021/01/06', '2021/01/13', '2021/01/20', '2021/01/27', '2021/02/03', '2021/02/10', '2021/02/17', '2021/02/24', '2021/03/03', '2021/03/10', '2021/03/17']
[3102.0, 3114.0, 3149.0, 3064.0, 3006.0, 2888.0, 2824.0, 3049.0, 3075.0, 3055.0, 3183.0]
list_by_day(df2, 4)['2021/01/07', '2021/01/14', '2021/01/21', '2021/01/28', '2021/02/04', '2021/02/11', '2021/02/18', '2021/02/25', '2021/03/04', '2021/03/11', '2021/03/18']
[3102.0, 3132.0, 3074.0, 3286.0, 3132.0, 2833.0, 2907.0, 3029.0, 3088.0, 3077.0, 3214.0]
list_by_day(df2, 5)['2021/01/01', '2021/01/08', '2021/01/15', '2021/01/22', '2021/01/29', '2021/02/05', '2021/02/12', '2021/02/19', '2021/02/26', '2021/03/05', '2021/03/12', '2021/03/19']
[2767.0, 3105.0, 3310.0, 3097.0, 3203.0, 3060.0, 2992.0, 2885.0, 3041.0, 3268.0, 3272.0, 3137.0]
list_by_day(df2, 6)['2021/01/02', '2021/01/09', '2021/01/16', '2021/01/23', '2021/01/30', '2021/02/06', '2021/02/13', '2021/02/20', '2021/03/06', '2021/03/13', '2021/03/20']
[2936.0, 3202.0, 3046.0, 3203.0, 3447.0, 3066.0, 3024.0, 2975.0, 2805.0, 3075.0, 2839.0]
list_by_day(df2, 7)['2021/01/03', '2021/01/10', '2021/01/17', '2021/01/24', '2021/01/31', '2021/02/07', '2021/02/14', '2021/02/21', '2021/03/07', '2021/03/14', '2021/03/21']
[3037.0, 2906.0, 3035.0, 3073.0, 2906.0, 2763.0, 2911.0, 3139.0, 2752.0, 2933.0, 3186.0]
總結: 預測時應該會從 n 日平均或星期來下手，不考慮使用 ML 模型，因為備轉容量似乎跟供電、負載無關，而且只預測 7 天，不能依靠大趨勢，參考前兩個禮拜內的資料或同為星期 x 的資料應該比較準
因為備載容量找不到甚麼相關度較高的屬性，所以我決定先用前 n 天的平均來做為我預測的值，例如: 若 n = 5 且預測目標是 3/23 ，我就拿 3/18 至 3/22 的備載容量拿來平均，作為預測的值
from sklearn.metrics import mean_squared_error
from math import sqrt首先我定義 orig_data 回傳以 begin_date 為基準取得 interval 天的備載容量
例如: orig_data('2021/03/10', 3) 即為取得 2021/3/10 到 2021/3/12 這三天的備載容量
def orig_data(begin_date, interval):
tmp = 0
for i in df2.index:
if df2.loc[i, '日期'] == begin_date:
break
else:
tmp = tmp + 1
return df2[tmp: tmp + interval]['備轉容量(MW)'].tolist()為了避開農曆新年，我只用以下日期來測試
test_lst = []
tmp_str = '2021/01/'
for i in range(15, 31):
test_lst.append(tmp_str + str(i))
tmp_str = '2021/03/0'
for i in range(2, 10):
test_lst.append(tmp_str + str(i))
tmp_str = '2021/03/'
for i in range(10, 16):
test_lst.append(tmp_str + str(i))
print(test_lst)['2021/01/15', '2021/01/16', '2021/01/17', '2021/01/18', '2021/01/19', '2021/01/20', '2021/01/21', '2021/01/22', '2021/01/23', '2021/01/24', '2021/01/25', '2021/01/26', '2021/01/27', '2021/01/28', '2021/01/29', '2021/01/30', '2021/03/02', '2021/03/03', '2021/03/04', '2021/03/05', '2021/03/06', '2021/03/07', '2021/03/08', '2021/03/09', '2021/03/10', '2021/03/11', '2021/03/12', '2021/03/13', '2021/03/14', '2021/03/15']
定義 model1 作為前 n 日平均的預測函數，以 begin_date 為基準預測 interval 天的備載容量，用 n 天前的資料平均，再針對假日做一些調整
def model1(begin_date, interval, n):
tmp = 0
for i in df2.index:
if df2.loc[i, '日期'] == begin_date:
break
else:
tmp = tmp + 1
lst = df2[tmp - n:tmp]['備轉容量(MW)'].tolist()
ret = []
mean = sum(lst) / n
for i in range(0, interval):
ret.append(int(mean))
return retevaluate 和 print_evaluate 是使用 Root Mean Squared Error (RMSE) 計算 model1 模型的誤差
def evaluate(begin_date, interval, n):
pred = model1(begin_date, interval, n)
actual = orig_data(begin_date, interval)
return sqrt(mean_squared_error(actual, pred))
def print_evaluate(begin_date, interval, n):
pred = model1(begin_date, interval, n)
actual = orig_data(begin_date, interval)
rtms = sqrt(mean_squared_error(actual, pred))
print('from:', begin_date, 'interval:', interval, 'days',)
print('actual data:', actual)
print('pred data: ', pred)
print('RTMS error:', rtms)以下舉一個例子，假設預測 03/15 往後的 7 天 (含 03/11)，取往前 13 天的平均，所計算出的 RMSE
print_evaluate('2021/03/15', 7, 13)from: 2021/03/15 interval: 7 days
actual data: [3074.0, 3087.0, 3183.0, 3214.0, 3137.0, 2839.0, 3186.0]
pred data: [3059, 3059, 3059, 3059, 3059, 3059, 3059]
RTMS error: 125.93819345785683
接下來我檢視單純計算 7 至 14 日均線 (即 model1) 預測後 7 日的 RMSE。
舉例來說， x 軸座標為 2021/03/10 ，對應到 n = 10 的折線的 y 座標，其 y 值即為對前 10 天備載容量平均，對後 7 天套用這個平均值所算出的 RMSE
n1, n2 = 7, 15
# test
result = []
for i in range(n1, n2):
tmp = []
for j in range(0, len(test_lst)):
tmp.append(evaluate(test_lst[j], 7, i))
result.append(tmp)
# show errors
fig = go.Figure()
for i in range(0, n2 - n1):
fig.add_trace(go.Scatter(x=test_lst, y=result[i], mode='lines+markers', name='n=' + str(n1 + i)))
fig.update_layout(title='Error', xaxis_title='日期', yaxis_title='RMSE')
fig.show()
- 由於星期一的備載容量通常較高，所以將預測的星期一的備轉容量拉高 30
- 由於星期六日通常是該周備轉容量最低的，所以將預測的星期日的備轉容量拉低 30 (數字 30 是我憑感覺亂選的，也許可以用暴力法挑出最優的數字，但有 overfitting 的疑慮)
def model1(begin_date, interval, n):
tmp = 0
day = 0
for i in df2.index:
if df2.loc[i, '日期'] == begin_date:
day = df2['星期'][i]
break
else:
tmp = tmp + 1
lst = df2[tmp - n:tmp]['備轉容量(MW)'].tolist()
ret = []
days = []
mean = sum(lst) / n
for i in range(0, interval):
temp = int(mean)
# ---
if day == 1:
temp = temp + 30
elif day == 6 or day == 7:
temp = temp - 30
ret.append(temp)
# ---
if day == 7:
day = 1
else:
day = day + 1
return ret
print_evaluate('2021/03/15', 7, 13)from: 2021/03/15 interval: 7 days
actual data: [3074.0, 3087.0, 3183.0, 3214.0, 3137.0, 2839.0, 3186.0]
pred data: [3089, 3059, 3059, 3059, 3059, 3029, 3029]
RTMS error: 123.77571882816343
n1, n2 = 7, 15
# test
result = []
for i in range(n1, n2):
tmp = []
for j in range(0, len(test_lst)):
tmp.append(evaluate(test_lst[j], 7, i))
result.append(tmp)
# show errors
fig = go.Figure()
for i in range(0, n2 - n1):
fig.add_trace(go.Scatter(x=test_lst, y=result[i], mode='lines+markers', name='n=' + str(n1 + i)))
fig.update_layout(title='Error', xaxis_title='日期', yaxis_title='RMSE')
fig.show()
經過微調後，3月份預測的 RMSE 幾乎沒變...
簡單來說酒是取同為星期 x 的備載容量來平均，例如: 預測 03/11 的備載容量，則拿同為星期四的 03/04、02/25、02/18 來平均。至於二二八連假這種被 drop 掉的，則會再往前取一個禮拜的資料來替代。
def list_by_day2(df, end, day):
lst = []
for i in range(end):
if df['星期'][i] == day:
lst.append(df['備轉容量(MW)'][i])
return lst
def model2(begin_date, interval, n):
tmp = 0
day = 0
for i in df2.index:
if df2.loc[i, '日期'] == begin_date:
day = df2.loc[i, '星期']
break
else:
tmp = tmp + 1
ret = []
for i in range(0, interval):
mean = 0
div = 0
lst = list_by_day2(df2, tmp + i, day)
lst.reverse()
if n > len(lst):
mean = sum(lst) / len(lst)
else:
mean = sum(lst[0:n]) / n
ret.append(int(mean))
if day == 7:
day = 1
else:
day = day + 1
return retdef evaluate(begin_date, interval, n):
pred = model2(begin_date, interval, n)
actual = orig_data(begin_date, interval)
return sqrt(mean_squared_error(actual, pred))
def print_evaluate(begin_date, interval, n):
pred = model2(begin_date, interval, n)
actual = orig_data(begin_date, interval)
rtms = sqrt(mean_squared_error(actual, pred))
print('from:', begin_date, 'interval:', interval, 'days',)
print('actual data:', actual)
print('pred data: ', pred)
print('RTMS error:', rtms)假設預測 03/14 往後的 7 天 (含 03/14)，對於 03/14 取往前 3 周同為周一的平均 (03/07、(跳過 02/28)、02/21、02/14)，03/15 依此類推，所計算出的 RMSE
print_evaluate('2021/03/14', 7, 3)from: 2021/03/14 interval: 7 days
actual data: [2933.0, 3074.0, 3087.0, 3183.0, 3214.0, 3137.0, 2839.0]
pred data: [2934, 2914, 3126, 3059, 3064, 3193, 2951]
RTMS error: 107.35654880550403
n1, n2 = 2, 5
# test
result = []
for i in range(n1, n2):
tmp = []
for j in range(0, len(test_lst)):
tmp.append(evaluate(test_lst[j], 7, i))
result.append(tmp)
# show errors
fig = go.Figure()
for i in range(0, n2 - n1):
fig.add_trace(go.Scatter(x=test_lst, y=result[i], mode='lines+markers', name='n=' + str(n1 + i)))
fig.update_layout(title='Error', xaxis_title='日期', yaxis_title='RMSE')
fig.show()
--- 03/18 ---
考慮到往後天氣有很高的機會越來越熱，我將過往周日以外的備轉容量全部加上 70 以利於計算上降低誤差，如果加上 70 之後仍小於 3000 ，我會將其值直接設為 3000。 (希望不會 overfitting...)
--- 03/22 ---
這周突然變冷，查了氣象發現有東北季風，所以可能要修正策略
def model2(begin_date, interval, n):
tmp = 0
day = 0
for i in df2.index:
if df2.loc[i, '日期'] == begin_date:
day = df2.loc[i, '星期']
break
else:
tmp = tmp + 1
ret = []
for i in range(0, interval):
mean = 0
div = 0
lst = list_by_day2(df2, tmp + i, day)
lst.reverse()
# ---
if day != 7 and day != 6:
for j in range(len(lst)):
lst[j] = lst[j] + 50
# ---
if n > len(lst):
mean = sum(lst) / len(lst)
else:
mean = sum(lst[0:n]) / n
ret.append(int(mean))
if day == 7:
day = 1
else:
day = day + 1
return retprint_evaluate('2021/03/14', 7, 3)from: 2021/03/14 interval: 7 days
actual data: [2933.0, 3074.0, 3087.0, 3183.0, 3214.0, 3137.0, 2839.0]
pred data: [2934, 2964, 3176, 3109, 3114, 3243, 2951]
RTMS error: 92.02328897776755
n1, n2 = 2, 5
# test
result = []
for i in range(n1, n2):
tmp = []
for j in range(0, len(test_lst)):
tmp.append(evaluate(test_lst[j], 7, i))
result.append(tmp)
# show errors
fig = go.Figure()
for i in range(0, n2 - n1):
fig.add_trace(go.Scatter(x=test_lst, y=result[i], mode='lines+markers', name='n=' + str(n1 + i)))
fig.update_layout(title='Error', xaxis_title='日期', yaxis_title='RMSE')
fig.show()
經過上述微調後，3月份預測的 RMSE 降低 20 到 30 左右
至此 model 大概已經定案了，就用降低過誤差的 model2 來預測，取 n = 3
首先填補 df2 到 2021/03/21 為止的缺失值放進 train_df
train_df = df2.copy()
date22 = datetime.datetime.strptime('2021/03/22', '%Y/%m/%d')
mar22 = pd.Series({'日期':'2021/03/22', '備轉容量(萬瓩)':298.0, '備轉容量率(%)':10.22, '星期':1, '備轉容量(MW)':2980, '日期2':date22})
train_df = train_df.append(mar22, ignore_index=True)建立模型
def list_by_day3(df, end, day):
lst = []
for i in range(end):
if df['星期'][i] == day:
lst.append(df['備轉容量(MW)'][i])
return lst
def model3(interval, n):
day = 2
ret = []
for i in range(0, interval):
mean = 0
div = 0
lst = list_by_day3(train_df, len(train_df), day)
lst.reverse()
# ---
if day != 7 and day != 6:
for j in range(len(lst)):
lst[j] = lst[j] + 50
# ---
if n > len(lst):
mean = sum(lst) / len(lst)
else:
mean = sum(lst[0:n]) / n
ret.append(int(mean))
if day == 7:
day = 1
else:
day = day + 1
return retpred = model3(7, 3)
print(['20210323', '20210324', '20210325', '20210326', '20210327', '20210328', '20210329'])
print(pred)['20210323', '20210324', '20210325', '20210326', '20210327', '20210328', '20210329']
[3168, 3154, 3176, 3275, 2906, 2957, 3104]
結束
預測備轉容量
台灣近年來環境：台灣近年來巔峰負載容量都是越來越多，也隨著季節有明顯的變化，而且工業用電佔整體很大比例。
從政府的平台上取得的資料，可以觀察到大約每7天為一個週期，而且巔峰尖峰負載和備載容量有很大的相關，而溫度又和巔峰負載有關系。
Warning：
台電的電力預測小組目標就是控制備載容量率，可以很明顯的觀察到近年來，台電的預測越來越準確，109年相比於108年，明顯感覺得到備載容量率有控制住
包裝資料
def create_data(past = 7, future = 7):
x_train = []
y_train = []
x_test = []
y_test = []
for i in range(past+future, len(data)+1):
t = data.iloc[i-past-future:i-future][['backup']]
y = data['備轉'].iloc[i-future:i]
if i < 2600:
x_train.append(t)
y_train.append(y)
else:
x_test.append(t)
y_test.append(y)
x_train = np.array(x_train)
y_train = np.array(y_train)
x_test = np.array(x_test)
y_test = np.array(y_test)
return x_train, y_train, x_test, y_test模型測試，測試要往前看幾天和往後看幾天，test data為3/14號往前28天
def mean_model(x_train, y_train, x_test, y_test, max_past_len, max_future_len, mode):
#test mode
if mode == "test":
count = data['備轉'][:max_past_len]
count = count.values.tolist()
predict_list = []
for i in range(7):
predict = sum(count) / max_past_len
predict_list.append(predict)
count = count.pop(0)
count.append(predict)
return predict_list
#train mode
else:
err = 0
result = []
for i, j in zip(x_test, y_test):
count = [k[0] for k in i]
for h in j:
predict = sum(count)/len(count)
err += (predict-h)**2
count.pop(0)
count.append(predict)
err = math.sqrt(err/max_future_len)
result.append(err)
err = 0
return sum(result)/len(result)
for i in range(1, 11):
for j in range(1, 8):
x_train, y_train, x_test, y_test = create_data(i, j)
print("past",i, "future", j, mean_model(x_train, y_train, x_test, y_test, i, j, "train"))結果：
past 1 future 1 151.41379310344828
past 1 future 2 179.79823390800894
past 1 future 3 194.7425855780763
past 1 future 4 203.48558309376423
past 1 future 5 207.76746222992378
past 1 future 6 211.75791090697138
past 1 future 7 212.6119237313435
past 2 future 1 145.1551724137931
past 2 future 2 165.03463789692958
past 2 future 3 179.59107739212575
past 2 future 4 189.58056634563434
past 2 future 5 195.32329523276132
past 2 future 6 196.85980399942432
past 2 future 7 193.38889932410663
past 3 future 1 139.44827586206893
past 3 future 2 164.6824806851891
past 3 future 3 177.51424193643348
past 3 future 4 183.5411259892367
past 3 future 5 187.84179979319882
past 3 future 6 188.19244060655242
past 3 future 7 182.485819567202
past 4 future 1 143.94827586206895
past 4 future 2 166.5319781008546
past 4 future 3 175.9036635424794
past 4 future 4 180.56602748536602
past 4 future 5 181.68210171774652
past 4 future 6 180.70728932521385
past 4 future 7 175.9194112913405
past 5 future 1 147.38620689655176
past 5 future 2 166.06213223809536
past 5 future 3 175.46447274466794
past 5 future 4 179.40721836744996
past 5 future 5 180.19841131107123
past 5 future 6 178.11205039279966
past 5 future 7 174.1325584915976
past 6 future 1 140.18965517241378
past 6 future 2 160.1665321934286
past 6 future 3 170.0038430455642
past 6 future 4 173.88728377277627
past 6 future 5 175.68718157267318
past 6 future 6 174.78914837789216
past 6 future 7 171.3865644366255
past 7 future 1 134.79802955665025
past 7 future 2 151.68731493174286
past 7 future 3 160.7376623067602
past 7 future 4 164.90945903286106
past 7 future 5 167.3367144671708
past 7 future 6 167.65450283134615
past 7 future 7 165.792102043254
past 8 future 1 131.2456896551724
past 8 future 2 146.59565201139836
past 8 future 3 154.73849836189487
past 8 future 4 158.96755695951373
past 8 future 5 161.2653762299115
past 8 future 6 162.0903252334757
past 8 future 7 161.09172009816754
past 9 future 1 135.32950191570885
past 9 future 2 148.6760313812843
past 9 future 3 155.61949568726467
past 9 future 4 159.24018478802412
past 9 future 5 161.20481867429885
past 9 future 6 161.93323458396787
past 9 future 7 161.19533601000532
past 10 future 1 135.27931034482756
past 10 future 2 148.03853073758447
past 10 future 3 154.05649266249495
past 10 future 4 157.33616928462044
past 10 future 5 159.24215239690537
past 10 future 6 160.45116390104045
past 10 future 7 160.0595167528463
以結果來說往前看7或8天似乎是最好的決定
- model:
- 多對多模型，利用過去7天，去預測未來7天
- model
#lstm model def buildmanytomany(shape): model = Sequential() model.add(LSTM(10, input_length=shape[1], input_dim=shape[2], return_sequences=True)) model.add(TimeDistributed(Dense(1))) model.compile(loss="mse", optimizer="adam") model.summary() return model
- training
x_train, y_train, x_test, y_test = create_data(7, 7) regressor = buildmanytomany(x_train.shape) print(x_train.shape) # 進行訓練 regressor.fit(x_train, y_train, epochs = 650)
- model
x_train, y_train, x_test, y_test = create_data(7, 7) regressor = buildmanytomany(x_train.shape) print(x_train.shape) # 進行訓練 regressor.fit(x_train, y_train, epochs = 650)
- predict
x_train, y_train, x_test, y_test = create_data(7, 7) regressor = buildmanytomany(x_train.shape) print(x_train.shape) # 進行訓練 regressor.fit(x_train, y_train, epochs = 650)
- 多對一模型，利用過去7天，去預測未來1天模型，並且連續預測7天
- build model
def buildManyToOneModel(shape): model = Sequential() model.add(LSTM(10, input_length=shape[1], input_dim=shape[2], return_sequences=True)) model.add(Dense(units = 10)) model.add(Dropout(0.2)) model.add(Dense(units = 10)) model.add(Dropout(0.2)) model.add(Dense(1)) model.compile(loss="mse", optimizer="adam") model.summary() return model
- build model
- training data
x_train, y_train, x_test, y_test = create_data(8, 1) print(x_train.shape) regressor = buildManyToOneModel(x_train.shape) # 進行訓練 regressor.fit(x_train[-100:]*10, y_train[-100:], epochs = 350)
- predict data
def lstm_model(x_train, y_train, x_test, y_test, model): err = 0 result = [] for i,j in zip(x_test, y_test): i = i.reshape(1, 8, 1) print(i*10) predict = model.predict(i) print(predict) predict = predict[0][0] j = j[0] err = (predict-j)**2 result.append(err) print(math.sqrt(err)) return math.sqrt(sum(result)/len(result)) print(lstm_model(x_train, y_train, x_test, y_test, regressor))
- data construct
def create_week_data(past = 3): # model = build() x_train = [] y_train = [] x_test = [] y_test = [] for i in range(0, len(data)-7*past): t = [] for j in range(past): t.append(data.iloc[i+j*7][['備轉']]) y = data['備轉'].iloc[i+j*7+7] if i >= 2586 and i <= 2613: continue if i < 2500: x_train.append(t) y_train.append(y) else: x_test.append(t) y_test.append(y) x_train = np.array(x_train) y_train = np.array(y_train) x_test = np.array(x_test) y_test = np.array(y_test) return x_train, y_train, x_test, y_test x_train, y_train, x_test, y_test = create_week_data(3)
- model
past = 3 x_train, y_train, x_test, y_test = create_week_data(past) model = baseline_model(x_test.shape) x_train = x_train.reshape(-1, past) x_test = x_test.reshape(-1, past) regr = MLPRegressor(random_state=1, max_iter=2000, hidden_layer_sizes=10).fit(x_train, y_train)
- predict
predict = [] for i in range(7): x = x_test[-14+i].reshape(-1, past) predict.append(regr.predict(x)[0])
- result
result = [2933, 3074, 3087, 3183, 3214, 3137, 2839] err = 0 for i in range(7): err += (result[i]-predict[i])**2 print(math.sqrt(err/7))
因為每週之間的權重佔比變化比較大，沒辦法完全收斂，所以最後的結果總是比lincc還高，尤其是3/14。
- 可以去尋找feature之間的correlation，比如說某幾個發電廠和備轉容量之間的關係。刪除離群值也是非常重要的。
- 使用Prophet & 資料時間尺度的選擇很重要，越近期的越重要。
Thank lincc and everyone!